Casey Handmer's blog

Casey Handmer May 2026

What?

Elon has recently (late 2025, early 2026) been talking about building many terawatts of orbital AI compute and launching some components from Moon factories with a mass driver. This is an old idea, enabled by the Moon’s relatively low gravity and lack of an atmosphere. See, for example, The Moon is a Harsh Mistress and The High Frontier. 

The fundamental problem with The High Frontier is that the set of products that can be made in space and sold on Earth while making money is very limited, due to the sheer cost and difficulty of accessing space. In general, they are observations and communication, which in both cases distributes the product using radio waves, which are much cheaper than physical return of artifacts from space.

In 2019, I wrote that Starlink was likely to be incredibly lucrative and I’m happy to see this is the case, with over $10b in revenue last year. Space AI takes this business model and ramps it up to 11. Why? Starlink has already established that converting a space solar photon into an electron and using it to relay bits of information around the world is extremely profitable. Space AI is a great way to vastly increase both the total demand for space data bits as well as the value per bit, as the tokens encoded by these bits have already proven to have stupendous economic value and apparently unlimited demand.

Why?

Starship promises a near future with launch costs to LEO of perhaps $100/kg. With electric propulsion, large quantities of cargo, including solar powered AI can be positioned anywhere in cis-Lunar space for an incremental cost beyond that. For AI hardware, most of the cost is in the GPU/TPU die, which contributes almost no mass, while most of the mass is in the solar panel and radiator, which cost (relatively speaking) almost nothing. 

Here’s a spreadsheet I put together last year with some basic first-principles analysis. At even $500/kg, launch cost is only 5% of the total satellite deployment cost, so a lunar mass driver is unlikely to drastically improve the economics of space-based AI, by reducing launch costs. 

It’s also unlikely to have low start up costs! 

Instead, we must look to a future where Starship costs stop falling from experience and economies of scale and rise to unaffordable levels, perhaps comparable to the Shuttle’s $50,000/kg, because of a constraint on launch capacity.

In my spreadsheet, I estimate that one Starship can deliver about 15 MW of solar power to orbit. Last year, China produced over 1 TW of solar photovoltaic panels. Supposing we weren’t constrained by chip fabrication and Starship was fully operational, it could launch 1 TW to orbit per year with just 67,000 launches, or one every 8 minutes. 

This might seem like a lot, but the world currently sustains about 100,000 commercial flights per day! In a world where SpaceX can turn around a launch site in an hour, only seven or eight pads would be necessary to keep up with this rate of launch, requiring a fleet of perhaps 10 boosters and a few hundred Starships. 

Nor would this launch rate defeat our global oil production. One Starship launch consumes roughly 10,000 barrels of oil (equivalent), and the world currently consumes 100 million per day. So 67,000 launches per year is less than a week of the world’s current supply of oil. It may require a few gas pipelines in Texas to be upgraded, and of course by the time this happens solar synthetic fuel will be a recognized and mature technology. 

There has been some speculation about damage to the upper atmosphere of the Earth caused by huge launch volumes. 

In any case, we’re talking about a launch volume of hundreds of thousands of Starships per year, or more than 10 million tonnes of cargo per year, with a total launch revenue of about a trillion dollars – equivalent to about six weeks of the global oil and gas industry!

The lunar mass driver must transcend this scale. 

What does the lunar mass driver drive?

The Moon is made of rocks. Primarily volcanic rock, similar to Earth basalts. As ores go, they are not preferred sources of metals on Earth. Though they contain nearly every metal – the net present value of the metal in 1 tonne of basalt is about $1300 vs the $20 price as crushed gravel – they’re mixed together and generally considered energetically infeasible to extract. We’re working on this at Terraform but the energy demand is a fact of life. 

In one model, the lunar mass driver fires raw rocks into Lunar orbit, to be processed in space using copiously available space solar power. In another model, moon rocks are pre-processed to increase their metal content, or even converted into finished products, before launch. Blue Origin has demonstrated a process to convert Lunar regolith (dirt) into a functioning solar panel, but it’s not clear what the energy return on energy invested for this process would be.

The Moon’s surface itself can be a tough place to do anything energetic, because it is subjected to 14 Earth days of shade during the long lunar night, followed by 14 days of unrelenting sun during the day. Any serious infrastructure will require serious power, either from extremely large nuclear reactors operated near the poles to create functional shaded radiators, or from energy beamed up from the Earth or from Lunar orbit, or both. 

For the following I’ll assume the driver is shifting dumb rock. It doesn’t change much but the g-tolerance of raw materials is a big plus! 

How big is my mass driver?

Let’s run some numbers around mass flow rate. We’re not going to the trouble of building a mass driver for no reason, we’re doing it to alleviate the burden on Earth of launching 1000 Starships every day, to push total human compute into the 10s of TWs incremental increase per year. 

So let’s assume one mass driver launches 10 million tonnes of rock per year. Taking into account rock refining losses and expansion this implies a Lunar fleet of a few dozen mass drivers, but you have to start somewhere. 10 million tonnes per year is 1 tonne every 3 seconds. 

We can calculate the total energy expenditure too. Delta-V to LLO is 1.6 km/s, and we assume that Delta-V is cheaper to come by in orbit, thanks to solar powered tugs. We only need the lunar mass driver to get rocks into orbit, where they are collected and moved to wherever they need to go. 

So total kinetic power is 0.5*mdot*v^2 = 450 MW, assuming 90% driver efficiency. The 10% waste covers ohmic losses, cargo-sled recycle weight, and active cooling. 

Why not stack some additional assumptions on top of this result? 

Let’s assume an equivalent price of $10/kg, reflecting that random rocks orbiting the moon are not quite as valuable to the customer as a finished satellite orbiting the Earth – where all the customers are. This implies that each lunar mass driver makes $100b/year. Assuming that 10% of this revenue pays for the power plant, this works out to $2.50/kWh, which is about 10x higher than a typical US rate payer in 2026. 

Can a 450+ MW nuclear power plant be built and operated on the Moon for a 10x cost increment relative to Earth? I’m not sure but it’s not forbidden by the laws of physics. 

For reference, a reactor of this scale would typically cost $2-4b on Earth and weighs perhaps 1000 T. 

I previously estimated that a radiator-constrained space reactor mounted into a Starship could generate perhaps 3 MWe, implying several hundred launches for the power of one mass driver. A much larger, monolithic space reactor of 500 MW scale would need to be completely re-engineered, requiring delivery, an enormous radiator (many hundreds of acres, assuming permanent shadow), and other hassles. But it could be done. 

Orbit, what orbit?

Low lunar orbits are, in general, unstable, due to the presence of gravitational anomalies called mascons (or mass concentrations) in various places corresponding to ancient impacts. There are, however, four classes of frozen orbits (~27°, 50°, 76°, 86°) on the Moon that are relatively stable, so provided the mass driver has a latitude and launch azimuth high enough to access these, the launched rocks won’t necessarily immediately return to obliterate the launch site two hours after launching. On a long enough timescale any passive payload launched from the surface of the Moon into Lunar orbit will run into the surface, so the trick is to launch payloads into converging bunches and scoop them up in orbit, performing circularization and/or relay transportation using some kind of orbital tug. It’s also possible to launch rocks from the Moon into more distant libration orbits, but I don’t cover that case in this post.

Technical implementation

The launcher in this image looks a bit like a rail gun. But rail guns pass current from rail to rail through the sabot (projectile enclosure) and suffer rail erosion at far higher rates than we can tolerate.

The lunar mass driver will be a maglev in disguise. 

In this section I draw on my ancient experience as a levitation engineer at Hyperloop back in about 2016. 

The job of the track is to accelerate a passive cargo-carrying sled to high speed over a short distance. The sled must be controlled over six degrees of freedom. Unlike a passenger maglev, where propulsion forces are a fraction of gravitational forces, a mass driver is optimized for highly efficient acceleration. 

Assuming a launch speed of 1.6 km/s, v^2 = 2 a s gives track length and acceleration as inversely proportional. 

If rocks can survive 1000 gs of acceleration (they can) then the launch track need only be 128 m long (or 256 m including the sled catching portion), greatly reducing its mechanical and construction complexity. Can an electromagnetic launch system deliver 1000s of gs? 

A helpful intuition pump for this is to consider how much force a permanent magnet can deliver. A commercial neodymium magnet can easily support 20x its own weight, but 2000x is highly non-trivial. Bear in mind that no matter how cleverly built, the track will have to be toleranced with non-zero gap between the sled and the maglev track. 

Add to this the fact that the acceleration is diluted by the mass ratio of the magnet to the rest of the system. I can imagine a launch sled which is 70% magnets, 10% structure, and 20% rocks, in which case even an optimal launch system would have to take a 30% haircut on acceleration. 

There are a few ways to skin this cat, but probably the easiest is with a synchronous linear motor along a pair of parallel tracks with the sled suspended in between. At any meaningful level of acceleration, gravitational forces from the Moon round to zero. It is possible to include, for example, some null flux electrodynamical suspension system for guidance, but if you’re using the sled magnets for anything other than acceleration then you’re throwing away performance and making the track longer. 

On this note, achieving anything like the necessary levels of magnetic shear force requires very very large, very flat magnets. In my model below, they are 20 cm wide, 2.8 cm thick, and 9 m long, entirely enclosed by an electromagnetic stator, and connected via high strength steel shear panels to the payload bucket. 

It’s not impossible to build long tracks but, having seen this close up at Hyperloop, it becomes extremely difficult to achieve sub millimeter alignment precision over long distances. To give a flavor of this, the Shanghai Pudong transrapid Maglev operated with a ride height of about 1 mm. It is built on an elevated steel reinforced concrete trackway whose foundations were dug to a depth of 80 m into the Pudong silt. Shanghai, like many cities, is built on a river delta. Literally thousands and thousands of tonnes of concrete and steel to form the track way. You would think that it would be stable enough that the track could be calibrated once during construction and then would be fine, but in fact the track had to be re-aligned, by a custom designed track maintenance vehicle, twice per day, to account for such perturbations as tidal deformation of the crust of the Earth. 

The Moon does not have saturated silty deltas or wildly varying tidal forces (being tidally locked) but it does have a very extreme temperature cycle over its 28 Earth-day day, so probably the best way to achieve dimensional stability, as well as a measure of meteorite protection, is to bury the track under a few meters of dirt.

(I asked AI to make a better version of this diagram but it wasn’t right. Motion is into the page. 200 kg of moon rocks can fit in a container 40 cm on a side.)

(Here’s an OnShape model. Below, diagrams of how the whole thing goes together.)

The cart oscillates back and forth on the track, launching rocks for ¼ of its cycle, with the rocks separating from the sled bucket at the midpoint of the track. The slow down step can recover some momentum as power, provided there’s a place to store it! At 1000gs, launch takes just 0.16s and the complete cycle (accelerate, release rocks, slow down, fly back to rock loading site) is as short as 0.64 s, not including time to reload rocks. At the 10 million tonnes per year rate, each launch would carry about 200 kg, with a total sled weight of 1000 kg loaded and 800 kg empty. This launch and recycle process would load the structure with up to 1000 T, that is 10 MN, of force, at about 1.4 Hz. That’s a pretty wild vibrational environment for the ground anchoring system to endure. 

Power = force*velocity, peaking at 16 GW at the middle point of the 256 m long mass driver, a lot more than the average demand. Power consumption averages 1.75 MW per meter, but much of this is recycled. Assuming 3% loss as ohmic heating, we’re still looking at 50 kW per meter, easily enough to justify an active coolant loop. For the purposes of this post, I assume any waste heat pumped and dumped in a permanently shadowed radiator used primarily by the power reactor. Beaming power up from the Earth would require a receiver rectenna several km across. 

Structurally, the entire launch rail must endure tremendous forces, variable loads (due to unbalanced cargo), impacts from micrometeorites and rock spray during rocket landings, electrostatic and thermal nonsense caused by the Lunar day night cycle, and lots of other things. Fortunately, it has no moving parts and no wear surfaces!

So while the mass driver could be brought to the Moon over several launches in pieces, it needs a sophisticated anchoring system to react out loads and cope with extreme thermal swings during the Lunar day/night cycle. 

What do you think, Claude?

In which I roast myself with the latest AI so you don’t have to. Also some more technical detail. 

Missing or under-treated

1. Sled recycle losses. You bury the second 128 m of deceleration in “recover some momentum as power.” Empty sled at 1.6 km/s = 1.02 GJ. At 1.56 Hz and 90% recovery, that’s 160 MW of pure dissipation, not in your 450 MW budget. At 80% it’s 320 MW. You either need to argue >95% round-trip electrical recovery (hard given switching losses on multi-GW pulses), or your reactor is closer to 700–800 MW. [CH: Correct, there’s a very steep penalty for higher parasitic power losses.]
2. Pulsed-power infrastructure. 16 GW peaks at 1.5 Hz with bidirectional flow. The capacitor banks / flywheels for this dominate the system mass and probably the cost. ITER’s pulsed power is tens of GW for comparable durations and it’s a building. On the Moon, this is the actual hard engineering problem, more so than the rail itself. Worth a paragraph. [CH: There are several ways to skin this cat.]
3. Catcher architecture. 56,000 launches/day into LLO requires tug coordination at obscene tempo. Velocity dispersion at exit sets apogee error: 0.1% velocity error → ±5 km position error after a half-orbit. Either you hit launch precision <10⁻⁴ (very hard with thermal drift on a 256 m structure) or your catcher needs km-scale capture and active pursuit. The catcher fleet plausibly outweighs and outcosts the driver. [CH: I have also waved my hands about what to do with rocks in lunar orbit – they still need to be converted into solar computronium somehow!]
4. Δv beyond LLO. Customer is in Earth orbit, not lunar orbit. LLO→GEO ≈ 3.9 km/s, LLO→LEO higher with aerocapture. You save ~9 km/s vs Earth launch but you’re not at the customer. The relevant comparison is (launch + tug from LLO) vs (Earth launch), not (launch) vs (Earth launch). Worth one paragraph stating the tug budget. [CH: As stated, the assumption is that Moon launch delta V is the expensive part we’re trying to overcome.]
5. Chemical-lander baseline. A dumb LOX/LH2 (or LOX/CH4) lander from the Moon to LLO needs ~1.8 km/s, mass ratio ~1.7. If you have any propellant production at all, chemical wins until volume is huge. Mass driver’s case rests on (a) no propellant feedstock available at scale, or (b) volume past the crossover. State this; the reader will ask. [CH: There’s no good reason to synthesize prop on the Moon and the gear ratio makes flying chemically propelled shuttles to and from LLO pretty silly.]
6. Reactor radiator. 450–800 MW thermal, polar shadow, ~700 K reject → tens of thousands of m². Your 2025 Mars-reactor post implies a tonne/m² class radiator. This is hundreds of tonnes of radiator alone, which dwarfs a 1000-T terrestrial reactor mass. The 10× cost premium claim wants more support. [CH: True, as a good rule of thumb, a space reactor’s radiator will weigh about as much as everything else put together.]
7. Magnet survivability at 1000g. NdFeB is brittle. Sintered blocks 9 m long under 1000g shear and oscillating tension fatigue is not obviously feasible. Either bonded magnets (lower B), or segmented with structural steel, which eats into your 70% magnet mass fraction. [CH: Actually building magnets and a sled structure which can make this work is serious engineering.]

Alternatives

The Moon’s lack of an atmosphere means you can also operate a lunar sling. A tower with extendable tethers whirls around to obtain high speeds without as high g loads. I see no reason why entrepreneurs couldn’t build a range of lunar launching solutions!

Conclusion

Lunar mass drivers do not violate the laws of physics. Their tech can be developed and tested thoroughly in labs on Earth at prices that are reasonable in comparison to the complexity and cost of Starship development. They are unlikely to be able to compete with Starship flying at any level of volume from Earth (where the chip fabs are) to space (where the infinite sunlight is), unless Earth launch is supply limited in some way. Order of magnitude, this is beyond 100 TW per year deployment. For reference, 10 TW of compute is roughly equal to the total economic output of the entire world’s supply of natural intelligences. We live in interesting times!

This post has benefited from generous contributions from Austin Vernon, Malcolm Davis, Sam D’Amico, and others who do not necessarily endorse the conclusions of this piece. I will follow Australian spelling conventions in this piece. 

[Edit: This piece was quoted by Eric Crampton in The Post on April 27, 2026.]

---

I’ve written a few pieces over the years about Australian energy and economic policy, and now I’m dipping my toes into defence. The usual disclaimers apply; this post represents my opinions only. I’m a dual US/Australia citizen, resident in Los Angeles, where I founded and run Terraform Industries, a solar synthetic fuel tech startup. I write this as someone who is familiar with the 2023 Defence Strategic Review, the 2024 and 2026 National Defence Strategies, and the 2024 and 2026 Integrated Investment Programs (IIP). While each successive document diagnoses the strategic environment with increasing accuracy it allocates capital against a threat model that no longer exists.

Australia spends nearly AUD $60B/year (US $40B/year) on defence, about 2% of GDP. The 2026 National Defence Strategy commits to raising this to 3% by 2033 and allocates $425B over the coming decade, but increased expenditure won’t buy increased security if it’s spent on weapons that are no longer determinative.

To avoid burying the lead, it is my contention that Australia should be able to achieve far more bang for its buck running competitive, lean, mean domestic weapons development programs focused on the newly demonstrated and increasingly determinative drones-of-various-kinds platforms. This is in stark contrast to the existing practice of spending >90% of acquisitions funding on foreign weapons platforms and much of the remainder on outrageously expensive failed development programs, such as the Hunter class anti-submarine frigate. This ship, which will spend eight years under construction until first launch in 2032, was based on a British design and will end up costing over $7b, per hull! 

Let’s contrast this to the US, hardly a paragon of lean defense spending. $7b would buy you a Nimitz-class nuclear aircraft carrier – including the planes. It would buy you the complete ground up development costs of nuclear attack submarines, with enough left over to pilot ballistic missile submarines and develop the original trident missile. It would buy you the complete ground up development of Falcon 9 and Starlink. That’s just for a single Hunter-class hull – Australia is currently planning to build six!

Why maintain a defence force at all? 

The primary purpose is to maintain Australia’s capacity for peaceful self determination indefinitely into the future. That is, absolute national sovereignty requires Australia to maintain the ability to secure its territory, deter hostile aggression, and control its destiny.

Without putting too fine a point on it, the purpose of the Australian Defence Force is to ensure that the calamity that befell Ukraine can never occur to Australia. While focused on Australia, this analysis could also be translated with minimal changes to other middle powers.

Australia is rich (US$69,360 GDP per capita), populous (28m + 1m diaspora), educated, peaceful, democratic, liberal, and financially stable. It has enormous territory and unmatched natural resources. It has much that foreign powers covet through either open or covert warfare. 

Australia’s current GDP exceeds that of every WW2 belligerent in 1940, even the US. It exceeds the combined economic power of all the Axis powers. And yet somehow, in 1940, in a world before computers, reliable diesel engines, modern healthcare, all the WW2 powers were able to churn out planes, tanks, ships, submarines, and munitions. The US produced nearly 300,000 planes. Even countries with terrible climates that are still poverty stricken in 2026, like Russia, were able to produce 160,000 aircraft back in the early 1940s. 

In contrast, wealthy, modern Australia was able to assemble 73 F-18 fighters from mostly imported components between 1984 and 1990, and nothing for the last 36 years.

This is a festering problem that is now inviting catastrophe. 

What the strategy documents say

The Albanese government’s position on Australian defence rests on three documents released in three consecutive biennial cycles.

The 2023 Defence Strategic Review (Smith / Houston) concluded that the ADF is “not fully fit for purpose” and that the historical ten-year warning window before major conflict no longer applies. It accepted 105 of 108 classified recommendations and identified six priorities: AUKUS nuclear-powered submarines, long-range strike and domestic munitions manufacture, northern base hardening, workforce retention, rapid translation of disruptive technology into ADF capability, and deepening Indo-Pacific partnerships.

The 2024 National Defence Strategy introduced two doctrinal inventions that now anchor Australian defence planning: National Defence (a whole-of-nation concept spanning maritime, land, air, space, cyber) and the Strategy of Denial (the cornerstone of Defence planning, intended to deter adversaries from projecting power through Australia’s northern approaches). Deterrence was elevated to Australia’s primary strategic defence objective, above the previously co-equal “shape” and “respond.” The accompanying IIP committed $330B through 2033/34.

The 2026 National Defence Strategy, released last week on 16 April 2026, builds on rather than departs from the 2024 framework. It raises the fiscal envelope to $425B over the decade, adds $14B over the forward estimates, and benchmarks defence spending at 3% of GDP by 2033. It lists seven IIP priorities: enhanced undersea warfare via nuclear-powered submarines, accelerated lethal maritime capabilities, expanded long-range strike, integrated air and missile defence (IAMD), expanded autonomous and uncrewed systems, counter-UAS for critical infrastructure, and resilient multi-orbit satellite communications. It claims lessons from Ukraine and the Middle East have been absorbed, and it emphasizes self-reliance, industrial resilience, and civil preparedness.

This is, on the face of it, a reasonable set of documents. The strategic diagnosis is correct. The Strategy of Denial is the right doctrine for a middle power: deterrence-by-punishment and deterrence-by-retaliation are beyond Australia’s resources, as Retired Major General Mick Ryan has noted. The rhetorical emphasis on sovereign industrial base, Ukraine lessons, and autonomous systems reads well.

The problem is the capability acquisition list that actually gets funded.

Why the strategy will fail in execution

The unifying defect of the current program is a mismatch between the rhetorical doctrine (denial, mass, self-reliance, Ukraine lessons) and the capital allocation (foreign-sourced exquisite platforms, long lead times, workforce-intensive crewed systems, single-digit hull counts).

Four critiques from inside the Australian strategic-policy establishment make this point from different angles, and I quote them here to establish that this diagnosis is not idiosyncratic:

• The Lowy Institute (2026) describes NDS 26 as “modest spending, welcome reforms” — a continuation of NDS 24 rather than a departure. It notes that Australian defence policy is developed within a narrow Canberra circle largely insulated from external scrutiny, and that Minister Marles’s dismissal of think-tank and retired-officer input at the National Press Club reflects a structural problem, not a tactical misstep.
• Meanwhile, the same author (retired Major General, CSIS fellow Mick Ryan) remarks on his Substack that AUKUS Pillar 1 plus conventional ADF modernization cannot both be funded at current spending — one will squeeze the other. He also documents that military star-rank officers grew 33% over the past decade while enlisted ranks shrank 1%, which is difficult to reconcile with the claim of a more effective, faster-moving force.
• Sam Roggeveen argues that increased spending may be necessary but buys us little if it only increases co-dependency on US operational capability, undermining the premise of improving self reliance! 
• Asia Pacific Defence Reporter summarized the ADF press release, but comments immediately identified the weaknesses with respect to environmental shifts since 2024 that the new document does not absorb: weakening US alliance commitments, reduced US Asia deployments, two theaters of successful low-cost mass drone warfare, and chronically underfunded British defence investment. Despite all of this, AUKUS Pillar 1 remains the centerpiece of Australian acquisition.

The argument I make below starts from these observations and pushes further.

The electric stack has inverted the offense/defence cost curve

The core technical fact that Australian defence planning has not absorbed is what Packy McCormick and Sam D’Amico call the Electric Slide: the five foundational technologies of the electric stack — motors, batteries, power electronics, sensors, and edge compute — have each decosted by roughly 100× over the past 30 years. The guidance electronics that in 1990 required a government munitions program now ship as the cheapest component in a disposable toy.

The practical result, demonstrated across Ukraine, Nagorno-Karabakh, the Red Sea, the cartel conflicts in northern Mexico, and now the Persian Gulf, is a cost-exchange regime in which a $500–$5,000 drone can plausibly destroy a $1M–$100M asset. Ukraine produced more than 2 million drones in 2024, and doubled that in 2025. Russia’s Black Sea Fleet, which in 2021 was significantly larger than the Royal Australian Navy, has been reduced by approximately 45% by an adversary with no navy at all. The dominant ships were destroyed by autonomous surface vessels and anti-ship missiles at a cost-exchange ratio on the order of 1:1000. Houthi operations have forced US carrier strike groups into standoff. Cartel drones routinely contest Mexican state control in Michoacán and Sinaloa.

This is not a speculative threat. It is the observed empirical base rate of every live conflict since 2020. The chart below shows just how cheaply an adversary can now asymmetrically convert exquisite war fighting systems into scrap metal. 

Australia’s existing force, and the force envisioned by IIP 26, is built to engage a different set of adversaries under a different set of cost-exchange assumptions. If a legacy system is operated as part of an integrated force with robust Integrated Air and Missile Defence, directed-energy weapons, resilient Command and Control, counter-UAS coverage, and updated tactics, it may retain some utility, albeit at enormous expense. The question is whether Australia has any of those enablers at scale, and whether NDS 26 funds them at the rate the threat demands. The honest answer to both is no.

Domain-by-domain

(With apologies for acronym soup, I have done by best to link/summarize/rationalize!)

LAND

SystemIOCUnitsUnit cost (A$)Vulnerability assessmentM1A1 Abrams MBT (tank)200759 (most transferred to Ukraine and subsequently destroyed)~$15MHIGH. Cold War tank optimized for maneuver warfare. Top-attack loitering munitions (Lancet-class), FPV drones with shaped charges, ATGM-armed UAS all lethal. Trophy APS not yet fitted. Weight precludes most regional deployment.AS21 Redback IFV (tank)2025–27129 planned~$27MMOD-HIGH. New Hanwha platform. Better protected than M113AS4 but at 42t still vulnerable to top-attack precision munitions. Active protection to be fitted.Boxer CRV 8×8 (tank)2025–26211~$12MMOD-HIGH. Lance turret 30mm provides some counter-UAS capability. No integrated APS.M113AS4 APC (tank)2007 (upgrade)~340~$3M upgradeHIGH. 1960s aluminium hull. No counter-UAS capability. Replacement overdue.Bushmaster PMV (armored truck) 2005~1,100~$1.5MMOD. Low unit cost makes individual losses tolerable.Hawkei PMV-L (armored truck) 2018~1,100~$2MMOD. Similar profile to Bushmaster.AS9 Huntsman SPH (K9) (howitzer) 202630~$25MHIGH. SPHs are the priority target for counter-battery UAS. Only 30 units — loss of a handful is operationally significant.M777A2 155mm Towed (howitzer)201054~$5MHIGH. Static when firing, slow to displace. Ukraine has proven this type extremely vulnerable.M142 HIMARS (rocket launcher)202542 (+48 ordered)~$8MMOD. Dispersible, shoot-and-scoot. GMLRS/ATACMS/PrSM is the real capability. First domestic GMLRS test-fired at Woomera (April 2026).NASAMS (LAND 19 Ph 7B) (surface to air missile)2025–27TBDTBDLOW (defensive). But: NASAMS is a short-range system based on AMRAAM. Does not close the IAMD gap. NDS 26 revived the MRGBAD program to layer above NASAMS — an implicit admission that the 20+ year Ground Based Air Defense (GBAD) gap has not been closed.MRGBAD (medium-range GBAD)TBDTBDTBDNew in NDS 26. Represents the first funded commitment to medium-range air and missile defence in two decades. Essential but late.ARH Tiger Attack Helo200422~$55MHIGH. Being replaced by Apache. Sustainment availability has been <40%.AH-64E Apache (helicopter)2025–2629~$80MMOD-HIGH. Class faces existential questions post-Ukraine. Standoff missile capability helps.CH-47F Chinook201514~$60MHIGH in contested airspace. Essential for logistics, no self-defence vs precision munitions.UH-60M Black Hawk2025–2740 ordered~$40MMOD-HIGH. Replacing Taipan. Same class vulnerability as all utility helicopters.

AIR

SystemIOCUnitsUnit cost (A$)Vulnerability assessmentF-35A Lightning II201872~$110MLOW-MOD. 5th-gen. Primary vulnerability is basing — small number of northern airfields are targetable and have no IAMD. ALIS/ODIN creates US dependency. AIM-260 JATM (air to air missile) acquisition now confirmed per IIP 26. LRASM (anti ship missile) to be integrated.F/A-18F Super Hornet201024~$90MMOD. 4.5-gen, not survivable vs modern IADS. Useful for standoff strike with JASSM-ER and LRASM (now operational). HACM (Hypersonic Attack Cruise Missile) development under US partnership. Block III upgrade in progress.EA-18G Growler201712~$100MLOW-MOD. Only non-US Growler operator. Extremely high value.E-7A Wedgetail (radar plane)20096~$350MMOD. Vulnerable to long-range AAMs (PL-15 class). Only 6 airframes. The future of this platform lies in MUM/T battle management for autonomous systems, not as a standalone AEW.P-8A Poseidon (sub hunter)201714~$250MMOD. Not survivable in contested airspace. ASW capability essential.C-17A Globemaster III20068~$330MHIGH if forward-deployed. Irreplaceable — 8 airframes, no planned replacement.C-130J Hercules200612 (+ 8 on order per IIP 26)~$130MHIGH if forward-deployed. IIP 26 expands fleet to 20.C-27J Spartan201510~$60MMOD-HIGH. Useful for dispersed archipelagic ops.KC-30A MRTT (tanker plane)20117~$300MMOD. Essential force multiplier. Only 7.MQ-4C Triton2024–254 (+3 planned)~$200MHIGH. Large, slow, non-maneuvering. Not survivable in contested airspace.MQ-28A Ghost Bat2024 (IOT&E)~6 test articles~$30–40MMOD. Points in the right direction, but at current unit cost not attritable by drone-war standards. Unit cost must fall 10–30× and production rate rise by orders of magnitude before Ghost Bat matters at the force level. As a learning lab and a template for evolved CCAs, more valuable.

SEA

SystemIOCUnitsUnit cost (A$)Vulnerability assessmentHobart-class DDG (guided missile destroyer) 20173~$3BMOD. Aegis + SM-2/SM-6/ESSM. Tomahawk recently added. Only 3 hulls. Vulnerable to saturation AShBM/AShCM and USVs.Anzac-class FFH (helicopter frigate)19967~$800M (original)HIGH. 1990s design. Being replaced by SEA 3000 Advanced Mogami-class frigates (11 planned, 10,000 nm range, 32-cell VLS).Hunter-class FFG~2031 (est.)6 planned (down from 9)~$7–8B per hull (not an aircraft carrier!!)Cost blowouts and delays, 32 VLS cells per hull. Entire program yields 192 VLS cells — one US Arleigh Burke Flight III carries 96. ASW-focused. A $45B+ program delivering two destroyers’ worth of missile cells.SEA 3000 Mogami-class2029–3011 planned~$1.5BNew in IIP 26. 32-cell VLS, 10,000 nm range. Japanese design, first 3 Japanese built. Doubles surface combatant fleet over the decade at $52–65B.Arafura-class OPV (offshore patrol)20246 (of 12)~$500MHIGH. OPV, lightly armed — some hulls entering service without main armament fitted. Constabulary only.Canberra-class LHD (baby carrier)20142~$1.5BHIGH. 27,000t amphib, minimal self-defence. Must be heavily escorted.HMAS Choules (LSD)20111~$150M acquiredHIGH. Landing ship, dock. Old, slow, poorly armed.Collins-class SSK (attack submarine)19966~$1.2B originalLOW-MOD. Quiet on battery. Historically poor availability. The most survivable major ADF platform if maintained and crewed.SSN-AUKUS (future)~2040s8 planned~$30–40B per (!)Not operational for 15+ years. Program cost consumes a huge share of envelope.Virginia-class SSN (interim) (nuclear attack sub)~20333 planned~$5–6B purchaseIf delivered on schedule, the most capable platform in ADF inventory. Crewing and basing challenges.Ghost Shark XL-AUV (robot submarine)2024 (prototype)TBD~$40–100MLOW. Anduril autonomous. Potentially game-changing. Scale is the question.

SPACE / CYBER

SystemIOCUnitsCost (A$)Vulnerability assessmentJP 9102 (MILSATCOM)~2027Multi-orbit~$3–4BRedefined in NDS 26 as multi-orbit for resilience. Number of satellites and timeline unspecified.DEF 799 (Space Surveillance)———Effectively disappeared from public reporting over the last two years. Likely quietly cancelled or downscaled. No sovereign space-based ISR.DARC (Deep Space Advanced Radar Capability)Partial—Joint AU/US/UK programTrilateral SSA capability out to GEO. Australia is a partner, not sole operator — co-dependent rather than fully sovereign.Jindalee OTH Radar (JORN)20033 sites~$2.5B totalHIGH. Ground-based, fixed, targetable — but geographically remote. Irreplaceable sovereign capability.REDSPICE (ASD Cyber)2022—~$10B / 10yrOffensive and defensive cyber. Doubles ASD capability. Arguably the best value-for-money investment in the ADF: asymmetric, scalable, sovereign.

In short, almost none of Australia’s current or near future weapon systems are useful against any adversary capable of obtaining the specific kinds of missile or drone warfare systems that are routinely fielded by such poorly funded outfits as second tier Mexican cartels, investigative journalists, Hamas, and Houthi rebels, as we have seen time and time again in Ukraine, Iran, and other places. 

This may seem unbelievable. So unbelievable that, for example, you can measure Russia’s state of denial in their loss of more than 2600 tanks to drone attacks in the last four years. Australia doesn’t have 2600 spare tanks to learn this lesson. We live in the future, and highly capable attack drones are significantly less difficult to build than, say, a motorcycle. 

The structural problem, restated

Australia’s capability posture is built around high-unit-cost, low-count, foreign-sourced exquisite platforms — a force structure appropriate to a world where precision strike was a US/USSR duopoly and tactical mass was a minor consideration. That world is gone. In the world NDS 26 claims to operate in — the post-Ukraine, post-Red Sea, post-Nagorno-Karabakh world — tactical mass is everything, and the cost-exchange regime rewards the side that can produce cheap guided munitions in volume.

Australia produces essentially none of these domestically. Approximately 90% of Australian defence acquisition spending flows offshore, primarily to the US. The 2024 NDS identified this as a problem and committed to sovereign industrial resilience; the 2026 NDS reiterates the commitment and adds some additional funding. Neither document commits to domestic manufacturing at the scale or tempo the threat model requires.

Despite spending $60b/year, foreseeable advances in drone weapons have rendered not only Australia’s legacy defence systems but almost all of its current generation of acquisitions obsolete.

On land we have vehicles, tanks, artillery, attack helicopters, all of which are extremely expensive cannon fodder for drones or guided rockets. As we have seen in Ukraine, when a $1000 drone can take out a $1m tank or $100m aircraft, the asset light combatant has a sharp advantage. 

On sea, we are planning to spend AUD$70b (1.5x the entire Manhattan Project!) on six Hunter Class Frigates that are far too few to defend our maritime borders and just as vulnerable to explosive jet skis as Russia’s black sea fleet, halved in four years by a conventionally far weaker adversary who doesn’t even have a navy, and who extracted casualties with a 1:1000 cost exchange ratio. Australia is rich but not that rich.

The only systems which are not floating boxes of explosives and sailors visible from space are the submarines, consisting of the rapidly aging Collins fleet and the AUKUS submarines that are as expensive and foreign as they are far off.

In air, Australia has about 100 foreign-built fighters. Which airforce are they built to engage? Australia doesn’t have a regional peer adversary who is going to slug it out fighter to fighter. 100 F-18s and F-35s could hardly stand up to Chinese air power and represent high value targets (particularly when on the ground) to irregular insurgent/proxy/asymmetrical combatants, against which they have struggled to engage in similar conflicts elsewhere. Israel has plenty of jets but they were not particularly useful in stopping rocket attacks from Gaza.

In all cases, this is hardware built for fighting yesterday’s wars. The electric stack has de-costed by a factor of 100-1000, putting sharply asymmetrical threats directly into the fight. In WW2, the Allies were ultimately able to achieve crushing air superiority and then destruction of enemy energy and transport infrastructure via the saturation Combined Bombing Offensive. Eighty years later, wars are once again decided by materiel production capacity, only adversaries can easily field 100 times as many aircraft and operate them with software rather than trained pilots. 

Despite spending nearly AUD$60b/year, Australia is functionally undefended and undefendable. The unstated premise of this conversation is whether Australia’s military is powerful enough to deter, that is, to compel the resolution of disputes through diplomatic channels, with strong military powers like China. But this isn’t how most wars are fought anymore. 

Instead, adversarial powers prefer to act with a winking deniability between networks of proxies, committing a litany of sub-threshold hybrid outrages that are calibrated to fall short of an open declaration of war while doing everything possible to degrade their opponents. 

Grey-zone threats the strategy underweights

The documents acknowledge “hybrid threats” but the IIP does not allocate against them at scale. The actual vectors being used against Australia and similar middle powers today include:

• Submarine cable cuts and seabed infrastructure sabotage, as seen in the Baltic and around Taiwan.
• Industrial espionage and IP theft, documented extensively across Australian universities and critical-tech companies.
• Cyber operations against critical infrastructure (Optus, Medibank, DP World) with state attribution in several cases.
• Covert police operations and diaspora surveillance, documented against the Chinese overseas security apparatus.
• Deliberate information operations across social platforms, especially during elections and referendums.
• Deliberate introduction of biological disease vectors and operation of domestic clandestine laboratories with foreign links and funding.
• Predatory dumping of both legitimate and contraband products to degrade domestic industry and social cohesion.
• Mysterious degradation of critical infrastructure. 

A further strategic problem: any adversary can use Australia’s much smaller, closer neighbours, or enormous swaths of uninhabited sovereign territory, or even existing infrastructure, as covert staging grounds. In a world where adversaries now routinely preposition containers of offensive drones within a mile of their target, how are Papua New Guinea, the Solomons, parts of Indonesia, and the broader Pacific arc, let alone the 99% of Australia with no significant human presence, meant to counter this threat? A few shipping containers of Shaheds or lightly modified commercial drones, released by a non-state proxy, would be sufficient to regionally neutralize Australian military power, with no warning, no formal declaration of war. This is not speculative. Hezbollah has conducted Shahed operations against Israel from Iraq and Yemen. Houthi operations have targeted commercial shipping and US naval assets at 1000+ km standoff. Wagner / Africa Corps conducts similar operations across the Sahel. The base rate of proxy-staged drone warfare is not hypothetical, rather, it is the modal form of contemporary conflict.

In 1943, Australian commandos were able to infiltrate Japan-occupied Singapore in a disguised fishing boat, sinking or damaging six ships. This sort of raid could only be done once, at enormous risk and limited impact. Today, a similar raid could be executed by a competent high school robotics team with autonomous boats, or deployed at literally 1000x the scale and 1000x less cost by a well-resourced military. 

What the 2026 IIP does not buy, and must

In the following, I advocate strongly for domestic production capacity. This does not mean autarky. It means possessing enough of the stack that either the remaining supply chains are highly fungible, or Australia has enough aggregate industrial power to earn a seat at the table when supply is constrained. 

(I drafted this post before the latest conflict in Iran. Then, I worried that Australia lacking bargaining power in a resource constrained environment would be hard to motivate. Now, with Australian agriculture on the verge of diesel starvation, it is only too obvious.)

None of what follows is technically difficult relative to Australian economic and technological capacity. The Soviet Union delivered most of it in the 1950s, at a fraction of Australia’s current real GDP per capita. Each program below could be operated for under AUD $1B/year — well within the existing defence envelope. 

Indeed, programs fail more often from over funding causing indigestion than underfunding. Committing to tight budgets and schedules is the key to success.

Must-have domestic capability

Orbital launch capability. SpaceX developed Falcon 1 in five years for under $100M. Rocket Lab repeated the feat with Electron a decade later. Gilmour Space Technology in Queensland is developing the Eris vehicle — the first Australian-developed orbital-class rocket — and represents the nascent beginning of sovereign launch. It is privately funded, chronically under-supported by the government, and has not yet reached orbit. Compare Rocket Lab’s trajectory, which benefited from early NZ government partnership and US customer access. Without sovereign launch, Australia lives under a sky controlled by others.

Domestic comsat, spysat, radar sat, and GNSS capability. SpaceX ships some of the most advanced satellites ever built, for less than $1M per unit. The US, China, India, Russia, Europe, and Japan each operate sovereign GNSS constellations. Planet Labs, a private company, operates hundreds of imaging satellites and offers sub-metre resolution. Umbra operates a SAR constellation resolving better than 1 m in any weather. The ADF buys or requests access to intelligence that any civilian with a credit card can purchase — and has no sovereign path to independent ISR because DEF 799 has effectively disappeared from the public IIP over the last two years. JP 9102 has been redefined to multi-orbit per NDS 26, but the number of satellites and timeline remain unspecified. Through DARC, Australia is a partner in deep-space SSA — genuinely useful, but co-dependent with the US and UK, not sovereign.

One million drones per month manufacturing capacity, with local supply chains and/or stockpiles sufficient for more than a year of sustained conflict. Ukraine produced over 2 million drones in 2024 and 4 million in 2025 — a wartime economy, certainly, but Ukraine’s pre-war GDP was a quarter of Australia’s. Australia has a strong drone innovation community (the current world drone speed record is held by an Australian hobbyist) but no production base worth the name.

Drones require motors, structures, batteries, power electronics, and controllers. Most of these parts can be mass-produced with startup capital in the ~$1B range per category. Standing up semiconductor fab capacity for controllers, MEMS sensors, and CMOS cameras at drone-adequate nodes would cost approximately $5B. The first step is supplier relationships and stockpiling; the second is a land-and-expand domestic fab strategy with enough strategic ambiguity about actual capacity that defence saturation or blockade are unacceptably risky for an adversary. Either Australia has the ability to produce its own industrial controllers, or it ends up — as Russia has — cannibalising white goods for guided-munition chips. Fabricating 1980s-era 8086-class processors is not technically difficult; those nodes are adequate for the vast majority of drone applications. But… Australia bulldozed its only functional fab in 2021 to extend the Sydney Metro.

Austin Vernon has described the necessary fleet architecture in his drone airforce essay. The point is that the key components are the same across drone types, which means that a single industrial base serves the entire mission set.

Missile defence. Israel, the US, and other allies have operationalised layered missile defence systems. What were infeasible science projects in the 1960s and unreliable demonstrators in the 1990s are now mature enough to shift deterrence calculus — not to guarantee leak-proof defence, which is physically difficult against a capable adversary, but to raise the cost of strike sufficiently to change the adversary’s planning. A system that intercepts more than 50–70% of incoming threats is a strategic asset; it does not need to be perfect to deter. Australia’s NDS 26 MRGBAD acquisition is the first real step in two decades, but it is a purchased point-defence capability, not a sovereign system. Building a domestic layered IAMD would cost a fraction of AUKUS and deliver deterrent value on a much shorter timescale. 

There are multiple rocket hobbyists on YouTube currently building rockets that, with some extra-legal tweaks, would be capable of missile defence. AI coding agents are more than capable of writing the entire software stack in an afternoon. Australia does not need to spend billions of dollars and wait decades to purchase this capability from foreign nations. 

Energy independence.

• Australia has oil and tight shale but has allowed domestic refining capacity to collapse. Two refineries are not enough. This is re-discovered every few years, including the current gulf crisis.
• Australia invented the modern solar module and then actively exported the technology to China. It is now time to play an active role in the production of the materials and technology that power the electric stack. Building GW-scale solar and inviting the world to smelt their aluminium and other metals in Australia is a path to both wealth and strategic weight.

Submarines.

• Surface naval assets are increasingly vulnerable against any adversary. Conventional wisdom maintains that surface ships can still function inside an integrated force given sufficient organic integrated air/missile defence (IAMD), electronic warfare (EW), counter unmanned aerial system (C-UAS), and coalition coverage but it is not the empirical pattern of recent conflicts. The Black Sea Fleet, operating inside its own anti-access/area denial (A2/AD) bubble with the full range of Russian EW and air cover, was reduced 45% by an adversary without a navy. The Moskva and Sevastopol strikes were the headline cases; the attritional campaign against patrol craft, intelligence ships, and the Kerch Bridge was continuous. Surface ships can be made more survivable. They cannot currently be made cheap enough for an attrition regime.
• Submarines retain genuine utility. They are used for anti-shipping, anti-submarine, special-forces insertion, strategic chokepoint interdiction, electronic intelligence, and missile launch. Crewed submarines perform all these missions, but autonomous submarines can perform many of them, and the technology is trending there — Australia’s Ghost Shark program is a leading example.
• Submarines require air-independent propulsion. The Collins class has demonstrated that Australia can operate large long-range diesel-electric submarines for decades, though transit speed and time on station are so limited that coverage of Australia’s top four strategic chokepoints would require a fleet of over one hundred submarines of this type. 
• The US Navy developed the first nuclear submarine powerplant in 1173 days for approximately $2.5B in current dollars, including two hulls and the adjacent unsuccessful sodium-cooled reactor work, as well as all the start-up costs associated with building the first ever nuclear power reactors, such as developing a supply chain for hafnium and vanadium from scratch. Seven decades of design heritage and modern computational tools later, Australia is being told that buying foreign-designed and mostly foreign-built nuclear submarine powerplants will cost $30–40B over 15+ years. Why do the error bars on this purchase exceed by a large factor the real world cost and time for long dead pioneers to do it for the first time ever? Brazil is developing sovereign nuclear submarine technology. A domestic Australian submarine powerplant effort, capped at $1B and four years, is technically feasible. Whether it is politically feasible is a separate question.

Must-have domestic participation

AI sovereignty. The near future will run on CCP AI or US AI. Choose wisely.

• Australia must retain the ability to make material contributions to US frontier AI, and therefore to derive special benefits from it.
• Australia does not need to run a nationalized frontier model or replicate TSMC. It does need to contribute enough of the stack to stay in the conversation.
• Australian expats are well-represented at every frontier AI company. A “Federation Fellowship” structured to repatriate senior talent on favourable terms could help, but would need significant follow through.
• Australia is geographically ideal for large solar-powered AI datacenters, but would need legal reform to bring Australian fair-use and training-data rules into alignment with US practice.

Nuclear weapons. This is the most politically fraught recommendation in this document, and I want to state it precisely.

The technical argument: producing a nuclear weapon is not especially hard given baseline industrial capacity. The 1964 Nth Country Experiment showed that three physics PhDs with no classified access produced a workable weapon design in 2.5 years.

The strategic argument: the 1994 Budapest Memorandum, under which Ukraine relinquished its nuclear arsenal in exchange for security “guarantees”, has become the canonical example of why middle powers without independent deterrents are structurally vulnerable. Having a nuclear deterrent guarantees a baseline of absolute national sovereignty that no alliance commitment can replicate.

The political and industrial argument: Australia has a tiny civil nuclear industry, no enrichment or separation capability, no testing infrastructure, and — at present — no political coalition willing to sustain the investment. That said, North Korea is hardly an economic powerhouse and was able to build plutonium weapons despite determined interference by its adversaries. The NPT and various regional agreements would need to be renegotiated, exited, or ignored.

The point of listing nuclear weapons here is to identify an asymmetry: the technical and industrial obstacles are self-imposed and reversible on a sub-decade timescale given political will. The political constraint is the binding one. If Australian political will to sustain sovereign security catches up to the realities of the post-unipolar era, the nuclear question will be on the table, and the preparation — civil nuclear industry, enrichment-adjacent capability, professional workforce — should begin now regardless.

Having a nuclear deterrent guarantees absolute national sovereignty. After WW1 and WW2, England and France did not hesitate for an instant to ensure they could never again suffer outrages from the industrial might of Germany and later, the Soviets. It is better to have it and not need it than to think “she’ll be right mate” and bequeath eternal slavery and damnation to your descendents.

Conclusion

None of the capabilities above are hard to build relative to Australian economic and technological capacity. The Soviet Union did them in the 1950s. Each program could be operated for under AUD $1B/year — a rounding error inside the existing defence budget, and less than a tenth of what AUKUS Pillar 1 is projected to consume annually by the late 2030s.

Unlike the current acquisition pattern, which sends most capital offshore in exchange for indefinite dependence on foreign industrial complexes for maintenance and support, a domestic weapons-platform development policy accumulates research and production expertise within Australia, where its value appreciates over time, including in the civilian economy. Expat Australians work at every frontier technology company on earth. Australian hobbyists hold world records in the relevant disciplines. The constraint has never been talent, capital, or technology.

The 2023 Defence Strategic Review diagnosed the threat environment accurately. The 2024 and 2026 National Defence Strategies identified the right doctrine — National Defence, Strategy of Denial, self-reliance, industrial resilience. The 2026 Integrated Investment Program commits $425B against these priorities and in doing so reaffirms a platform-centric acquisition model that the post-Ukraine evidence base does not support.

Australia is not undefendable in principle, but it is undefendable against the threat model the documents themselves describe, using the capabilities the documents themselves fund.

The Strategy of Denial is the right strategy. The question is whether the capability program actually delivers denial against a cost-exchange ratio of 1:1000 in favour of the adversary. 

The answer to that question, visible in the IIP line items, is obviously not. That has to change.

Australia’s future national sovereignty could topple at any moment, not through conventional conquest but through the slow attrition of deterrent credibility that invites exactly the kind of sub-threshold coercion the Strategy of Denial is evidently unable to prevent. 

If Australia fails to aggressively correct course toward domestic defence tech production immediately — not the next biennial National Defence Strategy cycle, right now — the last vestiges of its existence as an independent political entity will soon vanish entirely. 

Australia’s Fiscal Point of No Return

Last September I wrote about Australian economic stagnation — how taxation past the Laffer curve’s point of peak growth, zero public sector productivity improvement since 2001, and the systematic eradication of manufacturing and ICT from the economy have combined to produce economic stagnation and flatlining living standards. The NDIS and similar social spending were the only parts of the economy “growing,” but because this spending is not cumulatively generative, it’s not able to sustain long term high quality public benefits. In other words, we’re burning the seed corn.

[Edit: This post generated a surprising amount of coverage, including on news.com.au, The Australian, the Daily Telegraph. I’m pleased that the problem is now obviously undeniable. The question is what to do about it.]

The post generated a question I couldn’t answer at the time: when exactly does this become irreversible? During my recent visit in December 2025, I paid close attention to the function (and dysfunction) of the service economy and was forced to the uncomfortable realization that government spending already controls and distorts such a large fraction of the market that it is not clear whether the Hayek knowledge problem is being solved at all.

In other words, the economy is producing gross shortages and excesses of many goods because government interference in natural pricing mechanisms is so excessive that information transfer through prices no longer exists in a functional way. This is clear in the case of housing, education, healthcare, and now fuel, but also occurs almost invisibly in nearly every other economic edge node that Australians routinely use.

This motivated me to investigate the situation more thoroughly. The answer is not unexpected but quite dire. Australia passed the point of no return in 2013, and is now 13 years into a probably irreversible zombie-fication of its economy.

[Edit May 2026: I have had a few questions about how Australia’s future command economy would reveal itself. Here’s a rough sketch. First, some part of the “basket of goods” will have economically or electorally unacceptable levels of unaffordability. Usually this occurs in the trifecta of housing, healthcare, or education, and is probably driven by housing unaffordability for reasons beyond the scope of this discussion. In any case, you’ll never see gnashing of teeth over, say, scarcity and high prices in the case of consumer electronics, because in these sectors a lack of overregulation and a robust private manufacturing base has driven constant improvements. For example, a top of the line TV is now cheaper than wall it covers! Or a computer is cheaper than the furniture it sits on.

In response to high prices, politicians will generally want to be seen to “do something” and increasing supply through, eg, deregulation or better industrial policy is not direct enough. So first they will announce a subsidy. This could be a first home buyers grant, rent support, direct transfer payments, benefits of various kinds, etc.

Subsidizing demand, of course, does nothing to increase supply if supply is artificially constricted by regulation or some other source of scarcity, so this only has the effect of driving up prices while simultaneously draining the public purse.

The next step, when the cost of direct subsidy is too much of a drain on the tax base but, in the usual way, politically impossible to roll back, is to reach for price controls. Some fundamental necessity is too expensive and subsidies haven’t solved the problem? Simply ban any vendor from providing the service for a price deemed unacceptably high by an increasingly powerful arm of economic control. Of course, price controls just exacerbate a lack of supply because they disincentivize suppliers from entering the market. This has occurred in housing (rent control), construction (anti-gouging provisions after natural disasters), health care (single payer pricing), etc.

When this predictably fails to solve the problem and repeated intrusions into the healthy function of a transparent and mostly free market have destroyed a previously vibrant industry, the government is left with no other options than to reach further into the autocratic closet and start nationalization. In Australia, most healthcare and education is already nationalized, housing has a strong government component, and the current discussion is around controlling price and nationalizing the few private healthcare operators who provide expert specialist services outside or adjacent to the public system.

The correct thing to do, however, is to back off and liberalize normal economic functions. If housing is too expensive, make it easier to build new houses – don’t just try to re-allocate scarcity.]

What GDP Actually Measures

Australia’s headline GDP is $2.75 trillion. It grew 1.3% in the year to March 2025. GDP per capita fell, for the ninth time in the last eleven quarters. The standard interpretation is that we’re in a “per capita recession” but the total economy is still growing. This is wrong in a way that matters.

GDP doesn’t distinguish between a dollar spent building a factory (i.e. becoming long term productive capital) and a dollar spent on a plan manager coordinating a support coordinator to arrange, say, an NDIS claim. Both show up the same. But only one of them leads to capital accumulation and becomes an engine of generational wealth growth.

Both the private economy and the government spend on a range of things with a range of long term productivities, but as a de facto welfare monopoly and monopsony with the unique power of taxation, it is unavoidable that the government appropriates and ultimately compromises and destroys a lot of productive capital. Well and good, provided the economy can regenerate it at least as fast as it is harvested. Is this the case, remembering that the goal is to shear the sheep, not skin them!

Let’s break down Australia’s economy by public and private spending.

Strip everything tax-funded out of GDP. Government healthcare. NDIS. Aged care. Welfare. Defence. Public administration. Debt interest. To be complete, we need to include the nominally-private sector that exists solely to bill the government. NDIS plan managers. Private hospitals billing Medicare. The entire private health insurance industry, a $25B/year regulatory artifact that adds 15% administrative overhead to redirect money from taxpayers through insurers back to the same hospitals. What’s left is the wealth-generating economy that compounds over time.

That tax-funded economy is about 34% of GDP and growing at 4–5.5% real, despite flat productivity since 2001. The productive remainder is ~66% and growing at maybe 1.0–1.5%, generously. Decompose per-capita GDP to exclude the care economy’s above-GDP growth, and the productive economy per Australian has been shrinking since roughly 2016. Not stagnating. Shrinking.

What’s propping up headline GDP? Two things. First, mass immigration, keeping total population growing and consumption increasing. Second, government spending on itself, the public sector’s 5%+ growth rate dragging up the average. But neither of these factors actually supports long term economic growth or increases the growth rate. Mass immigration dilutes wealth and increases strain on public sector services that lack a market mechanism to respond to demand or to innovative on productivity. And public expenditure on welfare and services, while serving laudable social goals, consumes rather than produces net new wealth.

Communism Via the Back Door

Here’s the part that made me uncomfortable.

If the only “growing” part of the economy is government-funded services, and government-funded services are by definition centrally planned — administered prices, administered eligibility, administered supply — then what we’re watching is the gradual adoption of a command economy through the back door. Nobody voted for communism. But every year the tax-funded share of GDP ratchets up another fraction of a percent, another tranche of the economy moves from price-signal allocation to bureaucratic allocation, and another cohort of workers shifts from producing things people voluntarily pay for to producing things the government has decided they should have. We are sleepwalking into a planned economy, funded by the shrinking productive sector that hasn’t yet noticed it’s being eaten alive. The Fabians would be thrilled. The rest of us should be terrified.

This isn’t hyperbole. The NDIS is already closer to central planning than to a market. The government defines eligible populations, approved services, price caps, and quality standards, then funds everything through a single agency. Medicare is a monopsony with administered prices. Aged care is shifting to government-set pricing through the Independent Health and Aged Care Pricing Authority. At some point the “private provider” wrapper becomes pure overhead on what is functionally a state employment and service delivery system. Australia won’t choose a command economy. It will discover it’s already in one.

The parallels are even stronger. In a capitalist economy, workers are free to move to work in different fields for different employers, they’re free to critique (productively or otherwise) their industrial sector, and generally enjoy liberty and the freedoms we take for granted. In a communist economy, a government bureaucratic official decides what you are paid, what you do, where and when you do it, how you can do it, controls access to clients/customers, controls licensing, insurance, and explicitly prohibits any form of criticism internally or externally, lest it harm public trust in the service. Government policy on provisioning and rationing of inherently scarce public services is routinely updated, but you’ll never hear a word of commentary or constructive criticism from the actual people who work in those sectors, which is strange until you realize that some unelected unaccountable bureaucrat can simply revoke your right to work in your profession with no due process, depriving you of your livelihood. When any society or sector moves to crush dissent, it destroys the feedback system necessary for the most qualified and motivated people to advocate for reforms that help everyone.

And here’s the kicker: a competent command economy would at least be cheaper. Strip out the plan managers, the support coordinators, the LACs, the NDIA assessors, the AAT appeal lawyers, the compliance officers, the fraud investigators — the entire intermediation layer consuming 15–20% of NDIS costs — and just employ the support workers directly. You’d lose consumer choice. You’d save $7–9B a year on the NDIS alone. The current system manages to combine the allocative inefficiency of central planning with the transaction costs of market intermediation. We have the worst of both worlds, and nobody seems to have noticed.

When Did We Pass the Point of No Return?

Four key dates matter. At each of these times, Australia locked in a different dimension of the trap.

2007–08 (Structural). Productivity growth permanently downshifted from ~1.5% to ~1.0% after the mining capex boom ended. From here, the productive economy’s growth rate dropped below the care economy’s structural growth rate and never recovered. The share of GDP consumed by tax-funded services started rising monotonically and hasn’t stopped. The mining boom had been masking this for years. Once the mask came off, the underlying rot was already advanced.

2013 (Political). The NDIS was legislated with bipartisan support. This created a fourth uncapped, demand-driven entitlement — joining Medicare, PBS, and the Age Pension — with no fiscal cap, no means testing, and no GDP-linked growth constraint. Originally scoped for ~410,000 people. Now at ~740,000. Projected to hit a million by 2034. The scheme costs $46 billion this year and is growing at 8–12% per annum. The 2023 Intergenerational Report projected NDIS costs reaching 6.3% of GDP by 2062 without reform — higher than projected health spending over the same period.

The issue isn’t that the NDIS is a magnet for fraud, though of course there is some and of course the optimal amount of fraud is non-zero. NDIS currently spends about $12b a year on administration, part of which is intended to limit fraud, while Bill Shorten called estimates of $2 billion per year in fraud “egregious.” The Grattan Institute found that projected fraud savings over three years total $424 million — less than 0.3% of scheme expenses. Fraud isn’t the core problem. The core problem is uncapped demand in a system without effective eligibility gatekeeping creating incentives for large swaths of society to see themselves as permanently disabled instead of pursuing the dignity of meaningful and sustained contributions to our society despite the depredations of time and fortune that will ultimately take us all.

2013 was the point of no return. The only variable was speed.

2015–17 (Fiscal). The per-worker extraction rate — tax-funded economy divided by productive economy, weighted for working-age population share — permanently exceeded per-worker productivity growth. Each productive worker started falling further behind every year, and hasn’t stopped. By 2024 the extraction ratio hit roughly 66%: for every dollar of GDP the productive economy generates, 66 cents goes to fund the tax-funded economy. That number is heading for 80% by the mid-2030s and 100%+ by 2040 — a mathematical impossibility resolved only by unbounded debt growth, money-printing, or collapse.

~2024 (Democratic). Healthcare workers (~2M), NDIS participants and families (~2.2M), pensioners (~2.6M), aged care recipients (~1.3M), and public servants (~2.1M), after overlap adjustment, add up to about 50% of the voting population. Once a majority of voters depend on the tax-funded economy, no democratic government can run on structural reform. Every election becomes a bidding war to expand public spending. This is the Italian/Greek path, except without the EU backstop. Meanwhile, the once vibrant and vital productive engine of economic growth through primary and secondary production is now a strip mined, abject minority of the economy, colonized from within, extracted until nothing remains.

Demographics: The Locked-In Accelerant

The model gets worse when you add demographics, because the previous numbers assume a constant working-age share, but this is not the case.

Australia’s total fertility rate hit a record low of 1.48 in 2024. Below replacement since 1976 — fifty years running. Each generation is roughly 30% smaller than needed to replace itself. The working-age share of the population falls from 64.5% today to ~57% by 2050. The 85+ population doubles by 2042, and an 85-year-old costs roughly 9x per capita in healthcare versus a working-age adult. That’s assuming that healthcare costs don’t continue to increase.

Fewer workers. Exponentially more expensive dependents. The denominator shrinks while the numerator accelerates. Every threshold in the model shifts 3–6 years earlier once you account for this. The per-worker burden reaches 80% around 2036 and exceeds 100% before 2040.

The 2040s tax base was determined by births happening (or rather, not happening) right now. Even if fertility magically recovered to 2.1 tomorrow, those children don’t enter the workforce until 2045, but the fiscal math has been terminal since 2013 and workers who entered the workforce in 2013 will be nearing retirement by 2045, if retirement is a thing our society will be able to afford then.

Immigration doesn’t fix this, it just masks it. Median immigrant income is $45,351 versus $52,338 for the whole population — a 13% discount. More importantly, 30–40% of migrant labor goes directly into the tax-funded economy: healthcare, aged care, NDIS, education, public administration. Another ~20% into low-productivity domestic services that exist to service population growth itself — circular GDP. No more than 40–50% enters genuinely productive sectors. Counting an imported NDIS support worker’s wages as “productive immigration” is double-counting: the government funds the NDIS plan, the plan pays the worker, the worker’s income shows up as GDP and immigration statistics as economic contribution. We are importing today’s workers and tomorrow’s dependents, and congratulating ourselves on the headline GDP number, but doing nothing to ensure Australia’s long term economic productivity.

The Feedback Loop

Government expands spending → shows up as GDP → masks private sector contraction → government sector competes for labor → wages bid up in healthcare/NDIS → crowds out productive sectors → tax base erodes → more government spending required → repeat.

Meanwhile the government suppresses provider prices through Medicare fee schedules while simultaneously creating insatiable demand through NDIS and aged care entitlements. The gap shows up as wait times, workforce burnout, quality degradation, cost-shifting to emergency departments, and the steady monotonic ratcheting up of government control over more and more aspects of healthcare without ever noticing that control destroys the dynamism which actually delivers cost reduction and service improvement. This dynamic does not show up directly in GDP.

This is Baumol’s cost disease fused with a monopsony that prevents price signals from doing their job, inside a democracy where the beneficiaries now constitute a voting majority. Every year the productive sector that funds the machine gets a little smaller, and the machine gets a little bigger, and the people who run the machine tell us the solution is to make it bigger still.

How Bad Compared to the US?

The natural comparison is the United States, which has its own healthcare cost disease. The US passed its structural point of no return earlier (2003 — Medicare Part D created ~$8T in unfunded liabilities with no revenue source) and its political point later (2010 — ACA). US healthcare spending hit $5.3 trillion in 2024 — 18% of GDP and growing at 5.8% per year versus 4.3% GDP growth. Administrative overhead alone runs $800B–1.1T/year, 3–4% of GDP. More than most countries spend on healthcare total. The US healthcare system is already past the point where a straight command economy — an NHS — would be more efficient for the bottom 80% of the population. That’s a remarkable thing to be able to say and it’s not wrong.

But the US has two things Australia doesn’t. The reserve currency, which lets it run 6%+ of GDP deficits more or less indefinitely. And a productive economy large and dynamic enough — tech, energy, finance, defence — to carry the spending overhead. For now.

The US hits a discrete crisis first (~2033, trust fund exhaustion forces Congress to act). But Australia’s structural position is worse because there’s no forcing function for reform and no reserve currency to monetize deficits. Australia will slowly degrade until an external shock, e.g. commodity bust, China slowdown, AUD crisis, forces emergency austerity.

Neither country’s productive economy is growing fast enough to escape. Both are at about 1.0% real in their productive sectors. Both need sustained 3.5%+ to outrun the tax-funded economy’s growth rate. No technology revolution in the history of either country has closed a gap that large on a sustained basis. The IT revolution came closest — about 1.3 percentage points of productivity acceleration — and it lasted eight years before reverting. AI would need to be almost twice as impactful and last three times as long. And Baumol’s disease guarantees that much of AI’s benefit in healthcare gets captured as more expensive treatments rather than cheaper ones. More diagnoses, more drugs, more procedures, more cost. Every healthcare technology innovation in history has increased total spending. AI will not be the exception.

Escape Velocity

The only thing that has ever produced a sustained 2+ percentage point productivity acceleration is a cheap energy revolution. Steam. Electrification. Petrochemicals. Each time, abundant cheap energy unlocked a step change in industrial productivity that grew the denominator fast enough to outrun the numerator.

Australia is sitting on the best natural resources in the developed world and doing approximately nothing useful with it at industrial scale. Indeed, Australia spent 50 years after WW2 painstakingly building domestic oil and gas autarky, and the 30 years since then tearing it apart, as revealed plainly during the current gulf crisis and also overlapping nicely with Australia’s recent three decades of economic stagnation.

The fiscal math doesn’t care about your feelings. It doesn’t care about your policy preferences. If Australians want sustainable publicly funded services, the fiscal math needs a denominator that grows at 3.5%+, and nothing short of an energy revolution has ever produced that.

I have recently been using Claude Code (Opus 4.6) to do a variety of software projects. One was the completion of an ill-fated attempt to create a custom watch face for my Garmin Epix Pro 47 mm.

MidJourney produced this conceptual image for me, which is what I was trying to capture.

I wanted azurite/malachite/copper colors, and hands for astronomical details. This would be almost impossible with a physical watch but pretty easy if you have a small screen to work with.

For reference, this is the mineral in question.

I was ultimately able to get something like this working.

Here we have a sun and moon hand that show location both over the Earth’s surface and in the sky with respect to two overlaid polar-projected maps. Hours in Roman numerals, and the sun hand revolves once per day. The Moon hand shows phases and both hands show eclipses when they occur. Gold, silver, and purple curves give the terminator/horizons for the sun, moon, and user respectively. With this you can easily deduce who can see which body and whether and when it will be day time. The background shows the brightest stars which rotate in sidereal time.

This turned out to be so much fun that I designed a few more.

This one shows time on Mars, with a background map showing shaded topo relief of the planet (there are no shorelines at present), and stars. The user’s location is given as the Phlegra Montes, one potential site for a base. The white patch to the left is the Tharsis massif. The dark patch to the right is Hellas. We have a sun hand, and Earth hand, and two moon hands. The red one is phobos and turns anti-clockwise, matching its real world behavior. The hour dial includes N (nulla) to mark the 39 minute leap hour to account for Mars’ slightly slower rotation than Earth. During the time slip, the minute hand splits. NASA uses a different convention with “Mars hours” etc that are a bit longer than Earth hours, but that’s not as fun for watch design. Because Phlegra Montes is the first landing site, its timeslip is around midnight. Other locations on Mars experience the time slip at different times.

Returning to Earth, I wanted something a bit less abstract.

This one was quite hard to get right, requiring some manual coding (shocking!) It has a sun hand, moon hand, minute and second hand as before. The face is oriented such that the sun is at its zenith over the user at noon in their respective time zone. The lit side of Earth is the “pale blue dot” image, while the night side shows city lights. The terminator is orange to mark the sunset/sunrise.

Finally, for completeness, the Moon.

For this one I went a bit minimalist. The face shows a shaded relief map of the lunar south pole out to 80 degrees south. Blue marks the permanently shaded regions where ice may be found. Red marks the hills and slopes from which Earth is visible and vice versa. IMHO, a base should be built here so we can see its lights from Earth. The Moon rotates about once a month so I have a sun on the rim showing its location relative to the pole, and a shaded ring showing the light side of the horizon vs the dark. The relief is shaded in accordance with the solar position. At the noon position, we have a complication showing the appearance of the Earth from the Lunar south pole at that time. Phase, rotation state. During lunar eclipses, its blue atmosphere turns red. In the shot above, the middle east and eastern africa are visible around their dawn – they are shown south up, as would be seen from the lunar south pole. The time is shown with three hands depending from the rim in the conventional manner.

If you would like to try these for yourself, I have the latest .prg files collected here. No warranty, no guarantees, etc. These are vibe-coded and free!

Let me know if you’ve made any cool custom watch faces!

I recently realized I hadn’t written much in this blog about the Fermi Paradox, though I do write about it elsewhere. So here is a quick note.

The Fermi Paradox, sometimes paraphrased as “where are they?” is a question about the apparent lack of intelligent alien life in the universe. The universe is extreme old relative to the speed of light at galactic scales. Light has been able to cross our galaxy a thousand times since the dinosaurs died out, and that was relatively recent (less than 2%) compared to the age of the Earth. So if life is common in the universe (it seems like it might be) and if intelligent life is an eventually winning strategy of evolution (it seems to be) and technological life follows from this (it has at least once) and technology leads relatively quickly to space travel and visible technosignatures (this is probably not the hard part) then why isn’t the universe teeming with alien life?

There are a bunch of potential solutions to this puzzle. I’ll mention a few before I get to my preferred one.

We haven’t looked very carefully.

Space is big. You just won’t believe how vastly, hugely, mind- bogglingly big it is. I mean, you may think it’s a long way down the road to the chemist’s, but that’s just peanuts to space.
Douglas Adams, The Hitchhikers Guide to the Galaxy.

We have only relatively tiny telescopes on one tiny planet looking out into a vast darkness. We have found only a few thousand exoplanets, of which just a handful might be able to support life. We have not a single spectra from an exoplanet atmosphere. Our nearest star Proxima Centauri has planets and we know almost nothing about them. For all we know, there’s already an advanced civilization there and we would not be able to see it. We’ve run various SETI searches for a few decades but again, barely scratched the surface. We could build much larger telescopes but even one the size of the Earth would hardly rule out intelligent life in our galaxy – much of which is obscured by dust.

At our current rate of technology, we’re not going to discover intelligent aliens unless they’re very close by and sending us very powerful radio signals, or they visit us directly.

Interstellar travel might be impossible.

The galaxy might be only 100,000 light years across and nearly 10 billion years old, but you and I typically travel at perhaps a meter per second, while light covers the same distance in just 3 nanoseconds. That is, the galaxy is relatively small if you’re a photon, and impossibly enormous otherwise. Our fastest space probes would take nearly 100,000 years to reach the nearest stars. Antimatter might be energetic enough to accelerate to close to light speed, but that doesn’t mean that interstellar travel is possible – colliding with a single dust grain would be very bad news. Perhaps the galaxy has a million technological civilizations, and they’re all trapped in their respective solar systems by the enormous gulfs of space.

The Great Filter.

Maybe intelligent alien life is rare because there’s some filter or set of filters that kills off life forms that get too advanced. This filter could be in our past (multicellularity, asteroid extinction, solar flares) or in our future (nuclear war, hostile aliens killing upstarts, AIs starving us to death, depopulation, loss of culture of exploration). But you only need one very powerful alien species to overcome these filters and then they can fill up the galaxy relatively quickly. As far as we can see, the galaxy is not full.

Near light speed travel is hard to observe for people at the destination.

This is my preferred explanation at present.

The most interesting stuff I’ve read about the Fermi Paradox is Robin Hansen’s work on Grabby Aliens, which uses the fact that the universe appears to be empty and that cultural selection on expansionist aliens would lead to their rapid spread if they did occur to conclude that intelligent life must actually be very rare (fewer than one species per multiple galaxies) or that evolution must be very slow.

There is an observational subtlety to alien observations, which is that when we look out into the universe we are observing only our past light cone. If grabby aliens were expanding at a high fraction of the speed of light (c), the light carrying information of their coming would be only just ahead of them. So even though aliens might be quite close, we wouldn’t see them until just before they arrived. In fact, there is quadratically more available space further away from Earth, so while a nearby alien species might reach us with their slower, first generation starships, any starships that get here from more distant parts of the galaxy are almost certainly the fastest, latest tech ones which overtook the slow ones on their way here.

The universe could be in three different states, observationally. What we observe (no aliens), aliens seen but not here yet, and aliens among us. But if the aliens we see are traveling at high speed toward us, the intermediate state (seen but not met) is unlikely to be longer than a handful of weeks. Choosing our present time at random, there is almost zero chance for humanity to find itself in a time where we’re aware of alien intelligences but haven’t yet met them. That is, Earth is 4.5 billion years old (no aliens), then one day the Vera Rubin Observatory sees a flash that turns out to be an alien spacecraft departing to meet us from 100 light years away, traveling at 99.9% of c. They arrive just five weeks later. For the remaining billions of years of Earth’s existence, we are in the world of aliens among us.

I think it’s physically possible to reach 99% of c with current human technology, so there’s no reason to suppose aliens with better technology would fly slower than this, and they could fly much faster.

I put together this chart a year ago. If relativistic aliens are flying towards us, we won’t see their launch until the light gets here, and if they’re right behind the light, they’ll be here soon after. For example, reading this chart, if they’re traveling at 99% c, we will see them only when they’re 99% of the way here. If they’ve traveled 1000 light years to visit us, we’ll see them (at best) 10 years before they arrive. We might not see them at all – 1000 light years is far enough away that some stars are too dim to see with the naked eye. Meanwhile, 1000 light years is a long way to go, so it’s fortunate that at high speed, relativistic time dilation kicks in and helps to pass the time. This is shown with the yellow curve. At 99% c, the 1000 light year trip only feels like 142 years. This is still a long time, so perhaps they will travel to us at 99.9% c. In that case, the trip will feel like only 45 years to them, and we will get a whole year of warning, assuming we see them launch 1000 light years away.

I think this factor is under-estimated when discussing the Fermi Paradox. If most of the planets in the universe are too far away for us to see alien life, then if we see it at all we’ll be seeing their space ships as they come to us. But we won’t even see them launch to us, even with perfect telescopes staring out into the galaxy, until they’re almost here. In practice this means that, in the grand scheme of human history, the phase between becoming aware of aliens and meeting them is vanishingly short.

This quirk is intuitively obvious in the context of supersonic planes – whose sound arrives after the plane.

How to use this chart: Select your speed on the horizontal axis, and decide on your travel distance. Then run up vertically to read off the distance-time multiplier (blue line) for visible travel time on the ground and (orange line) the apparent travel time for the traveler due to time dilation. For example, let’s say we’re doing 99.5% c over 500 light years. Then we’re going 0.5% slower than c, so the delta t multiplier is 0.005*500 = 2.5 years, while the subjective travel time is 0.1*500 = 50 years. We will be in flight for 502.5 years, we will arrive 2.5 years after our light, and on board we’ll feel just 50 years pass by.

The Grabby Aliens hypotheses points out that expanding alien civilizations appear as circular regions in the night sky where, for example, we can observe spectral changes in stars or their planets, given a sufficiently powerful telescope. For an expansion speed that’s small compared to c, this gives the correct intuition. But, at higher speeds, the apparent angular size before contact shrinks. You might think that you’d see the alien sphere expand through stars in your field of view until it surrounded you, but in fact the light from their arrival at nearby off-axis stars is still on its way to you when they arrive. So the apparent shape of their expanding sphere, looking into our past light cone, is a cone whose narrowness increases with flight speed. In the extreme case, we would see nothing even with a perfect telescope. It’s quite hard to see things thousands of light years away!

There are a couple of other aspects to the Fermi paradox. It seems to me that the Fermi paradox can be at least partly explained if either relativistic interstellar travel is relatively easy, or any kind of interstellar travel is basically impossible. I think the intermediate case is ruled out quite well by even our limited observation.

I favor the first explanation. The implication is that the night sky is not full of alien civilizations because they’re expanding so fast that the period of time between our feeble telescopes being able to detect expansion and them actually arriving is extremely short. This does, however, imply that no traveling aliens could have occurred in our galaxy in the past billions of years, right up to barely 100,000 years ago, when our ancestors first started leaving Africa. There is still no good reason for this to be true, other than the anthropic principle.

Accordingly, when we look up and wonder where are the intelligent aliens, we can know two things for sure.
1) Our telescopes are bad and we should feel bad.
2) They could be passing Betelgeuse (700 ly away) right now on their way here and we would still not have seen their departure. If they’re going fast enough they could be closer and brighter than Alpha Centauri and we still wouldn’t have seen them yet.

And if we could only detect them at Betelgeuse, picking up a thruster signature with blue-shifting indicating 0.99 c travel speed, they’d be here in about 7 years (running just behind the light announcing their arrival) – an even more laughably ridiculously short period of time for us to know we’re not alone and have not yet shaken tentacles.

With Vera Rubin telescope up and running, we’d have a chance of detecting incoming relativistic spacecraft out to maybe 1000 LY, which means 10 years warning at most. If they can hit 0.99c, why not 0.999c?

A quick note collating a list of blogs I’ve written on the topic. I’ll update it with good third party write ups as I become aware of them.

Context: SpaceX announced a refocus on lunar development. My best explanation for this is that SpaceX wants to accelerate human occupation of space which, requiring enormous resources to sustain, means accelerating capitalism in space. Five years ago, building a city on Mars was conceived as a philanthropic venture and the best a market could offer was a set of tools to maximize the utility of each marginal tonne of cargo upmass. Today, exploding demand for power-intensive AI applications provides enough of an upside to justify producing components in space, providing the economic engine necessary to justify and fund the trillions of dollars necessary to build and sustain space factories.

The High Frontier: A Technical Critique. Gerry O’Neill’s 1976 far sighted classic imagines the MVP space colony supported by a (not yet launched) Space Shuttle, adequate to bootstrap the sci-fi future we all want. Just one problem: The economic “anchor tenant” was to be space-based solar power, and Space-based solar power is not a thing. Generating solar power in space to beam to Earth to compete with ground-based solar, gas combined cycle, etc, is “out of the money” by a factor of about 100 million.

In fact Starlink, a constellation of about 10,000 satellites with a combined power generation capacity in the 100s of MW, shows exactly how all viable space-based economic systems work. Send microwaves to Earth-based receivers containing not raw power, but sufficiently valuable binary bits of information. Typically, this is Earth observation data (eg Planet, Umbra) or telecommunication data (eg Starlink). It turns out that the economic value of a received Watt of Starlink microwave power is about a billion times higher than the marginal value of a Watt of electrical power, and so Starlink actually makes money. A lot of money – currently about $10b per year and growing.

The economics of orbital “datacenters” or essentially glorified Starlink satellites with a bunch of GPUs attached are likely to be even better than Starlink. Why? Because the value of a byte containing an inference token from some AGI is much higher than the value of a byte containing some error correction on a P frame of your TicTok video. Taking early 2026 data, the economic utility (price) of inference could easily be 100x the cost for ground-based datacenters. Ignoring the rapidly stiffening elasticity of land supply for solar-powered datacenters on the ground, space-based inference might cost twice as much as ground-based, but that still leaves 98% margin for profit. In other words, those of us who have been looking toward the growing maturity of the fully re-usable Starship for some time and wondering just how much demand it could possibly induce can rest easy: We’re unlikely to saturate our demand for intelligence and like Falcon and Starlink, Starship and orbital inference will print money. Some more details are spelled out in this post: Direct current data centers.

So far so good, but what is the limiting factor? Starlink is currently about 200 MW. Starship can lift orbital power generation to about 100 GW with perhaps 10,000 launches per year, or about one per hour. To go much beyond this, most of the satellite mass needs to be produced in space. The Gerry O’Neill concept calls for the extraction and possible refining of raw material on the Moon, its launch into cis-Lunar space with mass drivers (vintage O’Neill video!), and final processing and assembly in large orbital space stations. Much less than 1% of the final satellite mass is GPUs, so they could be conceivably imported from Earth for many years to come.

The advantage of processing in space is that unlike the Moon, a factory in space will have power 24/7/365, while most of the Moon is shaded for two weeks at a time during the lunar night. Converting the Moon’s crust (lots of alumino-silicates, fortunately) into aluminum and silicon is a power-intensive process.

Here are my notes on Starship Moon Base Design Principles. The short version is: avoid excavation and construction as much as possible.

No matter what, the Lunar base will still need plenty of power to run life support, crush rocks, and shoot them into space. I’ve written two posts on Lunar power. The first (Powering the Lunar base) looked at the challenge of providing consistent power using conventional methods throughout the Lunar day and night, even in polar areas where a few mountains get a bit more sun throughout the year. The second examines Powering the Lunar base with Earth-based microwaves. Not everyone agrees with me but I think the math is quite clear – beaming power from Earth-based power plants to the Moon is 1000x cheaper and much faster than the next cheapest option, and allows what cargo mass we do send up to be focused on production of ores, not operation of finicky and labor-intensive power plants.

Here are a few posts on operations with Starship. Long duration propellant stability in Starship. Lunar Starship and unnecessary operational complexity calculates all the “gear ratios” for transporting mass to and from the Moon using Starship with all the different orbits and engine combinations you could possibly want.

This leaves one final area of discussion: What about NASA?

I wrote Artemis can succeed using Starship around the time of the Human Landing System (HLS) award, pointing out that Starship (or a sufficiently capable Blue Moon system) routes around the damage of NASA’s default architecture, rendering SLS, Orion, and the Lunar Gateway redundant and thus enabling their speedy deletion from the critical path.

This is not a purely academic matter. In NASA is Worth Saving, I argued that NASA’s fundamental mission is to preserve the light of freedom for the future of humanity, a message that’s often lost beneath parochial and mundane concerns which continually erode NASA’s operational capacity.

Regarding NASA’s default architecture of SLS and Orion, which promise to cost $100b and land zero people on the Moon without the HLS contractors, I have written: SLS: Is Cancellation Too Good, SLS is still a national disgrace, and NASA’s Orion Space Capsule is Flaming Garbage. At the time of writing we stand on the cusp of NASA’s attempt to fly four humans around the Moon on Artemis II, a mission all but certain to generate entirely new classes of agonizingly embarrassing and hopefully not tragic system failures for hapless program managers to attempt to excuse and explain away with such phrases as “within the family of expected outcomes” and “it’s a learning process”. That is, if we can shift the goalposts fast enough to complete a wet dress rehearsal without bricking some critical piece of hardware.

The difference between SpaceX’s plan to commercialize lunar activity and the legacy systems we’ve been saddled with could not be more clear.

Casey Handmer, Matt Weickert

Originally posted at the Terraform Blog.

This post explains our current views on how humanity will achieve Kardashev Level 1 status by exploiting the full energy resources of an entire planet. More specifically, how pure solar+batteries will power AI scaleup beyond gas turbine manufacturing limits.

[Edit: Thank you for the nice write up, Ana!]

It is an extension to my earlier post of March 2024 on using solar to power AI datacenters, and a response of sorts to the Scale Microgrids paper that showed a mix of solar and gas could reduce emissions for the developers and operators of next gen AI datacenters. In that paper, Kyle Baranko, Duncan Campbell and co-authors showed that around 90% solar with local natural gas backup generators would be the fastest way to get power. In this work, we show that taking this trend to its obvious conclusion and deleting all the legacy fuel-based power components can be even faster and cheaper. We also include a discussion of space-based inference. 

Let’s examine this problem from first principles. What is silicon cognition?

You can call it a tensor core, a Blackwell, a GPU, but these are all versions of the same thing. A sliver of silicon with billions of transistors, through which cascade a torrent of electrons converting the entropy of a few volts to the entropy of information generation, and the entropy of waste heat. A GPU is a very complicated switch that regulates current flow, with some other side effects.

For the foreseeable future, the GPU will be the expensive part, currently valued at around $50,000/kW. All it needs to continue to operate is an infinite supply of moderately spicy electrons, that is, a DC power supply at a few volts. Given that making power is much simpler than thinking, the job of the power supply is to be uncomplicated and relatively cheap. In no universe should providing power be the hard part.

Solar and batteries are a natural match to this demand. A solar panel is a slice of silicon (without logic gates) that absorbs solar photons and drives electrons uphill. To a good approximation, a solar module is a constant current source that maxes out at about 40 V. A battery is a reversible chemical reaction that stores and releases electrons, and to a good approximation is a constant voltage source. Modern lithium chemistries hold at about 3.9 V across nearly their entire state-of-charge range. 

For logistical reasons related to the relative scarcity of copper in the crust of the Earth, it makes sense to operate solar cells, batteries, and GPUs in series so that the entire system runs at about 1000 V and each electron can be reused a few hundred times. 

Our radical claim is that, in the limit, Earth-based AI compute will look like this: 

By area, thousands of acres of solar panels. 

By cost, a pile of GPUs. 

In the limit, Earth-based AI compute will be a direct current (DC) solar array connected to a DC battery bank connected to a DC GPU rack. 

This approach brings numerous other advantages: 

• No grid connection. 
• No moving parts. 
• No turbines. 
• No gas connection. 
• No nuclear fuel. 
• No emissions. 
• No power conversion. 
• No transformers. 
• No inverters. 
• No power transmission. 

None of these parts make the AI smarter, and all of them can potentially intrude onto the critical path. 

Delete.

It sounds nice in theory, but how can this work?

The key metric to optimize is tokens per dollar. For example, take Scale Microgrids’ work on a 90-10 AI system, increase the size of the solar and battery farm enough to get to 99+% uptime, delete the gas power side, and compare overall economic productivity. A gas system that’s used only 1% or 0.1% of the time still costs time and money, and that’s the core reason why deleting it can end up reducing overall cost. 

The graph below shows the tokens per dollar landscape for two hypothetical solar powered AI systems, one with a gas powerplant (blue) and one without (orange). Both have a solar array (size given in nameplate multiples of the peak AI load) and battery size (given in hours of capacity at full load). 

The key insight is that there are two stable attractors. One with a pure gas energy supply, with solar and battery supplementation for vibes, CO2 reduction, or marginal capacity expansion. The other with pure solar and batteries, no gas. The pure gas system capex is minimized with no solar and batteries, as natural gas itself is relatively cheap given no preference for emissions reduction. But the two manifolds intersect along a frontier, and beyond that the solar array and battery are capable enough that it’s actually cheaper to delete the gas powerplant entirely.

This tradeoff does not come at zero cost. In exchange for deleting the cost, complexity, and schedule risk of a gas powerplant comes the sizable land demands of a solar array. To a rough approximation, 15 acres of solar are required per MW of DC AI load. For reference, the USA has about 150 million acres of unpopulated desert west of the Mississippi, enough for 10 TW of AI development. 10 TW is much more than total global electricity generation today. There is plenty.

On the other hand, while fracked gas is relatively abundant (for now) the turbines that convert it into power are hard to make, hard to ramp, and largely already spoken for. If AI seeks growth beyond the production ramp of turbines, it is clear which way the wind is blowing. 

Before we get to the methods section, I’ll give a rough heuristic for performance. Assuming an on-off binary state on the load, a 15 MW solar + 15 MWh battery can get to ~99% utilization anywhere in the US south west, but is that good enough? The short answer is yes – maximizing tokens per dollar spent, or ROI, justifies throttling demand on a few of the longest, coldest nights of the year. 

But it’s actually better than that. Remember that a GPU is a glorified silicon switch intermediating the flow of electrons downhill. Power consumption is proportional to clock frequency multiplied by the square of the voltage (P ~ f V2). GPU power consumption is not fundamental: Token production rate is. If we’ve deleted DC-DC converters then voltage is set by the state of the battery, and frequency is controlled by software. This means that a 3% reduction in token production rate can buy us a 9% reduction in power consumption. So the math changes from 99% utilization to more like 99.7%. This shifts the economics around solar and battery plant sizing considerably, given that GPU frequency modulation allows for a 3x discount in actual utilization and token production.

There is one other implication of these wildly capable and versatile solar+battery AI data centers. They have enough power to operate at full, or nearly full, capacity for the entire year. For 10 months of the year they are oversupplied, and can provide electricity and low grade heat (from their cooling systems) to neighboring customers essentially for the marginal cost of power transport. These could be seasonal or intermittently friendly loads such as the synthetic hydrocarbons and primary materials being pioneered at Terraform, and/or local communities. At Terraform, we believe that power should be as cheap as possible. 

Methods

Epistemology. How is it possible for this lightly evolved monkey to know these things?

You will need:

A year (at least) of real time solar data from a target location. This is data for an EW fixed tilt array in Texas that we generated by feeding fixed south tilt data into a slightly non-trivial geometry model.

A solar PV module IV curve model. This is based on the JAM72S30-540/MR/1500V but they’re all pretty similar.

A Li-ion discharge voltage curve. 

A frequency-power curve for a typical GPU.

Plug these all together in a model that charges the battery when the sun is up, provided the panel voltage is high enough. We initially simulated a system with no power electronics whatsoever, but found that battery charging efficiency was inhibited when the battery state of charge was low, because pulling the panels to a lower voltage actually decreased their efficiency. Given that MPPTs are not that expensive, we could put them back in.

Then, provided the battery can deliver power, the GPUs are powered and we count how many tokens are generated.

Throw in a basic “governor” that throttles the GPU when it predicts the battery will be exhausted before dawn. 

This graph shows performance over a ten day period in winter. Note how the governor throttles output early on the fourth day by rationing power until the following morning. The cubic power consumption of GPUs means that throttling a little bit early is much better for token production than running full blast into a wall and then dropping to zero production until the sun comes back up.

Now run thousands of simulations for every combination of battery size and array size, measuring overall utilization of the load. 

This chart shows yearly utilization of a GPU asset given solar and battery sizes, including our basic governor. Note that a “steepest” ascent starting at zero solar and zero batteries turns first on adequate solar, then adequate batteries, then marginal solar, then marginal batteries. This reflects the shape of the resource curve and the degree of exploitation required to get to the marginal nth 9 of reliability.

This chart shows curtailment reduction with adoption of the minimum viable governor vs some naive on/off operator, showing 2.3-2.6x improvement, which is close to the 3x implied by the GPU’s cubic power consumption. This governor is not very sophisticated, for example, it has no ability to take weather prediction into account. It merely assesses the time of day, the state of the battery and of solar generation and curtails GPU utilization accordingly. 

Throwing in assumptions about capex, we can assess capital efficiency.

This chart shows token production per dollar (in arbitrary units), showing a rather broad peak with considerable flexibility. Adding too much solar or batteries degrades capital efficiency – the correct response is to add more GPUs in this case. 

Because the peak is so broad, there is freedom to choose for one additional preference. That is, we can alter the size of the array and the battery by 20-30% with respect to the load and still get much the same return on capital. Given that land is finite, we may want to maximize tokens per acre while holding development cost constant, which puts us towards the lower edge of the peak in the diagram above. Then, holding land use and tokens per dollar equal, adding more battery towards the bottom right of the peak increases absolute token production on a fixed GPU and solar array asset base. This mirrors actual operational optimization, which is to say, pave all available land with solar, then add GPUs and batteries until revenue peaks. 

At last, the machinery to perform a comparison with a gas or gas-solar hybrid system is in place. Plug in some assumptions around GPU cost, solar cost, battery cost, gas turbine cost, gas fuel cost, and amortization period, and you can produce this chart.

Here we assume that GPUs are $50,000/kW, batteries (including all ancillary power electronics) are $200/kWh, solar is $200/kW, gas turbines are $2500/kW, gas is $55/MWh, and we’re amortizing over 10 years. 

The chart suggests the possibility that under some set of assumptions, it’s actually cheaper to delete the gas power system entirely, so what are those assumptions? For any given cost, select the peak utility point for solar array and battery size and marginalize across these parameters.

The left side of this chart shows where a pure solar system is the best value. As a rough rule of thumb, this is where 1.3 x battery cost + solar cost < $500/kW(h). 

As a sanity check, in early 2026 we’re seeing large scale integrated battery storage systems ship from China for well below $150/kWh, while the cheapest industrial scale solar systems are going in for under $200/kW-DC-nameplate. This is beyond the critical cost threshold – delete the gas system.

Let’s recap.

It is possible and even optimal to run a datacenter on pure solar and battery. The optimum level of availability is between 99% and 99.9% utilization, with the balance taking the form, primarily, of throttled use rather than lights out. 

The pure solar+battery data center is cheaper than a gas-assisted or pure gas data center if solar and battery costs per kW are <~10% of the turbine/gas generator cost. For example, if a 1 GW gas turbine costs $2.5b and the solar array costs $250m/GW and the BESS costs $250m/GWh, then the utility is roughly at parity. 

We estimated the opportunity cost of a month of delay on a 1 GW gas turbine being delivered at about $20m, which probably isn’t high enough to justify deleting the component out of pure suspicion, provided that you are confident it will be delivered eventually. 

On the other hand, we found that it is possible to commission a pure solar+battery data center with high utility and then backfill additional GPU capacity if/when a gas turbine becomes available. 

There is broad latitude for design flexibility within the peak utility (tokens/$) range. For example, there are points with equal utility that have 20% less solar or significantly more availability, depending on secondary constraints such as land availability.

Given that, long term, trends will favor pure solar+batteries and the performance relative to complexity is already favorable, there is an argument that one hyperscaler should probably move aggressively in that direction, so as to obtain differentiation. 

We investigated what the performance hit would be for a pure solar+battery DC power system that deleted non-computational silicon, that is, inverters, converters, and MPPTs between the solar and battery components, and even in the racks themselves. We found that the power system performance relative to cost improved with the deletion of AC and DC-DC conversion components between the arrays and batteries. Again, this is a nod to the future we will converge on.

Space AI

Late 2025 saw much speculation about space-based AI. It seems to me that SpaceX, with their incumbent advantages in launch and Starlink hardware expertise, may be able to ship gigawatts of inference compute into Earth orbit for something like 2x the per token cost of ground-based AI, but that this would still be quite profitable. Why bother? It’s a separate delivery and distribution channel that isn’t congested by the usual permitting and regulatory nonsense at play on the surface, or at least, a different and uniform set. And if you have unlimited launch upmass it helps to have a profitable use case, like Starlink or orbital AI, to soak up that supply. 

Let’s list their respective advantages and disadvantages.

Space AI

Positives:

• Simplified and unified regulatory regime.
• High altitude (800 km – 2400 km)  dawn dusk sun synchronous orbits are never in shade (except momentarily during rare lunar eclipses) so don’t need batteries.
• Don’t need to be cheaper than ground AI, as long as they’re cash flow positive.
• Infinite source of marginal launch demand can fill in gaps from other customers for a very large rocket.
• Passive stabilization with dihedral solar arrays is possible.

Negatives:

• Launch cost is not free.
• No maintenance capability.
• Higher latency (50 ms worst case).
• Radiation can accelerate GPU degradation.
• Thermal environment requires relatively large radiators and (probably) active cooling.

Ground solar AI

Positives:

• Mostly a land development play, plenty of unused land on the ground.
• Batteries are relatively cheap compared to GPUs and enable operation overnight, getting cheaper all the time. 
• Almost certainly cheaper than launch and space rating of components.
• Don’t require a million tonnes a year to low Earth orbit to deploy.
• Easier to maintain/retrofit.
• Lower latency (closer to end users).
• Default human Earth surface environment is less hostile to hardware.
• Can easily cool using HVAC and air. 

Negatives:

• Regulatory/permitting is painful, byzantine, and locally variable.

Fundamentally this is a bet that GPUs are so valuable on a per gram basis that even launching them to space helps improve the economic utility of a Watt of solar power. 

Finally, an application of space-based solar power that can justify something like the vision of Gerry O’Neill. To be clear, this is because the value of a watt of space-transmitted microwave power encoding an intelligent token of data is about a trillion times higher than the value of a watt of space-transmitted microwave power competing with your local power plant to supply the grid.

Conclusion

Whether in space or on the ground, turbines are irrelevant to reaching Kardashev Level 1. The fastest growing AI will win, and the fastest growing AI must delete all non-essential parts. The only essential parts are a solar array, a battery, and the GPU itself. 

Aside: Carl Sagan’s extension of Kardashev Levels from integers to reals

Planetary power: 1016 W -> K = 1
Stellar power: 1026 W -> K = 2
Galactic power: 1036 W -> K = 3

One step on the Kardashev scale is equivalent to increasing power consumption by a factor of 10 billion. 

What is the current level of humanity?

Global electricity production: 3.5 TW -> K = 0.65.
Global fuel consumption: 20 TW -> K = 0.73.
Global planetary surface use for agriculture: 13% -> K = 0.91.

If all of Earth’s land was paved with solar PV at 26% efficiency -> K = 1.01. 
If the entire Earth including oceans was paved with solar PV at 26% efficiency -> K = 1.06.
If we fill the unshaded dawn/dusk sun synchronous orbital (SSO) band (800 km to 2500 km) with SpaceX AI satellites, 10^17 W available -> K = 1.04.
All of Earth plus the SSO orbital band -> K = 1.085.

Convert the entire Moon into 2 kg/m^2 solar inference at Earth’s solar orbital radius -> K = 1.91.
Convert Mercury into a Dyson sphere (7 kg/m^2 density) @ 26% PV efficiency -> K = 1.9998.

No way to get to K2 without a slightly more efficient solar panel! 

Printable pdf.

It’s been a few years since I wrote a broad post on energy, so I’m providing an update in one easy to read place. More detailed specific posts on energy are here. 

If you want to work on the future of energy, we’re hiring for a wide range of opportunities at Terraform Industries. 

Current fuel mix and uses

The US consumes about 100 quadrillion BTUs of energy per year. Of this, about 80 start life as coal, oil or gas, and roughly a third of the energy mix serves the electrical grid. Less than 1% is food, reflecting our enormous energy wealth in comparison to our pre-industrial forebears. 

Energy mix and outlook

We’re in the middle of a period of rapid transition. Much will be much clearer in retrospect. This is how I think it will shake out.

Primary production will be mostly solar

Solar is decosting at an accelerating rate. 

The production-weighted learning rate is 48%! Module cost falling up to 20% per year, twice what it was five years ago. 

Coal is shrinking. Nuclear is flat. It doesn’t take Sherlock Holmes to see which way the wind is blowing.

Solar power production will locally feed the grid, but also provide power behind the meter, and beyond the meter in off-grid developments. 

Batteries everywhere, the grid will shrink

Do we need to drastically expand the grid in order to metabolize renewables? No.

Batteries have been winning for about a decade and the gap is only increasing. 

Battery production is growing while costs continue to plummet. 

Will we build more batteries? Yes. Per capita allocation of batteries has already increased from a few grams to a few tonnes in a single generation. It will only continue up and to the right. Demand will not saturate until after it stops being induced. 

Democratized battery ownership is good for freedom

Will batteries be deployed behind the meter, at the point of generation, or within the grid? 

Yes. 

Batteries at solar arrays allow higher utilization of offtake grid connections, matching evening power consumption. Batteries behind the meter allow granular, independent assurance of power continuation. Batteries in devices and vehicles. Batteries in houses. Batteries on power poles. Batteries at substations. Batteries in schools. Batteries in appliances. More! 

Already we are seeing adoption of behind the meter batteries such as the Tesla Powerwall for individual consumers who can justify the expense relative to the hassle of utility power cuts. In a world where every consumer can choose the size of their battery, it doesn’t make sense to spend 10x the money trying to keep the distribution grid at 99.9% production. Less developed markets are pointing the way – Pakistan has cut most domestic power consumption over to solar and batteries in about two years. This is analogous to the growth of cell phones in developing countries that never ran copper phone lines to every house. This trend puts more value and market power in the hands of individual consumers. In the limit, market power will shift from the monopoly electricity utility to amorphous confederations of illegible behind-the-meter demand in the form of networked batteries. 

Datacenters will be the nexus of electricity production growth

In 2025, headlines scream that datacenters are pushing prices up and consuming all the power. I think datacenters are exposing the rot in a moribund power generation and delivery industry which has proven unable to meet demand in recent years. But it is a moot point.

Datacenters are already building their own captive power plants. As AI demand outstrips production of gas turbines, hyperscalers will turn to offgrid solar+battery power systems, which are already competitive with pure gas or gas+solar in the sunnier parts of Earth. 

Depending on location, 10x overbuild of solar and batteries are sufficient to hit >99.5% uptime for the GPUs. 

Datacenters will be net power sources for their communities

On the flip side, these captive solar power plants will be curtailing approximately 75% of their generated power and will be able to provide net power on all but a few days per year. That is, 99% of the time, which is substantially higher utilization than any conventional thermal power plant. 

Within the next five years, market power between utilities and datacenters will flip, with DCs becoming the preferred load growth power generation partner. 

To spell out the implications, this means that consumers will get access to extremely competitive (cheap) power most of the time, and some combination of utility-owned and privately owned batteries will be needed to smooth out the gaps, as they would be anyway.

Solar datacenters will ultimately be pure DC constant voltage systems

Solar PV modules are approximately constant current sources. Lithium ion batteries are approximately constant voltage sources. GPU power consumption scales like the cube of token production. Why fill the entire thing with DC-DC converters and inverters? In the limit, it’s all a single piece of silicon. 

Substantial cost improvements are needed to make space AI competitive

If SpaceX or a competitor can ship inference compute to a 560 km unshaded sun-synchronous orbit which is 80% 1 kg/m^2 solar arrays by mass and 80% compute by cost, then it should be possible to make money. Otherwise, we can expect to see compute being developed on the ground. 

Electricity power markets should evolve toward real time and local prices

Real time matching of supply and demand will require responsive time- and location-based pricing. Different regulatory regimes are already experimenting with versions of this. In general, regulatory insistence on unphysical pricing schemes are a choice to socialize the costs of pathological markets. Almost by definition, capital allocation in opaque or non-existent markets with unresponsive prices will be less optimal, driving costs higher and increasing the value of the available price arbitrage.

I pre-register my belief here that electricity governance markets will bifurcate. On one side, we’ll see those that embrace a steady cadence of pricing reforms, allowing effective competition between many private operators of generation, storage and transmission assets, pushing prices down. On the other side, increasing prices for consumers will drive increasingly desperate governance measures that allow far less competitive storage operators to extract vast rents from the difference between real world power conditions and the conditions approximated by some legal framework.

Just remember, the universe does not care about how we encode our opinion as to how we think the world should work. It has given us an infinitely powerful sun and a planetary crust composed largely of silicon. What we do with that is up to us. 

Seasonal load variation – summer

Hot climates see increased loads due to air conditioning during the summer. Solar power systems also produce more power during the hottest days of summer, so this is a complete non-issue. Essentially all suburban houses can easily run their own ACs off rooftop solar, so we don’t even need an expansion of power distribution capacity. As an exercise, figure out how cold a standard issue suburban house could be made with a rooftop system. 

Seasonal load variation – winter

Nearly all winter load increase occurs in cold climates which also suffer a reduction of solar power in winter. This load is for heating.

Current battery technology is marginal for seasonal power storage. Conventional wisdom would dictate that we either need radically cheaper batteries, greatly expanded overbuild of solar or wind generation, continued burning of fuel for heat in winter, or a bunch of winter-only power plants that will have the same terrible utilization economics as seasonal-only batteries. 

These ideas overlook one important fact. Storing electricity for months is economically difficult, but storing heat is easy. Austin Vernon has been building ultra-low-cost thermal energy storage at Standard Thermal. Essentially a giant hair dryer blows hot air into a large pile of sand during the summer months with abundant cheap power. In winter, the fan switches direction, extracting heat. The storage medium can be made almost any size and is self-insulating. You can think of it as artificial geothermal power storage – in fact it has several strengths that geothermal lacks, like the ability to cheaply build and renew stored power.

While heat pumps can achieve higher efficiency, consumer uptake has been much lower than expected because fundamentally, a heat pump is a 20 year bet on future power prices that most homeowners are unwilling to make. 

Synthetic fuel is our path to chemical energy abundance

At Terraform Industries, we’re pioneering the technology to convert cheap solar power, air, and water into synthetic natural gas and other hydrocarbons. Within the next five years, solar cost reductions will drive our process to be cost-preferred in all hydrocarbon import markets, and geological sources of oil and gas will never again be able to compete. Our grandchildren will be swimming in copious cheap energy and wondering what all that drilling was for.

We believe that the path forward is lime-calcite captured CO2 + electrolyzed H2 to make CH4 and CH3OH (methanol). Methanol can be upgraded via a wide variety of existing petrochemical processes to make DME, ethylene, propane, gasoline, kerosene, and almost anything else you can imagine.

Hydrocarbon usage patterns will change a lot

In 2025, most gas is used for electricity generation, while most oil is used for cars, trucks, ships, and aircraft. 

Solar is going to continue to displace all other primary electricity generators. And electric cars and trucks will continue to dominate growth in ground transportation.

By 2045, natural gas will be used as LNG primarily for high performance supersonic aviation, shipping, and industrial heat. 

Methanol will be used as the universal industrial chemical precursor for plastics, paints, fertilizers, adhesives, as well as specialty fuels. Kerosene will service the legacy aviation fleet. Internal combustion piston engines will ultimately go the way of the piston steam engine. 

The United Kingdom needs wind

The only highly populated industrial country unable to trivially meet its electricity and synthetic fuel needs with solar alone is the United Kingdom, due primarily to a high population density and high latitude. The Nordic and Baltic countries are tiny by comparison. 

Among other problems, the UK needs to decide if it wants the future where energy is cheap and it is rich, or the future where energy is expensive and it is poor. 

If the former, it is time to get serious about large scale deployment of wind power, using home-grown vertically integrated technology at prices as low as $10/MWh. It is not forbidden by the laws of physics. 

Ultracheap solar power will eventually change mining

They don’t want you to know this, but rocks are made of metal oxides, and infinitely abundant commonly occurring rocks such as basalt contain basically every metal you could ever want. 

With sufficiently cheap power, we no longer need to travel to the ends of the Earth to build mines. Instead, build a solar powered rock refinery at your local gravel pit. 

Coastal deserts will be irrigated with desalinated water

Israel already does this at scale. But much of the coast of Australia, Chile, Peru, Namibia, South Africa, Mexico, Saudi Arabia and other gulf states have essentially infinite quantities of cheap land, free solar power, and sea water. Democratized solar desalination technology can turn any and all these areas into arbitrarily lush paradises with <1% of the available land under solar arrays. 

What have I missed?

I’ll add topics here. 

Originally posted on November 29, 2025 by Dr. Christine Corbett Moran

When asked on The Cheeky Pint podcast how we educate our children, my husband Casey Handmer replied “benign neglect.” It’s a cheeky answer that captures something real: we’re neither tiger nor helicopter parents. But after seven years and three kids (with a fourth on the way), we’ve developed a more deliberate approach. Here’s what we’ve learned.

The Family as a System

If I’ve taken anything away from seven years of parenting, it’s this: the family is a system. Treating it as such, rather than asking “is X better than Y?”, leads to more effective choices.

For working parents who want to scale both career and family, three system-level principles matter most:

• Reliability is non-negotiable: Even in a flexible workplace, unexpected time off creates disadvantages and logistical hassles. Any out-of-home option needs backup childcare (nanny or easy-to-access service) for inconvenient hours, school breaks, and illnesses. The younger the child, the more frequently you’ll need backup.
• Over-invest in childcare: Optimizing for convenience, reliability, and family-fit is worth spending more than you’d think, even more than one parent’s salary in extreme examples. This keeps both careers on growth trajectories, builds retirement savings, and makes having more children feel manageable rather than overwhelming. The cost of a nanny grows sublinearly with the number of children.
• Scalability matters: Decisions that work for one child may not scale to two or three. Time spent driving to individual activities multiplies. Costs of private schooling multiply. We’ve consistently preferred options that work better as our family grows.

The Caveats

This blog details our family preferences alone, which also change over time as we grow in number and maturity. We’re two startup executives with three kids (7, 5, and 2) and a fourth on the way. Our children are somewhat precocious, but otherwise don’t have special needs thus far.

My intention is that this post may provide helpful inspiration to find or tune your own family preferences. If I’ve taken anything away from the past seven years of parenting, it is that it’s helpful to think in terms of global, rather than local, hill climbing towards a better situation. With that, let’s kick things off.

Infancy (Birth to Age 2): Swapping Daycare for Nanny

Emily Oster’s Cribsheet covers from birth through early preschool, and is an excellent data driven read. A key takeaway regarding these years is:

Relative to pregnancy, there are fewer things here where the data will tell you what to do or avoid. Your family preferences will be more central.

What this means is that very few specific interventions actually matter for outcomes on a statistical level. The items which do are the overall quality of parenting and family life.

As new parents in 2018, we had a lot to learn and had relatively simple criteria. Both of us wished to continue our careers with little interruption, so a parent staying at home was off the table. With that our initial considerations were:

• Does it support both of us going back to work?
• Is it convenient?
• Is it sufficiently high quality and within our means?

One additional thing we tried to avoid was an abundance of electronic or “blinky” toys.

Childcare center

In the early days with our first child, we went with a child care center associated with JPL, the Children’s Educational Center, which fit the bill across the board. We both worked at JPL in person and it was just a minute away, on the way to work. I was able to visit at lunch to breastfeed. A fortunate accident of arriving on the CEC after our simple search was that as first time parents, we’ve since adopted much of their approach. They used a combination of RIE (Resources for Infant Educarers) and Montessori style methods as the children grew, as well as a focus on outdoor play even from the youngest age. Some key takeaways from their approach that we’ve consistently adopted:

• Treating even the youngest infant as a unique human being, not as an object. This means talking to them, walking them through procedures, and sensitive observations of them to understand needs
• Free movement vs seats, walkers swings or bouncers
• Increasing independence through independent play and with meals (drinking early from cups, self feeding).
• Open-ended, safe, passive toys.
• Spending lots of time outside, with adequate sun protection.

We practice this at home too. Almost all of our house is freely accessible to all of our children, with adults only stuff kept either in a dedicated room or on higher shelves. Mattresses are initially on the floor and don’t have cot sides, only the stairs have gates for crawlers who haven’t yet learned how to safely descend. For infants, free movement means plenty of time on the floor where they can navigate independently. We also set things up generally such that the children can roam. There are no locks (outside of on unsafe areas like our office or garage). And after graduating from a bassinet, the children sleep on floor and later elevated beds they can get in and out of independently.

Although broader data doesn’t support this approach providing uniquely better outcomes than alternatives, thinking of the family as a system there are many key advantages for our family, that have shown up early but been enhanced as our family size grows.

• Treating infants as humans makes caring for them more enjoyable, and makes it easier to spot their human needs and communication methods earlier
• Talking to infants helps their language development and even before you could imagine they might understand, and seems to sooth them
• Getting rid of dedicated entertainment and rather emphasizing observing (and intervening for safety) as the core parenting responsibility, makes parenting at this age more enjoyable. The infant will ask for attention when they need it, vs it being foisted upon them.
• Self feeding and drinking from cups means mealtimes are more scalable.
• Being able to get out of bed by themselves is easier for potty training and sibling interaction, and entertaining themselves on the rare occasion they wake before us.
• The focus on simple toys means a house with aesthetic, useful, playthings that stand the test of time and span ages, appealing to our kids as they grow and even us as adults. Our favorites that begin at this age are magnetiles, blocks, a railway set, and duplo.

Why we switched to a nanny

With our subsequent children, we both had moved on from JPL; I worked remote and Casey quickly transitioned into an in person role leading his startup. Meanwhile, our income had grown along with the size of the family, so the costs and conveniences of a nanny came much closer to childcare, which ran us about $4000/month. We were able to hire a full time 1:1 caregiver, focused on ages 0-2, but available to assist with the other ages when out of school. With two (and soon to be more) children this more effectively supported us going back to work, was of the utmost convenience, was extremely high quality, and while expensive was within our means. The more children we had, the more this method made sense financially, logistically and emotionally, so we’ve “signed” our caregiver for the next 5 years or more.

Advantages of 1:1 care for our family system:

• Attention sustainability: Neither Casey nor I can focus joyfully on an infant for their full waking hours. A professional who has chosen this calling can supplement our active attention.
• Sibling time: When our nanny’s hours overlap with other children’s schedules, we can spend 1:1 time with the older kids.
• Breastfeeding without pumping: Working from home with the infant at home made this seamless.
• Zero transition friction: Dropoffs and pickups take seconds, no packing, no drama with a steady caregiver.
• Household support: During nap times, our nanny helps with laundry, dishes, light cleaning, and inventory management. When we’re off work, we focus on children instead of chores.
• Built-in backup care: As older children attend school, the nanny handles dropoffs, pickups, and care during scheduled or unscheduled days off.
• 99%+ work reliability: With kids in mixed settings, the chance someone is ill or off school is high. The nanny provides backup, so we consistently meet work commitments.
• Upside for larger families: Without the short term difficulty of sole caregiving, we revel in family life, rejoice in our children, and want to have more! As we have more children, our childcare cost does not grow significantly. This means that once we’ve built this into our budget, having additional kids doesn’t substantially impact it further.

Infant Care Conclusion

Even considering the educational benefits of our childcare center, having experienced both, I’d choose a nanny for infants given the option. The emotional benefits for infant and parent, plus the system-level advantages, outweigh the learning benefits for first-time parents.

Recommended reading: Janet Lansbury’s Elevating Childcare (use it as a menu of options, not a rigid philosophy) and Loose Parts.

Preschool

While the data is mixed whether communal care before preschool has positive, neutral, or negative impacts on outcome, some preschool starting around age two or three will improve the ease of transitioning to school (Cribsheet). Preschool could have a positive or negative impact on the family system depending on the setup. We again were influenced by our early childcare setting near work, which focused on outdoor play and child-led exploration. Some of the ways in which it aligned with our family’s focus were:

• Social focus over academics: No drilling on letters, reading, or math. Our kids absorb academic content at home: we’re doing math, sounding out letters, and reading together. What they needed was learning to interact appropriately with peers and adults. Casey and I are unusual adults who didn’t have great social skills at that age, so this was the value-add. Plus, once academic focus starts for bright kids who are already “ahead,” it can be difficult to study alongside peers months or years behind. This has to start sometime, but there’s no reason it needs to before kindergarten.
• Child-led exploration: with familiar routines and lots of self-directed play
• Supervised risk-taking: cooking, climbing trees, jumping from heights
• High quality staff with low turnover
• No screens or “blinky” toys
• Authoritative style:  Authoritative parenting combines high expectations with high responsiveness: setting clear boundaries and rules while being warm, supportive, and responsive to the child’s needs and emotions. In contrast, other common styles include authoritarian which is high expectations/low responsiveness, and permissive is low expectations/high responsiveness.
• Close proximity: 5-10 minutes drive or walking distance, so we don’t spend our lives in the car

How we found it: The NAEYC filter

When we needed something more local, we initially struck out. LA has social pressure for “feeder preschools” among our peer group, plus options with very different focuses, rigid schedules or heavy academic drilling.

In researching, I discovered our original preschool was NAEYC accredited. NAEYC accreditation is essentially a “gold standard” certification that preschools earn by meeting rigorous quality benchmarks focused on developmentally appropriate practices. Using this as a filter led us to a fantastic local option that happened to be a co-op.

The co-op advantage

In a co-op, parents volunteer in the classroom roughly three hours per week to supplement two lead teachers. There are day-consistent volunteers (I always took Wednesdays) with more responsibilities, and floaters who keep a general eye on things and escalate to permanent staff for anything major.

Compared to our previous preschool, the overall staff quality was higher especially post COVID. Parental volunteers were highly educated, empathetic parents who shared our philosophy (hence ending up at the same place), rather than entry-level staff. The permanent teachers had generally been there 10+ years. The atmosphere was more DIY, but they were excellent stewards of parent money, focusing on what matters for development over Instagram aesthetics.

Working from home with teams across the US, I shifted my schedule to start on East Coast hours a few days a week, stopping at close of business Pacific to accommodate Wednesday volunteering. Casey, lacking this flexibility, fulfilled his volunteer hours doing construction projects for the school.

Volunteering was sometimes tough work, but I got to see my kid and benefited from informal education in early childhood development watching the pros navigate the classroom. I’ve worked as a consistent volunteer in the 4-5 age classroom, the 3-4 age classroom, and soon the 2-3 age classroom. I’ll have learned all preschool ages firsthand.

Preschool Care Conclusion

Attending some preschool before formal schooling is probably good, but the data isn’t strong enough to stress if waiting makes more sense for your family. If you want a feeder school for private education or don’t have niche requirements, any school with good staff is probably fine.

If you have similar priorities to ours: play-focused, developmentally appropriate, searching for NAEYC accredited schools is an excellent starting point. Picking a school within 5-10 minutes of home has significant family-system advantages. If you can volunteer or participate in a co-op, the education in your own parenting development is valuable, and I believe investing time to volunteer at this stage benefits both parent and child more than volunteering at later stages.

Recommended reading: Cribsheet by Emily Oster

Kindergarten and Elementary: Maximizing the “Typical School Experience”

Our kids track 1-4+ grades above in reading, math, and other core academics, with the gap seeming to accelerate over time. They remain grade-level in handwriting, social skills, executive functioning (homework and test-taking diligence), and rule-following.

Why we chose public school while we still can

We want to take advantage of the time when this gap is manageable to focus on grade-level skills and give them a “typical school experience” as long as possible. We expect their schooling will need to deviate from typical more in the future.

There’s a perception that parents drive the acceleration of kids tracking above grade level. My experience is the opposite: kids like this thirst and hunger for knowledge, so acceleration is kid-driven. You’d no more deny it to a child thirsting for it than you would a glass of water.

Our values driving our elementary school choice have been

• Convenience: Close to home with minimal driving.
• Typical experience: Age-similar peers, similar routines, focus on basics over bells and whistles.
• Flexibility: Amenable to various options for academic acceleration so we can keep this mode of education as long as it makes sense.
• Quality: Basic quality of educators and system, with good odds of matching with teachers who work well with our kids.
• Cost: This is where our systems thinking diverged from preschool/infant care. Investing heavily in a nanny or preschool makes life easier on the whole family: worth every dollar. Private elementary school wouldn’t increase reliability or convenience for us while offering marginal educational benefit and decreased normalcy. A nanny benefits multiple children for approximately the same cost; private schooling scales up with each child. When private school offers significant educational benefit and normalcy is no longer possible anywhere, we might reconsider (stay tuned!).

How it’s working

We chose the local public school. It’s walking distance or a short drive, has good ratings (matched by our kids being paired with great teachers), and is approximately free. All teachers are trained in gifted education, and while we don’t expect their clustered grouping approach (as opposed to academic streaming) to be especially useful, teachers have been flexible and creative about differentiated options. Our eldest does math warmups two grades ahead and accesses curriculum three grades ahead during math study time. Our younger gets books tailored to her reading level, alongside kids in her classroom receiving more or less advanced material during reading time.

They’ve made friends, including some similarly curious kids. Much of their school time involves moving between activities, recess, and enrichment classes: art, gardening, PE, music, classroom runs, alongside handwriting, spelling, and socialization with age peers. That is to say, the majority of their day doesn’t require differentiated curriculum, and the portion which does they blend in well with others receiving the same (at higher or lower levels).

The homework approach

There have been hiccups. We weren’t paying attention to kindergarten homework with our eldest until he complained math was boring. When I looked, I realized we needed to give him challenges (see Enrichment section below). After providing after-school opportunities to learn math in a structured way, he hasn’t complained since. In later grades we paired with his teacher to provide some of these opportunities at school as well.

Homework continues to be straightforward for both children. We debated requesting or inventing more challenging homework but decided against it. Instead, we go deep on the homework they do have, focusing on executive functioning, diligence, and improving at-grade-level skills. The math may be straightforward, but writing answers clearly, checking work carefully, paying attention to details, and completing and turning it in on time may not be. Since it is straightforward, this typically doesn’t take much time for anyone involved, and the time invested seems worth it for the returns.

Why this works for our family system

Public school requires no driving logistics that multiply with family size, and costs nothing while our children can still thrive there. The kids are picking up key social skills in additional to their educational advances and have the opportunity to experience a rite of passage shared by most Americans.

Recommended reading: The Family Firm by Emily Oster

Enrichment and Extracurriculars: Prioritizing Family Time What we skip and why

We prioritize family time together, keeping weekends largely free. This means we can spend time as a family, taking advantage of the big family we’ve built, while keeping logistics manageable. Carting one child around from activity to activity can be overwhelming; carting a large family to individual activities can be prohibitive.

As simple as it sounds, this choice seems fairly unique among our friend group. As trivial an objection to scaling a family as it might seem (“Johnny never had a sister because we wanted him to play t-ball”), I get the sense it plays a huge factor when people with one or two heavily-scheduled kids say they could never imagine having a bigger family.

We opt out of group sports at elementary age. Families with sports as a core value would make different decisions. We do physical activities together: climbing, biking, hiking, walking, and physical play. We’ve had the fortune to move to a walkable community with sports options, so as the children grow older they can walk themselves to practices in areas of their interest.

What we offer at home

We focus enrichment activities on what’s available at home.

• Piano A teacher comes once a week to our house and teaches our kids. We are a musical family and value the kids having piano as a foundation to electively build upon.
• Math enrichment. While I prepare dinner, the older kids have the opportunity to self-study from books and online curricula. I’m available to mentor. It’s gotten to the point where my eldest enjoys mentoring my younger, which also cements concepts for him: win-win-win. We use a combination of:

1. Beast Academy: a math curriculum with comic books and online/offline options. It’s rigorous, creative, allows self-study, and is fun for our kids.
2. Adventures with Mr. Math: Zoom classes focusing on using analytical reasoning skills to solve difficult math problems and puzzles. This doesn’t teach math but rather teaches problem solving. The homework and coursework provide an awesome challenge.
3. Custom problems: We dive deep on exciting problems together some of which are hard for us to solve as mathematically advanced adults. The kids don’t always have the tools in their toolbox, but we approach them together.

• Chess: Our eldest takes group chess classes at any level through International Chess Academy via Zoom. This wasn’t in the plan, but he participated in a chess activity at his homework club and when it was discontinued requested to find another option. We were delighted to stumble upon one online that has worked well
• Strewing. We participate heavily in “strewing” both accidentally and on purpose:

Strewing is the intentional, casual placement of interesting materials (books, puzzles, art supplies, science kits, etc.) around the home environment where children will naturally encounter them. The key is that it’s done subtly—you’re creating an enriched environment without directing or requiring the child to engage with the materials.

We have a great variety of material on diverse topics at home and regularly visit the library to check out and strew more. As a result, we regularly hear “I’m sorry I got stuck in a book!” when following up on why a key activity (getting dressed, taking a shower, putting on shoes) wasn’t completed in a timely manner. There are worse problems to have.

A Brief Note On Screen time

We take a selective approach to screen time, emphasizing tools over passive consumption, and focusing on moderation. The thinking: learning to use these powerful tools is a modern skill, and in the right contexts, they can function as an Illustrated Primer in service of personal growth and education. The kids use screens while in view.

• Outside the home: No screens. Including none on roadtrips.
• Passive media at home: We watch together occasionally: rocket launches, documentaries, Veritasium, chess videos, Mark Rober (limited: more edutainment than education), and other media of broad interest that is education adjacent. The kids don’t watch media alone.
• Active learning with screens: Starting around age two, we introduced limited computer or iPad time.

Useful apps and sites we’ve found:

• Ages 2-4: Khan Academy Kids on iPad. Free, ad-free content mixing fun with education. Extensive library of books with read-aloud option available offline.
• Ages 4+: Beast Academy online options. The online version is particularly engaging and offers read-aloud mode for pre-fluent readers. The kids use the books as well, but when given the choice gravitate toward the computer version.
• Post-fluency readers: Scratch.mit.edu for coding concepts, Replit for AI guided programming (limited, more edutainment than education), Google Colab for advanced coding concepts, Onshape for 3D printing, ChessKid for chess, Google Docs for typing and notes, Gmail for communicating with grandparents.

Finding resources: the Davidson community

I have a friend whose children seemed so similar to ours, we swapped educational tips frequently, and I have her to thank for stumbling upon many. On top of that, she introduced us to the Davidson Young Scholars community which offers a variety of free resources to families with gifted children as well as a community forum. Since her kids were eligible and seemed so similar to ours it was no surprise when we qualified as well. This has been helpful in seeing that there is no perfect educational fit for similar kids, as well as discovering new potential enrichment areas.

Bottom line on enrichment

Our approach maximizes convenience (everything at home or via Zoom except one piano teacher visit), keeps weekends free, and scales well with family size. Kids pursue what interests them without the pressure or logistics of scheduled activities all over town.

Recommended resources: Beast Academy, Adventures With Mr. Math, International Chess Academy, Davidson Young Scholars Program and Forums

Beyond Elementary: Our Evolving Plan

We haven’t crossed this bridge yet as our children are solidly thriving in infant care, preschool, and elementary school. But based on the Davidson Forums and our research, beyond elementary is where things get complex, requiring a tapestry of approaches with none of the off-the-shelf options (whether private or public, gifted or not) being a perfect fit.

Our current thinking

We’re inspired by research including Bloom’s two sigma problem, the finding that students taught one-on-one with mastery learning perform two standard deviations better than students in traditional classrooms. We’re also drawn to the observation that high-quality 1:1 tutoring alongside self-directed learning can be significantly less expensive than private schooling.

We’re not setting aside money for private school at the middle school, high school, or college level. In our experience, this can be a family size limiter (“Jane never had a younger brother because we wanted her and siblings to signal high status by attending $40k/year schooling for a decade”). We expect to prefer investing that capital differently and forego status signaling, focusing on educational outcomes instead.

That said, for the right educational outcomes, we’d consider investing. Our priorities will likely be:

• Actual learning over credentialing: Brian Caplan’s The Case Against Education argues that the majority of education’s value at upper levels comes from signaling versus actual learning. We’re inclined to focus on the learning.
• Kid -driven and kid-led: We expect our children’s education to be driven by our children’s needs, not our own desires for status or projections of our desires onto theirs.
• Mix and match approaches: We expect to weave together coursework at various levels, 1:1 tutoring, self-directed study, and possibly dual enrollment or online options.
• Maintaining social connection: Whatever we do, we want our kids to have peers and community, not just academic advancement in isolation.
• Flexibility as needs evolve: What works at 12 may not work at 15. We’re prepared to adjust.

We don’t have it all figured out

We’re learning from families a few years ahead of us and staying open to creative solutions. I’ll write an update to this blog when we have more life experience to back up our initial thoughts.

Recommended reading: The Case Against Education by Bryan Caplan

Executive Summary

• Infancy: Nanny or 1:1 caregiver if you can swing it. Daycare also works fine developmentally, pick one you can learn from and have backup childcare options.
• Preschool: NAEYC accredited preschool close to home. Prioritize social skills over academics.
• Elementary: Local public school if it matches your values. We considered this when choosing where to live.
• Enrichment and Extracurriculars: Prioritize family time over individual extracurriculars. Offer optional academic enrichment at home. Thoughtful, limited, supervised screen time.
• Beyond Elementary: Mix and match schooling options focused on actual learning over credentialing. Weave in coursework at various levels and 1:1 tutoring rather than private school.
• Key Resources: Elevating Childcare, Loose Parts, Cribsheet, The Family Firm, The Case Against Education, Beast Academy, Adventures With Mr. Math, International Chess Academy, Davidson Young Scholars

Printable PDF of this post. 

“Space is big. You just won’t believe how vastly, hugely, mind-bogglingly big it is. I mean, you may think it’s a long way down the road to the chemist’s, but that’s just peanuts to space.” Douglas Adams, The Hitchhiker’s Guide to the Galaxy

Our vision for the future has humans traveling between planets much faster than our ancestors sailed across oceans, but no existing rocket technology can achieve that. We’re going to need something significantly more energetic, and antimatter is the key. 

In August 2024, I wrote a primer examining the merits of antimatter propulsion from first principles. In this post, I will lay out in some detail a specific plan for an antimatter “Manhattan Project”. If we don’t do it, someone will. 

Credit: Avatar: The Way of Water. Canonically, these spaceships use antimatter fusion drives. 

Why?

Chemical propulsion is the standard for launch today, while some satellites use electric propulsion for station keeping. 

Credit: Trevor Mahlmann for SpaceNews

Starship is terrific but it’s not capable of flying to a nearby star at 30% of the speed of light and then landing. For that we’re going to need something far more energetic, and we are fortunate to have it.

Antimatter is powerful because its embodied energy gets to use the equation.

E = mc2.

c2 is a very large number, approximately 1017 = 100000000000000000. When antimatter encounters ordinary matter, it annihilates completely converting a tiny amount of mass into a huge amount of pure energy. This is 100-1000x more energy than even the most energetic nuclear fission reactions. 

With rocket propulsion, the distance you can go is determined by the change in velocity, Δv (“delta vee”), you can achieve with all the fuel you brought. The Tsiokolsky rocket equation boils down to Δv ~ 2 ve, the exhaust velocity, while thrust is given by the mass flow times the exhaust velocity, T = ṁ ve. Exhaust velocity is usually expressed as specific impulse (Isp = ve/g) and measured in seconds. It is the amount of time a given propellant can generate 1g of thrust. A more powerful propellant can provide the thrust for longer.

This graph shows the rough domains of existing and hypothetical propulsion systems. Chemical can generate terrific thrust, but is limited by low specific impulse. Electric propulsion can achieve much higher specific impulse, but is plagued by low thrust. This applies also to nuclear electric propulsion systems, which enjoy all of the hassles of nuclear reactors and still don’t achieve the desired high thrust, high Isp, high power operating mode. 

Credit: Adapted from graph by Frans Ebersohn. 

An antimatter rocket cycle can bridge this gap. Like chemical propulsion, there’s no upper limit to thrust. And given that the default antimatter reaction product is hard gamma rays, there’s no real upper limit to Isp either. If humans ever find a way to cross the gulf between stars, it will be with antimatter powered propulsion.

How 

I could spend another 10,000 words singing the praises of antimatter propulsion, but if you’re not bought in at this point, why bother? Let’s focus on the how.

Usually, when talk begins of exotic propulsion methods, discussion immediately centers on particle accelerators and superconduction magnets. Hold it right there! We’re trying to launch this on a rocket. Let’s conceptualize around a Starship upper stage, so we’re talking 1000 T of propellant, 100 T of structure, and 10 T of engine. Launch is a dynamic environment, which means everything needs to be able to withstand shock and vibration. I like particle accelerators as much as the next guy, but let’s begin by deleting as much complexity from the critical path as possible. 

Let’s discuss the various parts of the antimatter problem that need to be solved: Production, storage, and use. 

Production

The model for antimatter is that it is produced on Earth using the power and skill of our entire industrial base. Like aluminum, but to a far greater extent, it exists as an extremely condensed form of stored energy that can then be readily transported into space. We cannot easily lift the entire grid of the US into space, but its 1.3 TW capacity, run for an entire year, condensed into antimatter, would weigh just 227 kg (less than 500 lbs), which is well within our launch capacities. 227 kg of antimatter is also easily enough to launch hundreds of enormous spacecraft to nearby stars, so we will begin with a somewhat more modest quantity. 

As of late 2025, humanity is able to produce antiprotons and antihydrogen in the thousands of atoms per day and millions in total. This is incredibly impressive even by the standards of a decade ago, but it’s roughly analogous to our plutonium production capacity in late 1940. We have a ways to go here.

Antimatter production is something like 0.000001% efficient. It requires quite large particle accelerators and vacuum storage rings. 

The good news is that even at this efficiency, I think it’s worthwhile to scale up production for deep space propulsion, which is astoundingly expensive and profoundly limited by default. Remember, getting some marginal Δv when you’re a long way from home is essentially completely inelastic. There are no other options, and the solar system is the size it is. 

Obviously the utility is enormously increased if the production cost can be brought down, so the even better news is that it’s hard to imagine ways of making it less efficiently than we already do. We’re very early. We’ve barely even begun to think of ways to do this better. For example, CERN recently demonstrated 8x higher production efficiency, with a fairly obvious hack. That’s almost an entire order of magnitude. Three or four more advances like this and we’ll really be getting somewhere.

Currently, antimatter is made by bombarding a tungsten target with a high energy particle beam, which produces a few antiprotons. Then, if they happen to be at the right energy and going in the right direction, we can capture and store them for a while, gradually slow them down, and combine them with antielectrons (positrons) to form neutral antihydrogen. When I was a child, no-one was even sure if this was possible. Even today, there is serious work underway to investigate if gravity works on antimatter the same way as on normal matter. Perhaps it doesn’t!

As told in “The Making of the Atomic Bomb”, Leo Szilard had been obsessed with nuclear power and weapons for many years. In 1938, he realized that uranium, uniquely of the naturally occurring elements, could support a chain reaction. This led to a letter signed by Einstein and delivered by Alexander Sachs to FDR on October 11, 1939. Even then, the Manhattan Project wasn’t officially begun until August 1942, nearly three years later.

I think it should be possible to make antimatter with better than 0.01% efficiency, which would make high performance flight to Mars, Jupiter, and Saturn possible at scale within existing spaceflight budgets.

Humanity is on the cusp of being able to make useful quantities of antimatter, and obviously controlling the technology is strategically vital – and not just for high performance rocket engines. 

Storage

Conventionally, antimatter is stored as a charged plasma in electromagnetic “storage rings”. It is non-trivial to design containment for a form of matter that instantly annihilates on contact with any ordinary matter. Storage rings are large, heavy, and persnickety. It would be ideal to find a more robust method for storing up to milligram quantities of antimatter.

In the interests of simplicity, I think the path forward might be electrostatic containment. A small, cryogenically cold vacuum chamber (similar to devices used for quantum computers today) stores antihydrogen as a liquid droplet or ice crystal. A small net charge allows active containment in a 3D electrostatic trap. Surface charge also modulates surface tension and partial pressure, by which atomic quantities can be emitted from the surface for use, similar to a bubble jet printer head. 

Diagram showing conceptual antimatter droplet containment system. A charged drop of liquid antihydrogen is held electrostatically between actively controlled electrodes. Surface charge is modulated with an electron gun. The system is held in vacuum and kept well below 20 K. Boiloff is directed out to the engine.

Here’s Gemini 3’s image model version. Not bad!

In concept, this is a relatively small, easy to build piece of laboratory equipment. It can be tested on regular hydrogen with no special hazards, simply by inverting the charge on the containment system. A hydrogen droplet storage and manipulation system could be built from scratch for less than a million dollars. Of the three parts, it is the cheapest to test and retire risk early.

Engine cycles

Here we come to the fun part. How to actually use this incredible form of stored energy?

The fundamental problem is one of transduction – the same problem with any rocket engine. The propellants are enormously energetic, but they really want to just make heat and noise and light. How do you get them shoving mass out the back at high speed, safely and reliably?

This particular problem underscores the difficulty of making nuclear fission propulsion work. Nuclear fuel is about a million times more energetic than chemical fuel. Given that E = ½mv2, we should expect a nuclear rocket to be able to deliver about 500,000 s of Isp at high thrust. 

The only known way of doing this is via Freeman Dyson’s infernal contraption Project Orion. Yes, it achieves high Isp and high thrust, by detonating thousands of nuclear bombs behind it on its way. 

Credit: PanzerSoldat_46

Less concussive methods of production instead rely on nuclear reactors. Nuclear thermal propulsion uses a nuclear core to heat passing hydrogen, achieving an Isp of around 900 s with a steep complexity and mass penalty and lower exhaust temperature than high performance hydrolox engines such as the BE-3. Why? Nuclear materials such as zirconium and hafnium aren’t stable enough at higher temperatures. In a world of Starships and orbital refilling, it’s hardly worth the effort.

Nuclear electric propulsion instead takes a nuclear power reactor and runs electric propulsion, which can achieve Isps well into the thousands of seconds, albeit at much lower thrust. In this case, we take the incredible power of nuclear fission and tie it down with a boring Brayton cycle nuclear thermal reactor, spin a turbine, make electricity, then run that through a Hall effect thruster or similar. Each step takes a 90% cut and pretty soon, we’re once again left with some watered down weak sauce propulsion system that, at best, can achieve lower acceleration than an infinitely cheaper and easier solar sail, at least anywhere inside the orbit of Saturn. It’s not even more compact, since any space nuclear reactor needs a giant radiator to keep the cold side of the heat engine cold – a radiator that is bigger, heavier, more expensive, and more complex than a space solar array that would generate equivalent power without any nuclear reactor at all. Do not optimize something that should not exist. Delete!

Antimatter thermal propulsion

Antimatter suffers from the same problems. But maybe we can find some way to use it without particle accelerators and superconducting magnetic fields?

We can afford to “waste” almost all the inherent energy, provided at least some makes it through to the business end of the rocket and produces a high thrust, high Isp result. 

The simplest method is thermal propulsion, similar to nuclear thermal, but better. Emit a stream of antiprotons into a block of high temperature refractory, such as hafnium carbide, with a melting point of about 4000 C. The antiprotons annihilate against the block, producing a stream of hard gammas that are absorbed by the block, heating it. Then flow a propellant through. Hydrogen has low molecular mass, which creates a higher exhaust velocity. But its storage density as a liquid is about 12x lower than water, which also doesn’t require cryogenic temperatures. Personally, I’m in favor of the higher thrust and higher mass fraction of a denser propellant, and I’m prepared to sacrifice some Isp. But if you’re determined to fly a brachistochrone trajectory to Pluto and back with a human crew, other trades may apply. 

For reference, an antimatter thermal cycle running steam can produce 900-1000 s of Isp, slightly better than a nuclear thermal rocket and without needing a local nuclear reactor. Hydrogen propellant can produce around 1700 s. Total achievable Δv is about the same, at about 24 km/s. This is easily enough to fly to and from anywhere in the solar system on a Hohmann cycle (slow, efficient) orbit. Very respectable!

The principle virtue of the antimatter thermal cycle is that it’s simple. No moving parts in the hot zone, and only a simple pump to flow water through a refractory block and into space. The main downside is that it leaves a lot of performance on the table. The antimatter thermal cycle can be thought of as a uranium gun type bomb. Simple, crude, underpowered. Today, we have dial-a-yield shelf-shable thermonuclear bombs with 100x the yield that are lighter and smaller. What might be the antimatter rocket equivalent of the plutonium implosion device? 

Antimatter-catalyzed fission fragment propulsion

An antimatter thermal engine is a good start but to unlock the solar system we’re going to need a method to get to higher thrust at much higher exhaust velocity. But hafnium carbide is about as hot as solid materials can get. We need a way to get much much more energy into the exhaust gas stream without relying on heat transfer from a convenient solid. We need some kind of antimatter-fueled afterburner.

Unfortunately, simply shooting antimatter into the exhaust stream won’t accomplish much. The antimatter will annihilate, producing a bunch of gamma rays that will zoom off into the universe. They can penetrate about 10 cm through a dense solid like hafnium carbide, and about 1 km through relatively hot, sparse exhaust gasses, even at the throat before expansion. Since engines are much smaller than this, we’re going to need a mechanism to transduct the extreme energy of gamma rays into a form that can further heat exhaust gasses.

Credit: Future of antimatter production, storage, control, and annihilation applications in propulsion technologies 

We need a “kinetic cascade”. I previously wrote about this quirk of mechanics in my post on orbital debris. In that case, dense, fast-moving satellites and debris in Earth orbit are not slowed down enough by the sparse atoms of Earth’s upper atmosphere. Instead, we can launch tons of 30 micron powder into the desired orbital regions. This powder greatly increases the flux of non-destructive momentum-sapping drag-inducing collisions, filtering debris below some density threshold out of orbit and pushing it lower, where it burns up in the atmosphere. Meanwhile, 30 micron powder is sparse enough that it is materially affected by residual atmospheric drag, and also gets de-orbited in a few months. The powder creates a bridge that enables momentum transfer from big things (debris) to medium-sized things (powder) to small things (air molecules). This is a kinetic cascade.

The same principle can be applied to antimatter propulsion.

Every middle schooler in America knows that neutrons induce fission in U-235, the lighter isotope of uranium. Both U-235 and neutrons can be hard to come by, mostly for the better. What only a few middle schoolers really understand is that antiprotons can induce fission in U-238, the inert naturally-occurring form of uranium used routinely in ceramic glazes. In both cases, the result are two large fission fragments, one typically larger than the other.

The actinides, including uranium, can all be fissioned by antiproton collision.

A chart showing daughter nuclide distribution. The daughters are usually also radioactive and decay over a period of seconds-to-days into much less radioactive products – none of which are relevant over the timescales for engine propulsion, which is far less than a second. 

The important point is that one antiproton can collide with a uranium atom, producing two highly charged daughters moving at 0.05 c, instead of two gammas. Yes, the fission consumes about 40% of the antiproton’s embodied energy, but two highly charged massive and highly energetic atomic nuclei are much much easier to work with than two highly introverted gamma rays. Instead of passing through kilometers of exhaust gas, fusion products stall in centimeters of air, enabling the remaining 60% of the energy to be dumped directly into a moving gas stream. This is the afterburner!

Cherenkov radiation is actually caused by much faster particles moving through a medium faster than the speed of light in that medium, but it gives a good intuition for the scales and ranges involved.

At these temperatures, the exhaust gas becomes a dissociated plasma. Plasmas can be controlled and directed with magnetic fields, but this is extremely challenging for high thrust engines with lots of gas flow. I think that a “film cooling” approach is probably best. Dump the heat into the core of the engine exhaust, and allow cooler steam to contact the metallic parts of the engine directly. The engine bell would naturally be regeneratively cooled anyway, using the inflowing water to absorb heat prior to flashing to steam in the injector. 

The way this actually works is that there’s a water injection plate at the upstream side of the engine “combustion” chamber, and in the middle is a pintle injector able to dispense almost microscopic quantities of liquid U-238 (it’s molten at these temperatures) and very microscopic quantities of antihydrogen. They interact instantly in a tiny volume, throwing out particles around 10 cm into the surrounding steam volume, then expand out into the exhaust. No magnetic fields. No particle accelerators. No radiation shielding needed.

Diagram showing how antimatter-uranium reaction at the tip of an injector produces energetic fragments that heat an advecting layer of dense steam, superheating exhaust gasses.

The neat thing about this approach is that the relative flow of antimatter and water can be modulated. The antimatter essentially sets the engine power, the water flow cools that, setting the temperature. Higher water flow = higher thrust at lower Isp. Lower water flow = lower thrust, higher Isp. Essentially any Isp between chemical ranges (400 s) and electric propulsion (5000-20,000 s) and even beyond is possible. 

It’s even possible to reduce the amount of antimatter required for a given engine and mission profile. Enriching the uranium supply with around 20% U-235 can create localized chain reactions, with hundreds of fissions and fragments per antiproton. This actually creates an amplification effect. Instead of fission consuming 40% of the inherent energy of the antiproton, it produces a 100x the output. Either way, the exhaust will always be mildly radioactive and thus probably undesirable as the engine of a first stage Earth-launched rocket, at least until social norms around radiation exposure (other than sunbathing) changes or better, we figure out how to upregulate our DNA repair machinery and achieve something like immortality. 

The fission fragments themselves, reflected with an electrostatic nozzle, would achieve an Isp of 1,500,000 s, enough to propel a spacecraft to perhaps 10% of the speed of light. But if you are attempting to optimize for travelling to the stars, there are other cycles that may work better.

To summarize the complete cycle, a small cathode ray electron gun modulates the surface charge of a stored droplet of cryogenically cooled antihydrogen, spalling off antiprotons. They travel through an injector, meeting tiny droplets of molten U-238, potentially with some U-235 alloyed in for extra spiciness. Fission daughter nuclei zoom outwards, colliding with millions of partly dissociated steam molecules in a compact area, superheating the exhaust. The exhaust expands out generating high thrust at high velocity, enabling missions with tens, hundreds, or even thousands of km/s of Δv. 

How much antimatter do we need? It depends on how fast we want to go. For a Starship-class vehicle with 20 km/s of Δv, we need barely a handful of antimatter and U-238 – that’s enough to boil 1000 T of water to 25000 K exhaust temperature. Same vehicle but now we want 100 km/s, enough to fly to Pluto and back in less than twenty years? 10 kg of U-238 and just 45 g of antimatter, both occupying about 500 ccs. 45 g of antimatter too expensive? Mix in 1 kg of U-235 and we can make do with just 0.5 g. 

This model is truly the “Heart of Gold” for advanced propulsion. 

Here’s a basic spreadsheet that calculates performance characteristics for the antiproton catalyzed fission engine cycle. 

Alternatives

There are numerous other ways to potentially chase the dream of high thrust, high Isp engines powered by antimatter. Restricting ourselves to recognizably useful systems that violate no known laws of physics, the following changes are possible. 

Delete antimatter

Antimatter production and storage is painful – no two ways about it. Can we do without it?

Yes, but you might not like it. Instead of U-238 we can use mostly U-235 (highly enriched uranium) and drive fissions with a stream of neutrons rather than antiprotons. The energetics are much the same, but now we need a sufficiently bright neutron source. This probably requires a 1 GeV particle accelerator, but there are some ideas for compact particle accelerators that could be launched on rockets. See, for example, the AWAKE experiment. 

These are at a similar level of maturity to antimatter production. The key difference is that if we can store antimatter, we can produce it in labs on the ground, whereas a neutron-driven subcritical fission fragment propulsion system would always need an at least shipping container-sized accelerator to generate the neutrons – and AWAKE makes nowhere near enough of them. 

Delete antimatter storage

Storing antimatter is highly nontrivial and, if containment fails, beyond catastrophic. Not quite nuclear explosion level but still, the warranty is voided. 

If we’re going to design and build a flight-ready 1 GeV-class high intensity particle accelerator, why not just generate anti-protons directly instead of spalling off neutrons?

A just-in-time antiproton production system, combined with a partially enriched uranium target, could potentially scale to much larger sizes. Why? If the impinging antiproton is fast enough, it will relativistically beam the daughter nuclides in the direction of impact, removing the requirement for sufficiently dense steam and/or electrostatic mirrors. This is one potential form of a relativistic interstellar engine. 

Superconducting magnets

LK-99 was a bust, but maybe it’s possible to build enormously powerful and compact magnets. In that case, we can delete the expansion bell and use a magnetic nozzle to contain and expand the superheated exhaust to provide thrust. Like particle accelerators, this almost certainly works better at a larger scale and higher Isps better suited to travel to the outer solar system and beyond. 

Conclusion

There are a number of very smart people wondering what NASA might do, now that commercial launch and reusable rockets have been effectively incubated and the end is in sight for SLS, Orion, and the ISS. Something deeply technical, requiring deep investment, world leading expertise, and that gives humanity the next big unlock. A Manhattan Project that will give us the entire galaxy. An answer to the ambitions of adversaries who are gaining fast on our existing technology stack. 

Even incremental improvements in any one part of antimatter production, storage, and use will deliver enormous benefits to our civilization. Relatively modest improvements across all three will unlock the solar system. Finally, in the limit, having the ability to condense the power of the sun into pure energy in the form of antimatter is about as far as our tech tree is likely to go. The end is in sight!