medium convictionactive · updated 2026-05-23T00:00:00.000Z
Terrestrial power-flat → chip-output exponential → orbital DC arbitrage
Electricity output outside China is roughly flat while chip output grows exponentially. Terrestrial solar at scale is regulatory-blocked. Space solar delivers ~5x output with no batteries and a permissive regulatory regime. Therefore the cheapest place to put AI compute in 30-36 months is orbital — even with the GPU-servicing cost penalty.
The chain
1
Electrical output outside China is essentially flat; only China is adding power capacity at scale.
elon-musk in 2026-02-05-dwarkesh-patel-elon-musk-in-36-months-the-cheapest-place-to-put: "if you look at electrical output outside of China, everywhere outside of China, it's more or less flat. It's very. Maybe a slight increase, but pretty close flat. China has a rapid increase in electrical output"
2
Chip (compute capacity) output grows exponentially. The chip vs. power gap is the binding constraint on AI buildout — not silicon, not capital.
elon-musk in 2026-02-05-dwarkesh-patel-elon-musk-in-36-months-the-cheapest-place-to-put: "the output of chips is growing pretty much exponentially, but the output of electricity is flat. So how are you going to turn them to chips on? Magical power sources. Magical electricity fairies"
3
Terrestrial solar at scale is regulatory-blocked. Even covering Nevada in solar panels requires permits that don't exist.
elon-musk in 2026-02-05-dwarkesh-patel-elon-musk-in-36-months-the-cheapest-place-to-put: "I think it's pretty hard to cover Nevada in solar panels. You have to get permits from. Try getting the permits for that"
dwarkesh-patel in 2026-02-05-dwarkesh-patel-elon-musk-in-36-months-the-cheapest-place-to-put: "So space is really a regulatory. It's really a regulatory play. It's harder to build on land than it is in space"
From 2026-05-12-cnbc-television-ron-baron-on-upcoming-spacex-ipo (Ron Baron, broadcast captions normalized to sentence case): "instead of having data centers on our planet where it's very, very difficult to get — the governor [of] New Jersey, because she said, our electricity bills are too high, we're going to stop this data center construction, making our electricity bills go up. Her roommate, when she was in Congress, is a candidate for governor in Virginia. She won. Same same idea."
From 2026-05-12-cnbc-television-ron-baron-on-upcoming-spacex-ipo: "people don't like it because not enough jobs and because the electricity bills are going up." Plus "water is short and electricity is a problem."
From 2026-05-08-all-in-podcast-elon-s-anthropic-deal-the-next-ai-monopoly (Chamath): "you have about 9 gigawatts that are supposed to come online this year. Almost 50% of it now is being protested." Quantifies the regulatory-block / NIMBY thesis for the *current* US 2026 buildout — half of planned capacity facing organized opposition. Reinforces Step 3 by tying it to a current-year number rather than a long-run trend
4
Space solar delivers ~5x more output per panel (no atmosphere, no day/night, no clouds) and eliminates the cost of nighttime batteries. Regulatory regime is permissive relative to terrestrial.
elon-musk in 2026-02-05-dwarkesh-patel-elon-musk-in-36-months-the-cheapest-place-to-put: "you're going to get about five times the effectiveness of solar panels in Space versus the ground. And you don't need batteries... the atmosphere alone results in about a 30% loss of energy. So any given solar panels can do about five times more power in space than on the ground. And you avoid the cost of having batteries to carry you through the night. So it's actually much cheaper to do in space"
From 2026-05-12-cnbc-television-ron-baron-on-upcoming-spacex-ipo (Ron Baron, broadcast captions normalized): "you don't need the cooling in space, number one, as long as you have giant radiators. And then electricity, which is a huge cost, you don't need that either because you're using the sun. So basically you're getting free, free electricity and free cooling once you get into space."
5
Therefore the cheapest place to put AI compute in 30-36 months is space. The GPU-servicing penalty (no on-orbit repair) is real but absorbed by GPU reliability past initial debug. Scaling beyond 1 TW of compute can only happen in space.
elon-musk in 2026-02-05-dwarkesh-patel-elon-musk-in-36-months-the-cheapest-place-to-put: "in 36 months, but probably closer to 30 months, the most economically compelling place to put AI will be space. And then it'll then get ridiculously better to be in space. And then the scaling. The only place you can really scale is space"
elon-musk in 2026-02-05-dwarkesh-patel-elon-musk-in-36-months-the-cheapest-place-to-put: "the reliability is. Actually, they're quite reliable past certain point. So I don't think the servicing thing is an issue"
From 2026-05-12-cnbc-television-ron-baron-on-upcoming-spacex-ipo (Ron Baron tracing the operator chain from reusable rockets through Starlink to orbital data centers, broadcast captions normalized): "so he comes up with the idea that you'll be able to launch a rocket and reuse it. And no one ever did that before… that's what enabled us to have Starlink… then we decided, hey, wait, there's a demand for SpaceX data centers." Plus on Starship as the next step: "because of the idea of going to space, we've now developed the Starship. And the Starship, once that works, is going to be 100 times…" [transcript cut].
From 2026-05-08-all-in-podcast-elon-s-anthropic-deal-the-next-ai-monopoly (Chamath, on SpaceX IPO valuation): "the biggest element [in the SpaceX bear case] is the on the come value around the orbital data centers. And by actually landing a bunch of terrestrial capacity, I think you start to blunt that because you can now start to say that even if the orbital data centers get delayed by a few months or a few quarters, even if the technological de risking of it takes longer, he now has a structural core business that will effectively subsidize his ability to train Grok." **Important framing shift for this chain: the Anthropic-EWS deal does NOT invalidate the orbital-DC thesis — it provides a structural cash-flow bridge that absorbs schedule slippage on Starship/orbital while the orbital deployment de-risks.** "Elon Web Services" becomes the near-term tradeable proxy; orbital DC remains the long-term tradeable proxy. Both are SpaceX-internal so the SpaceX IPO captures both legs
gavin-baker in 2026-05-20-podcast-invest-like-the-best-gavin-baker-watts-and-wafers-invest-like-the-best: **"I do want to reframe Orbital Compute because I think when people hear Data Centers in Space, they picture a Pentagon sized building in space. They're like, well, we can't do that. That's not what it is. A Blackwell rack weighs 3,000 pounds, it's 8ft high, it's 4ft deep, 3ft wide. It's racks in space... You keep it in a sun synchronous orbit. So those solar panels are always in the sun."** Independent practitioner (Atreides CIO) corroborating Musk's thesis but with a sharper engineering frame: it's not data-centers-in-space, it's racks-in-space connected by laser links through vacuum. Starlink V3 already operates at 20 kilowatts; a Blackwell rack is 100 kilowatts; SpaceX "seems very confident they're going to go right to 100 to 120." This is the most concrete deployment shape of the thesis articulated by a non-Musk operator-investor. Conviction upgraded from low to medium on the Baker corroboration.
gavin-baker in 2026-05-20-podcast-invest-like-the-best-gavin-baker-watts-and-wafers-invest-like-the-best: **"Starship is going to change the space economy in ways we cannot imagine. And particularly if regulation becomes a constraint to data centers, none of it's going to matter. You're going to sell as much orbital compute as you can make."** Baker explicitly endorses the regulatory-block-forces-orbital framing — Step 3 is operative. Per Baker, the Watts shortage "will probably begin to alleviate 27, 28 and then I think Orbital Compute will really solve that" — providing a timeline for the transition from terrestrial bottleneck to orbital-supplemented compute.
What would falsify this
- Step 1: If US/EU baseload power capacity grows >5%/yr through 2027 (nuclear, geothermal, fast solar permitting), step 1's 'flat' framing is wrong.
- Step 4: If orbital DC pilots fail thermal/radiation hardening (Starcloud, etc.) within 24 months, the chain's economics premise breaks.
- Step 5: If no commercial orbital DC reaches GW-scale by end of 2028, the 30-36 month timeline is falsified regardless of long-run thesis.
Contradictions / tensions
- Aggressive 30-36 month timeline conflicts with Starship's track record of schedule slippage.
- GPU-servicing question: Elon dismisses it, but extending GPU MTBF to 5-7 years (typical orbital DC lifetime) is unproven.
- Capital cost of orbital launch + radiation hardening + thermal management may dominate the 5x panel efficiency win at small scale.
- **Source-independence caveat on the Baron corroboration.** Ron Baron is Baron Capital's CEO and an early SpaceX investor (~$15B current position per the same source). He restates Musk's argument largely beat-for-beat — independent voice on the framing but not on the underlying thesis. The NIMBY-politics evidence is genuinely new and concrete; the rest of Baron's contribution is Musk-aligned reinforcement and should be discounted as such for conviction purposes.
Implications
- Speculative-bullish for SpaceX-adjacent supply chains (launch cost is the gating cost; Starship economics are the variable). No clean public stock proxy today; **the pending SpaceX IPO becomes the public proxy** once it lists — see spacex and spacex-ipo-index-inclusion-mechanic.
- Bearish for over-built terrestrial DC providers IF the timeline holds — but 30-36 months is aggressive and the consensus prices in terrestrial-only scaling.
- The **NIMBY-politics evidence** Baron added in May 2026 (NJ + Virginia governors blocking data center construction on electricity-bill politics) is a new and concrete mechanism for *why* step 3's regulatory block actually bites. Permitting was the prior framing; political bills-blamed-on-DCs is the operative force.
- Reinforces the AI-power-bottleneck thesis — the chain treats power, not silicon, as the binding constraint, which is consistent with the nuclear-baseload concept.
- If true, weakens Step 3 of the Taiwan-chokepoint chain (frontier-AI silicon bottleneck) — orbital compute could relax the power constraint that makes the silicon constraint binding.
Companies
Concepts
Open questions
none