이미지: X — AI 리더 발언 영상 갈무리
Summary
- Elon Musk posted on X that orbital computing would become the only way to scale AI by around 2029
- The remark referenced a simulation model called "AI Compute Keystone" from Mach33 Financial Group
- The model projects that by 2040, dedicated infrastructure of about 500GW would process 800 quadrillion tokens, generating $3.2 trillion in revenue from the infrastructure layer alone
- 발언자
- 일론 머스크
- 발언 시점
- 2026-08-14(UTC 16:49)
- 핵심 주장
- 궤도 컴퓨팅이 2029년경 AI 확장의 유일한 방법이 될 것
- 제시 근거
- 지상의 전력 가용성 문제·인허가 지연
- 인용 모델
- Mach33 Financial Group 'AI Compute Keystone' (2026-08-13 공개)
- 모델 전망치
- 2040년까지 80경(800,000 quadrillion) 토큰, 약 500GW 전용 인프라, 인프라 계층 매출 약 3.2조달러
- 확산 경로
- 오언 루이스(Owen Lewis)가 스레드 인용, 머스크가 이를 재게시
Musk: "We can't keep expanding on the ground"
On August 14, Elon Musk posted a short line on his X account: by around 2029, orbital computing would become the only way to keep scaling AI. He cited two reasons — power is becoming increasingly hard to secure, and building data centers on the ground keeps running into permitting delays. Musk wrote, "Power and permitting issues mean orbital computing is the only way to scale."
The remark wasn't a new claim of his own but a reference to someone else's analysis. The post Musk linked to was a thread by an account named Owen Lewis, introducing a computing infrastructure simulation model from Mach33 Financial Group. Lewis highlighted the section on orbital data centers as the most important part of the model, emphasizing the line: "Orbital data centers, treated as a niche today, will become a requirement for AI scaling in the 2030s."
Mach33's "Keystone" model maps out 2040
On August 13, Mach33 Financial Group released a market simulation model called "AI Compute Keystone." The model forecasts supply and demand for AI computing infrastructure from 2026 to 2040, notably covering both ground-based and orbital infrastructure together. According to Mach33, the model ran 5,000 simulations across thousands of assumption combinations to account for uncertainty.
| Metric | Mach33 Forecast (2026–2040) |
|---|---|
| Cumulative token throughput | About 800 quadrillion tokens |
| Dedicated infrastructure scale | About 500GW |
| Infrastructure-layer revenue | About $3.2 trillion (excluding model/app layers) |
The company said these figures represent revenue projections from the infrastructure layer alone, excluding revenue from the model or service layers built on top. At 500GW, the scale is substantial even compared to current total global data center power consumption, and whether that much power can realistically be sourced from ground-based infrastructure alone lies at the heart of the debate.
Ground-based bottlenecks — power and permitting
There's context behind Musk specifically citing "permitting issues" as a reason. His own company, xAI, has already faced controversy over unpermitted turbines at its data centers. It was confirmed that SpaceX pushed back by another year its plan to remove all unpermitted turbines installed at xAI's data center — a case that illustrates how the pace of securing power for AI data centers is now outrunning permitting procedures. The point that power has become the core bottleneck in the AI infrastructure race is underscored by this very example.
Bringing new gigawatt-scale power online on the ground requires clearing three layers of process: expanding power plants, connecting to transmission grids, and obtaining local permits. Given that each stage commonly takes years, the math suggests it would be physically tight to meet hundreds of gigawatts of demand within the 2030s. This speed limit on ground infrastructure is the backdrop against which the orbital computing scenario has emerged.
Editor's take
This single tweet is hard to dismiss lightly, because of the timing against Musk's other ventures. SpaceX has already built out a satellite communications network in Earth orbit; the same SpaceX recently delayed removing xAI's unpermitted turbines and just completed its acquisition of Cursor. The fact that the person calling orbital computing "the only way" is also the founder of the company that can launch payloads into space most cheaply invites reading this remark not just as a forecast, but as a strategic move positioning his own assets.
Viewed in a longer arc, this isn't a new idea. The concept of orbital data centers has circulated among academics and startups for years, on the grounds that solar power efficiency is far higher in space than on the ground and that the vacuum of space can be used for cooling. Until now, though, the idea kept getting deferred to "someday" because of the barriers of launch cost and communication latency. What's different this time is that "someday" now has a concrete date attached: 2029. With firms like Mach33 putting out simulations quantified at 500GW and $3.2 trillion, the discussion is shifting from speculation to calculation.
In practical terms, there isn't much domestic companies can do right now. Orbital computing requires an entire stack of satellite manufacturing, launch, and communications infrastructure to be in place, leaving little room for ordinary companies to enter. But what this debate does suggest is that power should be the very first item checked in any AI business plan. It's no longer just about how many GPUs a company can secure — increasingly, it's about how quickly it can obtain the power and site permits to run those GPUs that determines the actual launch schedule for services.
In the coming weeks, this tweet is likely to spur a wave of funding announcements from orbital data center startups citing it as justification. At the same time, friction between big tech and local regulators over ground-based power and permitting issues will likely continue to make headlines. Whether orbital computing actually arrives by 2029 remains uncertain, but this remark once again confirms that, until then, power remains the real bottleneck for the AI industry.
