SemiAnalysis explains why SpaceX could reach 10GW of AI compute by 2027

In episode 24 of SemiAnalysis Weekly, published 9 August 2026, Jeremie Eliahou Ontiveros, Reyk Knuhtsen, and Jordan Nanos examine SpaceX's reported plan to reach 10 gigawatts of AI compute by the end of 2027. That means data-centre capacity drawing up to 10 billion watts. Their case rests on scarce compute earning exceptional returns before conventional projects can arrive. They trace whether SpaceX can assemble the sites, power, chips, customers, and financing quickly enough—and where each assumption could fail.

SemiAnalysis estimates that frontier-model APIs can generate about $100 million per megawatt each year, far above ordinary long-term infrastructure pricing. That gap supports premium short-notice capacity, but the panel says the real test is whether SpaceX can finance and build roughly 8GW next year.

Watch the related SpaceX clip on X

• SemiAnalysis contrasts long-term infrastructure pricing of roughly $12 million to $13 million per megawatt with its estimate that frontier-model APIs can generate about $100 million per megawatt annually.
• SpaceX can charge more when it supplies large amounts of compute immediately, while 90-day cancellation terms reduce customer risk.
• The estimate combines measured open-model workloads with a simulation of frontier models. The panel expects newer hardware to increase token output per megawatt.
• Financing pushes conventional capacity 12 to 18 months away. A builder willing to risk temporary underuse can capture the premium when demand arrives first.
• The panel considers selling the capacity less doubtful than executing the build, especially adding roughly 8GW next year and securing enough chips, sites, and workers.

SemiAnalysis says suitable sites, turbines, reported chip orders, prefabrication, and unconventional equipment make a rapid build physically plausible. The speed comes with a trade-off: lower efficiency, redundancy, and service guarantees than a conventional data centre.

• The team's search found several large warehouses and pipeline-adjacent plots, about 7GW of unidentified turbines, and reported chip-side orders for 5GW to 10GW the following year.
• Rapidly available, pre-assembled Chinese electrical, cooling, and switchgear equipment could ease supply constraints if customers accept unconventional infrastructure.
• Labour remains a major bottleneck. The panel estimates Colossus 2 used about three times fewer peak workers per gigawatt than leading fast builders, partly through prefabrication.
• Warehouse conversions, remote generation, and private transmission could avoid parts of the normal powered-land process, even if they sacrifice efficiency.
• Customers seeking one large cluster within months may accept less electrical and generator redundancy and lower service guarantees. One speaker cautions that the lost protection still matters.

The panel names Microsoft as the clearest large customer because OpenAI-linked demand gives it a valuable immediate use for compute, while an earlier construction pause left a near-term capacity gap. SpaceX's 90-day cancellation terms could bridge that gap without another long commitment.

• Microsoft's access to OpenAI intellectual property places it among the few companies that the panel believes can earn frontier-model economics from the capacity.
• A 2024–25 Microsoft construction pause, large OpenAI commitment, and long lead times leave it short in late 2026 and early 2027 despite renewed building and leasing.
• SpaceX capacity would let Microsoft pursue the claimed high-value token opportunity immediately instead of waiting for projects due in late 2027 or 2028.
• Either party can cancel with 90 days' notice, making the service a six-to-twelve-month bridge instead of another binding long-term infrastructure agreement.

The panel argues that contracted AI capacity could generate enough operating cash to finance much of SpaceX's expansion, with Nvidia as a possible funding partner. Nobody knows the financing structure, and the plan still requires enough chips and highly profitable frontier AI tokens.

Watch the related NVIDIA clip on X

• Existing contracts covering roughly 1GW to 1.5GW could generate about $50 billion in annualized revenue at very high margins, creating cash to reinvest.
• On the panel's assumptions, monetizing 5GW at about $50 billion per gigawatt annually could produce roughly $250 billion of operating cash flow, with a GPU paying back in under a year.
• Nvidia is a plausible financing partner because SpaceX has committed to its hardware and Nvidia has reason to defend its ecosystem. The speakers say nobody knows the exact structure.
• Building power-and-cooling-ready data centres first could create more customer and GPU-financing options, although chip supply may still fall short.
• The financing works only if frontier AI tokens stay highly profitable. The panel puts current margins at 85% to 95% and does not claim they will last.