SemiAnalysis is valuing SpaceX’s AI buildout through a different lens: not how many GPUs it controls on paper, but how quickly large blocks of compute can actually be turned on. In that framework, the company may be able to charge a time premium if it can deliver hundreds of megawatts of AI capacity months ahead of alternatives.
Focus shifts from GPU count to time-to-compute
The WhiteLine article says AI infrastructure deals are often judged first by GPU totals, but chips are not the only scarce input. Data center timelines for land acquisition, power connection, substations, cooling systems, and server halls can run longer than the chip upgrade cycle. That is why SemiAnalysis centers its model on “time-to-compute” — how long it takes, starting today, for a large pool of compute to become operational.
Under that logic, SpaceX can prepare power and data center facilities first, then buy newer-generation GPUs closer to the point when electricity is ready. The company could then decide whether to use that capacity internally or lease it out based on demand. The advantage is optionality in energized capacity: it does not have to commit too early to older chips, and it can allocate near-term available compute to customers willing to pay more. SemiAnalysis estimates that delivery times for some SpaceX projects could be compressed to roughly three to five months.
Why customers may pay up for immediate access
The article says Google will pay SpaceX about $920 million per month starting in October 2026 for roughly 110,000 NVIDIA GPUs and other compute resources. It also says Anthropic previously agreed to purchase SpaceX compute for $1.25 billion a month.
Those figures are far above standard long-term cloud contracts, and the report gives a simple reason: customers cannot wait. Their own data centers are still not online, while training and inference demand is already in front of them. In that setting, what they are buying is less a conventional cloud service than time — early access to compute when it is needed, much like paying an expedited fee for spot transport capacity during a peak season.
The article also notes that this pricing may not last. Google’s agreement, once conditions are met, allows either side to terminate with 90 days’ notice. That makes current pricing look more like capacity insurance during a shortage than a level that can be extended indefinitely using long-term cloud contract assumptions.
The assumptions behind a $300 billion ARR scenario
SemiAnalysis outlines an aggressive case in which SpaceX could approach 10 GW of compute capacity by the end of 2027 and generate about $300 billion in annual recurring revenue.
That valuation depends on several conditions. SpaceX would need to keep delivery times at the level of a few months. NVIDIA would need to supply enough newer-generation GPUs. And once customers bring their own data centers online, the market would still need to support renewals for high-priced contracts.
By this reading, SpaceX’s most important AI asset is not just GPUs. It is energized capacity that can be delivered in the near term. The article says the key figures to watch next are powered GW, revenue per GW, and contract renewal timing. Only if those numbers continue to hold can the short-term premium from time-to-compute become a durable business.
The article was sourced from WhiteLine and translated by Wu Blockchain.