If SpaceX provides the compute capacity, Microsoft could monetize it.
Elon Musk made some very big predictions and threw around some very big numbers during the SpaceX earnings call last week. SemiAnalysis, a research boutique, says, "We believe that number is real."
Elon Musk's first earnings call for SpaceX as a public company was notable for its grand ambitions. Using the word "conservatively," he outlined a plan for SpaceX to expand its current AI compute capacity of around 1.4 gigawatts by something like 6 to 8 gigawatts next year, with the possibility of exceeding 10GW.
One potential consequence of this could see Microsoft $(MSFT)$, currently facing an enormous compute gap this year and next as it tries to service OpenAI's frontier models, becoming the primary off-taker of SpaceX $(SPCX)$ compute capacity.
This is the theory of the research team at SemiAnalysis, an independent research and advisory boutique operating out of San Francisco. In a Substack post published August 7 they explored Musk's ambitious proposals that "shocked the world, once again," but the team of Jeremie Eliahou Ontiveros, Reyk Knuhtsen and Jordan Nanos think those proposals are plausible.
To build up to 10GW of compute capacity will require somewhere between $300 billion to $500 billion of capex by the end of 2027 but SemiAnalysis reckons Musk has ways and means of bypassing typical datacenter construction constraints. And what the team calls "large-scale, near-term compute" is such a scarce commodity that it's priced at a significant premium, something like $50 billion per GW per annum.
That sounds like a lot, and it is, but the frontier AI labs like OpenAI and Anthropic can make a "good living" off it. How much? Utilizing proprietary inference simulators and coding benchmarks, SemiAnalysis arrives at $100 billion per GW of annual recurring revenue.
Aside from OpenAI and Anthropic, however, SemiAnalysis highlights one other company that is capable of printing such economics: Microsoft . Ontiveros and team believe Microsoft can generate the same revenue and margins as the two established players, but without paying the same training costs owing to its 27% stake in OpenAI and the deal it has negotiated with it.
SemiAnalysis writes: "Put simply, Microsoft has a giant incentive to procure as many megawatts as possible, as fast as possible. The potential impact is Microsoft Azure accelerating revenue growth from around 42% at present to in excess of 100% in 2027." For Ontiveros this represents "a once-in-a-generation opportunity that SpaceX (SPCX) is well-positioned to serve."
Microsoft' demand for compute capacity is booming
The numbers and the forecasts sound, as SemiAnalysis readily acknowledges, "insane," but they emphasize they think these targets are achievable for two reasons: first, Microsoft is already preparing for an epic datacenter ramp, having signed lots of deals and expecting to sign more; second, with a 90-day cancellation policy, there is "zero balance sheet risk," according to SemiAnalysis.
Moreover, the report predicts that the tricky part of the equation - SpaceX funding the capex - is manageable through support from Nvidia (NVDA) by means of vendor financing, and also selling large-scale compute with a three-to-five month lead time.
In pre-market trading Monday SpaceX shares were trading 2.4% higher than Friday's close at $136.48 while Microsoft stock was fractionally lower at $498.47.
-Jules Rimmer
Comments