Anthropic will pay SpaceX $1.25 billion a month for compute at the Colossus 1 facility in Memphis through May 2029, a fact disclosed not in Anthropic's own confidential IPO filing but in SpaceX's prospectus [1]. The point no one is saying aloud: the second-largest frontier AI lab is now buying $15 billion of annual compute from a rocket company that owns a rival model lab, and the hyperscalers were either not asked or could not deliver.
The disclosure asymmetry matters more than the number
The $1.25 billion monthly figure is large but legible. What is harder to explain away is where it surfaced. Anthropic's S-1 is confidential. SpaceX's is public. Anthropic's single largest known operating commitment, a $15 billion annual line item running for three years, became visible to the market through a counterparty's filing [2]. Either Anthropic's lawyers were comfortable letting a rival's prospectus do the disclosure work, or they had no choice because SpaceX needed to book the revenue line to justify a $1.75 trillion valuation and a $75 billion raise [3]. Both readings are awkward, and both tell you the same thing: the compute commitment is now load-bearing for two separate IPO narratives, neither of which Anthropic controls.
Compare this to the standard hyperscaler arrangement. When OpenAI signs with Microsoft, the contract sits inside an equity-and-credits structure that obscures cash flow and binds the parties commercially. When Anthropic signs with AWS or Google, the same pattern applies. The SpaceX deal has none of that architecture. It is a 90-day-terminable cash contract with a fixed end date [4]. That is a spot-market purchase at industrial scale, which is what you do when long-term suppliers cannot meet your order book.
Why the capital stack matters
The investor list on Anthropic's May 2026 round is the part of this story that will occupy consultancies for the next two quarters. Blackstone. Brookfield. GIC. D1 Capital. General Catalyst. Insight Partners [5]. The first two are not venture investors. Blackstone is the world's largest alternative asset manager. Brookfield is one of the largest infrastructure investors on earth, with deep holdings in power generation, transmission, and data center real estate. They do not write cheques into software companies at $965 billion post-money valuations because they think this is a software company. They write them because they think the underlying asset class is infrastructure, and they want exposure to the cash flows that will eventually have to underwrite the power, land, and silicon.
That is a meaningful re-pricing of what AI compute actually is. Sovereign wealth and infrastructure capital arriving on the cap table of a model lab, at a valuation that more than doubled from $380 billion in February to $965 billion in May [6], signals that the smart money has stopped treating frontier AI as a hyperscaler-mediated business and started treating it as a direct infrastructure play. The corollary is that hyperscalers no longer sit between the capital and the asset. They sit beside it, and on some workloads, behind it.
The margin question nobody is answering
The number Anthropic has not disclosed is the one that determines whether any of this works. Nate Elliott at EMARKETER put it plainly: gross margin is the figure no one outside Anthropic has ever seen, and it is the figure that determines everything [7]. A $47 billion annualized revenue run rate against a known $15 billion annual SpaceX bill, before any other compute, headcount, or training cost, leaves a margin structure that is either spectacular or catastrophic. CNET's reporting that the AI industry has spent more than twice what it has earned, with Nvidia the only company in the black across the chain, suggests the catastrophic case is the base case for most of the field [8].
This is why the IPO timing matters. Gil Luria at D.A. Davidson said it directly: OpenAI and Anthropic are racing to go public before capital runs out [9]. Morningstar's Michael Field made the same point: public equity is now the cheapest available source of cash for these companies in a rising rate environment [10]. The implication for hyperscalers is uncomfortable. If the model labs raise tens of billions in public equity in 2026 and 2027, they no longer need hyperscaler balance sheets to fund their compute. They can write the cheques themselves, to whoever has capacity, which is less and less often the three cloud providers.
The counter-case, taken seriously
The strongest version of the bear case runs as follows. The SpaceX deal is a capacity overflow arrangement, not a shift in the supply chain's centre of gravity. Hyperscalers dominate inference, which is where the actual revenue lives. Enterprise integration, identity, security, and edge deployment remain owned by AWS, Azure, and GCP, and no model lab is going to rebuild those stacks. Alphabet's market cap moved from $2.3 trillion to $4.54 trillion in twelve months, which is not the price action of an eroding moat [11]. Sam Altman has been explicit that public listings are financing events, not strategic pivots. And EMARKETER notes Claude is "simply not competitive in consumer AI," trailing ChatGPT and Gemini on projected U.S. users, meaning Anthropic's enterprise-only posture may cap its ceiling regardless of how it sources compute.
The case is real and worth holding. But it conflates two different things. Inference economics and enterprise stickiness are not in dispute. What is in dispute is who underwrites the next generation of training infrastructure, and on what terms. Meta's market cap fell from $1.76 trillion to $1.55 trillion over the same period that Alphabet's doubled [12]. Public markets are not uniformly bullish on hyperscale AI capex; they are discriminating. And Microsoft's exclusivity with OpenAI has, by its own admission, evolved as both companies look to other partners [13]. The hyperscaler moat around inference is intact. The moat around frontier training capacity is the one losing ground, and the Anthropic-SpaceX contract is the visible evidence.
The bear case also underestimates how quickly the capital base for AI infrastructure is changing. When Brookfield underwrites a power-and-compute campus directly, AWS is not the intermediary. When SpaceX's Colossus 1 sells $15 billion of compute to a rival lab, Azure is not the intermediary. Each of these is a single arrangement. Three or four more, and the question of who sits at the centre of frontier AI infrastructure becomes genuinely open.
What to watch
1. Anthropic's S-1 gross margin disclosure when it goes public. A clearly positive gross margin makes the SpaceX contract affordable and the bypass model durable. A negative or thin gross margin means the IPO is a cash rescue and the thesis weakens materially. The number will be public within 90 days of the S-1 going non-confidential.
2. Whether any of the three hyperscalers announce a counter-arrangement with a non-hyperscaler compute supplier in the next two quarters. A Microsoft or Google deal with an independent GPU cloud or a sovereign data center operator at multi-gigawatt scale would confirm that hyperscalers themselves have concluded their internal capacity cannot meet frontier demand. Silence confirms the opposite.
3. The post-May-2029 disposition of the Colossus 1 contract. If Anthropic renews or extends before mid-2028, the SpaceX relationship is a fixture. If the contract runs off and Anthropic returns to AWS or Google for equivalent capacity, the bypass was tactical and the hyperscaler moat holds. The 90-day termination clause means the signal will arrive earlier than the formal end date.
Sources
[1] https://www.cnbc.com/2026/06/01/anthropic-ipo-s1-prospectus.html
[2] https://www.cnbc.com/2026/06/01/anthropic-ipo-s1-prospectus.html
[3] https://www.reuters.com/business/ai-giant-anthropic-confidentially-files-us-ipo-2026-06-01/
[4] https://www.cnbc.com/2026/06/01/anthropic-ipo-s1-prospectus.html
[5] https://www.reuters.com/business/ai-giant-anthropic-confidentially-files-us-ipo-2026-06-01/
[6] https://www.theguardian.com/technology/2026/jun/01/anthropic-ai-ipo
[8] https://www.cnet.com/tech/services-and-software/anthropic-ipo-first-filing-ai-revenue/
[9] https://www.reuters.com/business/ai-giant-anthropic-confidentially-files-us-ipo-2026-06-01/
[10] https://www.ynetnews.com/business/article/bkkdmt0lfg
[11] https://www.ynetnews.com/business/article/bkkdmt0lfg