SpaceX's IPO filing lays it bare: Anthropic pays Musk's xAI $1.25B a month for compute — while xAI loses $2.4B a quarter

What the filing disclosed

As part of taking SpaceX public, its S-1 registration statement put on the official record a set of facts that had circulated as reporting since May but were never formally confirmed. Per Data Center Dynamics, TechCrunch, and KraneShares:

• Anthropic pays xAI $1.25 billion per month for compute capacity.
• The capacity is ~300 megawatts at the Colossus 1 data center near Memphis, Tennessee — roughly 220,000 GPUs.
• The contract runs through May 2029 and could exceed $40 billion in total value.
• Either party can terminate with 90 days’ notice.
• Separately, the filing shows xAI lost $2.4 billion in Q1 2026, up from $936 million a year earlier, driven by data-center capex.

Because this is an SEC filing, it’s a materially more reliable source than the press reports that preceded it — the numbers are now legally attested, not leaked. That’s the “sources or skip” gold standard.

Why this matters

1. It shows that compute, not rivalry, is the organizing principle of AI right now. The safety-first lab that refused Pentagon work over autonomous-weapons concerns is paying Elon Musk’s company $15 billion a year to rent data-center space. That’s not a contradiction — it’s the clearest possible statement of priorities. Access to power and GPUs is so scarce, and so slow to build, that even bitter competitors trade it. You cannot conjure 300 megawatts and 220,000 GPUs on demand; the power contracts and construction take years. Anthropic buying ready capacity from a rival was simply faster than building its own — and in 2026, speed of capacity is the whole game.

2. The $2.4B quarterly loss reframes every “who’s winning” narrative. xAI’s loss nearly tripling year-over-year — to $2.4 billion in a single quarter — is a reminder that the AI race is being run at a staggering cash burn. The SpaceX/xAI merger, the $60B Cursor acquisition, the Colossus buildout — all of it runs on losses that would sink most companies. Renting capacity to Anthropic ($1.25B/month) is partly how xAI defrays that spend. It’s a revealing loop: one lab’s compute bill funds a competitor’s data-center losses, and both depend on the arrangement continuing.

3. The 90-day termination clause is the detail traders and buyers should notice. A ~$40B contract that either side can exit on 90 days’ notice is not the ironclad commitment the headline number implies. It gives Anthropic flexibility (if it builds or buys cheaper capacity, it can walk) and gives xAI exposure (a huge revenue line that could evaporate in a quarter). For anyone reading SPCX’s IPO on the strength of “guaranteed Anthropic revenue,” that clause is the asterisk — the revenue is real but not locked.

4. It’s the balance-sheet reality behind the prices you actually pay. Every trend we’ve covered on the product side — GPT-5.6’s cheap Terra/Luna tiers, Claude Opus 5 at half of Fable 5’s price, the custom-silicon push — is a response to these numbers. When capacity costs $1.25B/month and labs lose billions a quarter, relentless pressure to cut cost-per-token isn’t optional; it’s survival. The consumer benefit (cheaper, more efficient models) is downstream of an infrastructure economy operating at eye-watering scale and loss.

5. It exposes the concentration underneath a “competitive” market. On the surface, Anthropic and xAI compete for users. Underneath, Anthropic’s frontier models partly run on xAI’s hardware, xAI is inside SpaceX, and SpaceX is a launch-and-satellite monopoly going public at a ~$1.75 trillion valuation. The AI “market” is more entangled than the model-vs-model framing suggests — a handful of infrastructure owners underpin nominally-rival products. For buyers thinking about resilience and vendor risk, that hidden concentration is worth understanding.

The bigger pattern: everyone rents from everyone

This deal looks strange in isolation — Anthropic paying its rival — but it’s one thread in a web of cross-dependencies that defines the 2026 compute economy. Anthropic doesn’t run only on xAI hardware: it also has a $40B+ Google investment and TPU capacity, memory-chip partnerships with Samsung, SK Hynix, and Micron, its own custom-silicon work with Broadcom, and — announced days before this filing surfaced — a deal for up to 2 gigawatts of AMD Instinct MI450 GPUs, with AMD investing up to $5B in Anthropic, roughly 6–7× the xAI capacity described here. OpenAI has its Jalapeño chip with Broadcom and cloud deals across Azure and AWS. Google is building Frozen v2 to ease its own capacity shortage.

The through-line: no single lab controls enough compute, so they buy it from whoever has spare capacity — including competitors. The result is a market where the “rivals” are financially entangled at the infrastructure layer even as they compete fiercely at the product layer. For a buyer, the practical implication is that model availability and pricing depend on a fragile, shared, capacity-constrained substrate — one that a single big termination or power shortfall could disrupt across multiple “independent” providers at once.

What this means for you

• You don’t buy compute, but you pay for it. These economics set your model prices, rate limits, and availability. The cheap tiers and efficiency models arriving all month are the visible edge of this invisible cost pressure.
• Read “guaranteed revenue” IPO framing skeptically. If you’re evaluating SPCX or the AI-infrastructure trade, the 90-day termination clause means the Anthropic revenue is contingent, not fixed. Weight it accordingly (and note we’re not offering investment advice — this is about reading the disclosure).
• Expect more such deals, and more disclosure. As labs approach public markets (Anthropic and OpenAI both filed confidentially), more of these compute arrangements will surface in filings. The infrastructure layer is where the real money — and the real risk — sits.
• Judge labs on products, not purity. Anthropic renting from Musk doesn’t make its safety positioning fake; it makes it pragmatic. Buy tools on what they do, not on the tidy narratives around who’s aligned with whom.

The honest caveats

• The deal isn’t brand new; the formal disclosure is. The $1.25B/month arrangement was reported in May. What the S-1 adds is legal confirmation and the xAI loss figure — meaningful, but don’t read it as a fresh deal.
• “Could exceed $40B” is a run-rate projection, not a guaranteed total. At $1.25B/month through May 2029 the arithmetic reaches ~$40B+, but the 90-day termination clause means the realized total could be far less.
• Colossus specifics blur across sources. Reporting variously references Colossus 1 (300 MW, ~220,000 GPUs) and, in some accounts, Colossus I & II; treat the exact facility split as approximate pending the filing’s precise language.
• xAI’s loss is one quarter, mid-buildout. A $2.4B quarterly loss during heavy data-center construction is not the same as a $2.4B/quarter steady state; capex-driven losses can compress once facilities are operational and monetized.
• This is disclosure, not endorsement. SEC filings are reliable for what’s attested, but the strategic interpretation here is ours; reasonable analysts will read the same numbers differently.

The grounded summary: SpaceX’s IPO filing quietly confirmed that the AI industry’s real substrate is a compute economy so scarce and expensive that safety-first Anthropic pays Elon Musk’s xAI $15 billion a year to rent it — while xAI loses billions a quarter building it. The model launches get the headlines; these balance sheets set the prices. Read the cheap tiers and efficiency models of the last month as exactly what they are: the visible output of an invisible, staggering cost.

Questions or corrections? Email Pick Right. Want the full list? See all news.