Anthropic Just Agreed to Pay SpaceX $45 Billion Through 2029. Every Claude API Call You Make Now Pays Elon Musk's Rocket Company.

On May 20, Bloomberg confirmed that Anthropic agreed to pay SpaceX nearly $45 billion for compute access through May 2029. The headline number is $1.25 billion per month, ramped down during a capacity build-up in May and June 2026 and at the contracted rate after that. The deal gives Anthropic access to roughly 220,000 NVIDIA GPUs across Colossus 1 and Colossus 2, totalling more than 300 megawatts of compute capacity. SpaceX's S-1 filing notes that the Anthropic deal accounted for 76% of its $10.1 billion in Q1 2026 capital spending.

If you ship anything on Claude (a SaaS, a coding agent, an internal tool you sell to mid-market companies), your unit economics now sit downstream of a neocloud controlled by Elon Musk. Most coverage is framing this as Musk-vs-OpenAI drama or as Anthropic-buys-its-way-out-of-the-compute-crunch. The solo operator angle is more concrete: the exogenous variables in your Claude pricing model just got weirder.

The actual structure, not the press-release version

The deal has four pieces worth understanding before you decide whether it changes anything about how you build.

First, the money. $1.25 billion per month is the contracted run-rate. The ramp during May–June 2026 means Anthropic pays less while the second cluster comes online. Across three years, the total lands close to $45 billion. That is, in round numbers, the same order of magnitude as Anthropic's projected 2026 revenue, and meaningfully larger than the cash Anthropic currently has on hand. Anthropic is funding this from its in-progress $30 billion round, which is partly why the round closed at the valuation it did.

Second, the hardware. 220,000 GPUs and 300 MW puts this in the same tier as the largest hyperscaler clusters, but it's a different relationship than Anthropic's existing AWS and Google Cloud contracts. Colossus 1 is the xAI cluster in Memphis. Colossus 2 adds next-generation GB200 hardware. The supply is SpaceX's, not AWS's, not Microsoft's, and the only other AI tenant of comparable size is xAI itself.

Third, the exit clause. Either party can terminate with 90 days' notice. This is unusual for a contract of this size, and it is the part of the deal I find most interesting. Anthropic is not locked in. Neither is SpaceX. Three years of revenue commitment with a 90-day exit on either side is structurally a long-dated month-to-month contract, not a take-or-pay infrastructure deal.

Fourth, the orbital footnote. The partnership announcement mentions both companies exploring multi-gigawatt orbital AI compute capacity: data centers in space, solar-powered, cooled by the void. This sounds like science fiction. It also sounded like science fiction when SpaceX said it would land orbital boosters. I would not bet my pricing model on space-based compute by 2029. I would also not bet against it by 2032.

Why this exists, in plain English

Anthropic's compute crunch has been building for two years. Demand for Claude grew faster than AWS and Google could provision dedicated capacity. The $4.8B Q1 revenue jumping to a projected $10.9B in Q2 only makes the math worse. You cannot grow inference 130% quarter-over-quarter if your cloud providers can only commit to 40% more dedicated GPU allocation.

SpaceX, meanwhile, has compute sitting around. Colossus 1 was built for xAI training runs that don't fill the cluster 24/7. Colossus 2 is being built for the same reason. Both clusters benefit economically if a second tenant buys the unused capacity. SpaceX's S-1 calls the Anthropic deal a way to "monetize unused compute capacity in its infrastructure."

So the deal is mostly rational on both sides. Anthropic gets capacity its other vendors couldn't deliver in time. SpaceX gets a $45B revenue stream that funds the rest of the cluster build-out. The structurally weird part is not the deal: it's that this market now has Musk-controlled infrastructure underneath one of his loudest critics' products.

What this means for your unit economics

If you charge customers based on a margin over your Claude API spend, three things change for you.

One: pricing pressure is no longer purely capacity-driven. Anthropic now has a fixed $1.25B monthly cost regardless of how many tokens it serves. The economic incentive is to drive more usage through the same compute, which means continued pressure to push prices down on volume customers. If you're a high-volume builder on Claude, you should expect to keep getting better pricing, not worse, at least through the ramp.

Two: the 90-day exit clause works both ways, and SpaceX has leverage. If a future political or commercial disagreement triggers SpaceX to use that 90-day clause, Anthropic does not have a replacement 220K-GPU cluster sitting idle anywhere. That tail risk does not affect daily inference: it affects what happens if you're building a business that materially depends on Claude being available at current prices in 2028.

Three: the contract concentration. SpaceX disclosed in its S-1 that the Anthropic deal was 76% of Q1 capex. That makes Anthropic functionally SpaceX's largest customer, by a wide margin, in a quarter where SpaceX is filing to go public. SpaceX has a financial incentive to keep this relationship working through its IPO window. After that, the incentives can drift.

What I'd actually do

I have a small Claude-based product running in production. Here's what I'm changing this week and what I'm leaving alone.

I'm not switching providers. Claude is still the best model for what I ship, and the structural risk of SpaceX cutting Anthropic off is real but not 12-month-real. The bet that you should make on a 3–4 year deal with a 90-day exit is that the first 18 months are stable. They will be.

I am building a model-agnostic abstraction at the API layer. Not for cost. For optionality. The week the Microsoft-OpenAI exclusivity blew up was the week I started taking model-portability seriously. If your code currently has anthropic.messages.create strewn across thirty files, that's the version of vendor lock-in that costs three weeks of refactor when you need to move in a hurry. Two days to abstract now buys you a 48-hour migration later.

I am also raising prices for any product priced on a Claude-cost-plus model. Not because I expect immediate token price hikes (actually the opposite, I think prices stay flat or fall through 2026) but because the volatility in the underlying supply chain is now measurably higher. The risk premium on a Claude-based product should be priced in, not absorbed.

And I am reading the SpaceX S-1 in full when the redacted sections clear. The deal terms that affect indie builders are not in the headlines. They're in the customer-concentration risk factors, the "material adverse change" language, and the disclosed renewal terms. That's where the surprise lives in 2027.

Where this argument could be wrong

The strong version of "SpaceX leverage now matters to Claude pricing" depends on assumptions that may not hold.

SpaceX may not actually be able to flex pricing or availability without breaching its own contract. The 90-day exit clause is mutual. If SpaceX exercises it without cause, Anthropic has a 90-day window to migrate workloads to AWS and Google, disruptive, but survivable. The clause is more useful as a negotiating lever than as a kill switch.

Anthropic also has not put all its compute eggs in this basket. The Amazon trainium expansion and the Google TPU contracts continue. SpaceX is the marginal capacity that lets Anthropic say yes to enterprise contracts it would otherwise have to defer. If SpaceX leaves, growth slows; the existing product does not turn off.

And the orbital data center stuff is a press release flourish, not a 2026 line item. I am budgeting nothing for it. Neither should you.

The honest take

If you ship on Claude, this deal does not change what you build this month. It changes what your contingency plan should look like by Q4 2026. Build the API abstraction now, raise prices on Claude-cost-plus products by 10–15%, and read the SpaceX S-1 when it clears redaction. The version of risk management that matters here is the one your accountant cannot model from the headlines.

The bigger question (the one mainstream coverage is dancing around) is whether having Musk-controlled infrastructure underneath the AI company most aligned against him is sustainable through 2029. I think it is, for one specific reason. Both sides need the deal more than they need each other to fail. That kind of mutual dependency is, historically, more stable than mutual affection. It usually breaks anyway, but it takes longer than the loudest commentators predict.

Author

Lukas

@lukcombinator

Sources

• Anthropic to Pay SpaceX Nearly $45 Billion for Computing Deal (Bloomberg)
• Anthropic secures $45bn SpaceX deal for Claude AI computing power (Verdict)
• SpaceX's $45 Billion Anthropic Deal Could Fund Its AI Buildout Three Times Over (Benzinga)
• SpaceX SPCX IPO S-1 Full Teardown (The VC Corner)
• Anthropic on Track for First Operating Profit as Revenue Surges (PYMNTS)