As a tenant, I know the quiet leverage in how you pay rent. My landlord once offered me a choice: a full year upfront or month-to-month.
The yearly deal protects the landlord’s peace of mind, while paying monthly protects my freedom. If the building goes downhill or a better deal pops up, I can pack my bags without losing a fortune.
Anthropic made the same choice with Elon Musk’s SpaceX. In February, Musk replied on X (the former Twitter) by calling the Claude maker “evil.” Seven months later, Anthropic’s potential compute bill with his company has nearly doubled.
Anthropic could pay SpaceX up to $84.5 billion for Nvidia-based computing power through 2029, according to a confidential IPO prospectus reviewed by Reuters. SpaceX’s own IPO filing in May pointed to a deal worth roughly $45 billion.
The size grabbed headlines, but the exit door matters more. Most of the SpaceX capacity can be canceled on 90 days’ notice, Reuters reported.
About 80% of Anthropic’s $518 billion in infrastructure commitments, by contrast, must be paid whether it uses the capacity or not.
Also read: Anthropic’s IPO results in an unusual contradiction for investors
Anthropic gave Musk the month-to-month lease
Look at who got the yearly deals. Anthropic has committed more than $250 billion in locked-in spending to Google, Amazon, and Microsoft over seven to 10 years, according to Reuters figures.
That split looks deliberate to me. All three cloud giants are Anthropic investors and established suppliers, so a decade-long commitment to them is easier to justify. SpaceX is a newer landlord whose AI unit runs a rival chatbot, Grok, and whose founder has mocked Anthropic in public.
So Anthropic acted like a careful tenant. It took the space, paid the rent, and kept a key to the back door.
Anthropic can cancel most of its SpaceX compute deal on 90 days’ notice, but 80% of its $518 billion infrastructure bill is locked in.
SpaceX shareholders carry the risk of that exit clause
The landlord in this story is a public company. SpaceX trades on Nasdaq as SPCX after raising roughly $86 billion in June, the largest IPO ever, according to CNN. Its growth story now leans heavily on renting AI computing power to outside labs, with Anthropic as the marquee customer.
For investors, it makes Anthropic’s spending decisions a direct input into SpaceX’s revenue line.
That business is expensive to build. SpaceX poured $15.8 billion into AI capital spending in the second quarter alone, according to S&P Global Market Intelligence. S&P found that AI also drove most analysts’ estimate upgrades after the quarter.
Here is what I think the market undervalues. SpaceX is building data centres designed to last for years, funded by revenue from a customer who can leave in roughly three months. Reuters noted that SpaceX’s May filing contained a similar 90-day clause, yet most coverage at the time focused on the monthly payment.
To be fair, the clause cuts both ways. While Anthropic’s demand outruns supply, walking away would hurt Anthropic more than SpaceX. The real danger arrives only if cheaper capacity appears elsewhere or demand cools.
Anthropic’s own investors face the opposite problem
SpaceX’s risk is a customer who can leave. For anyone weighing Anthropic’s planned IPO, the problem is that Anthropic mostly can’t.
Anthropic told investors that future demand for advanced AI will be “limited principally by the availability of compute,” according to Reuters. It is betting hundreds of billions of dollars that this holds true for a decade.
The early numbers tell a mixed story. Anthropic’s revenue grew about twelvefold in 2025, but its operating loss still widened to $8.06 billion, Reuters reported.
The risk sits in the contracts. If demand slows, Anthropic can hand back SpaceX capacity but not most of what it owes the cloud giants, which are also its investors.
PitchBook analyst Harrison Rolfes flagged that loop in May. Cloud giants that invest in AI labs are also “the vendors collecting the compute bills,” he said, according to Morningstar. That leaves Anthropic little room to renegotiate if the market turns.
More Anthropic:
- Microsoft just sent a strong message to Anthropic
- Anthropic’s IPO results in an unusual contradiction for investors
- Anthropic’s $2 trillion IPO comes with a $518 billion bill
AI compute is splitting into anchors and swing suppliers
Step back, and Anthropic’s filing reveals a two-tier compute market taking shape. Anchor suppliers get decade-long commitments that can’t be canceled. Swing suppliers like SpaceX win big contracts with far weaker guarantees.
That distinction should change how investors value companies selling compute. A dollar of locked-in revenue and a dollar that can vanish on 90 days’ notice are not the same dollar. I’m not convinced the market has fully priced that difference yet.
Telecom investors learned a version of this lesson in the early 2000s. Fiber networks built for booming internet traffic outlasted many of the customers expected to fill them. AI compute is not fiber, but the risk of building ahead of demand looks familiar.
The next test arrives when Anthropic’s prospectus goes public and investors can read the termination terms themselves. Musk’s insult made this story memorable. The 90-day clause will decide who holds the leverage if the AI boom cools.
Related: Anthropic’s $2 trillion IPO comes with a $518 billion bill
