Anthropic Discloses $84.5 Billion Compute Agreement with SpaceX
nashnova research
Anthropic's IPO prospectus reveals a compute agreement with SpaceX worth up to $84.5 billion through 2029 — nearly double the $45 billion previously disclosed — but most terms allow cancellation with 90 days' notice, making this a demand signal, not locked-in revenue.
How did $45 billion become $84.5 billion?
In May, SpaceX's IPO filing first disclosed that Anthropic agreed to pay $1.25 billion per month for up to three years — roughly $45 billion if fulfilled in full.
The latest prospectus raises the ceiling to $84.5 billion, extending the horizon to 2029 — an increase of about 87.8%.
This means → Anthropic has widened both the spending window and the dollar cap, signaling that its projected demand for model training and inference compute is scaling faster than it anticipated just months ago.
But most of the agreement can be canceled with 90 days' notice. In plain terms = this is a "could spend up to" framework, not a "will spend exactly" commitment.
Can Anthropic afford to burn this much?
In 2025, Anthropic's revenue grew to roughly $4.6 billion — about 12× the prior year — while operating losses exceeded $8 billion.
Compute and infrastructure spending hit $7.33 billion, roughly triple the 2024 figure and about 58% of total operating expenses of $12.65 billion.
This means → for every dollar earned, Anthropic spent about $1.60 on compute alone. Revenue is surging, but cash burn is surging faster.
Including Google, Amazon, Microsoft, and other compute partners, Anthropic projects at least $518 billion in total compute spending over roughly the next decade, with about 80% non-cancelable or pay-regardless.
What does all this compute buying mean for users?
Anthropic says the SpaceX deal and other new compute arrangements let it raise usage limits on Claude Code and its API and improve capacity available to paying subscribers.
This reflects a transmission chain: more compute delivered → broader product capability → more paid usage.
In plain terms = buying more GPUs is not about stockpiling — it is about ensuring the product can absorb more users and sell more subscriptions. Compute procurement is essentially a bet on future commercial revenue.
Who else is buying compute from SpaceX?
Google has signed a deal for roughly 110,000 Nvidia GPUs and associated compute, paying $920 million per month from October 2026 through June 2029 — roughly $30.36 billion over the full 33 months.
Musk disclosed on September 25 that the Colossus 1 and Colossus 2 AI data centers house about 780,000 AI GPUs combined, with plans to add Nvidia GB300 clusters in phases — potentially reaching 1.44 million GPUs total, though the late-December batch still carries uncertain conditions.
This means → SpaceX is evolving from a rocket company into a major AI infrastructure provider, and the Anthropic and Google contracts are simultaneously validating the commercial appeal of that new role.
How much of this will actually materialize?
The $84.5 billion is a ceiling, not confirmed revenue. The cancellation clauses mean SpaceX's actual take depends on Anthropic's commercialization trajectory.
Google's contract likewise includes delivery requirements and early-termination provisions; SpaceX's planned late-December GPU expansion batch also carries uncertain conditions.
In plain terms = everyone is signing massive compute agreements, but these deals function more as "demand-intent ceilings." How much actually lands depends on whether AI companies can turn their models into sustainably profitable businesses.
市场有风险，内容仅供研究参考，不构成投资建议。