Anthropic is buying computing capacity from its competitor
Anthropic and SpaceX have already been working together for several months. This convergence began to become visible when Anthropic signed enormous infrastructure contracts as Bitcoin miners turned to AI.
In May, Anthropic officially announced that it had secured access to the full capacity of Colossus 1, with more than 220,000 Nvidia GPUs and approximately 300 MW of power. The agreement was subsequently expanded to Colossus II. Anthropic said at the time that the additional capacity would allow it to increase the limits of Claude Code and its API.
Documents filed by SpaceX with the SEC offer a different picture: approximately 325,000 Nvidia GPUs are covered by its agreements with Anthropic. The customer must pay $1.25 billion per month through May 2029, with reduced payments during the initial ramp-up. Over three full years, the theoretical amount approaches $45 billion. However, either party may terminate the agreement with 90 days’ notice after the initial period.
So where does the $84.5 billion figure come from?
Reuters, which reviewed Anthropic’s confidential prospectus, describes a separate agreement that could reach $84.5 billion with xAI for Nvidia capacity. Since February 2, that distinction has become more subtle: SpaceX acquired xAI, which is now its wholly owned subsidiary.
Presenting the $84.5 billion as a simple SpaceX contract is therefore not entirely unreasonable at the group level, but it conflates two different perspectives. Anthropic’s document refers to xAI. SpaceX’s documents, meanwhile, detail the monthly contract linked to Colossus.
$518 billion to power Claude
The truly staggering figure is even higher: $518 billion.
This is the minimum amount of infrastructure commitments Anthropic has listed for the coming years. Reuters estimates that around 80% are tied to largely non-cancellable contracts, regardless of actual usage levels.
Google accounts for at least $111.1 billion, Amazon for approximately $110 billion, and Microsoft for $31.4 billion. Broadcom is associated with around $161.2 billion in leasing obligations. xAI, AMD, Akamai and other providers add to the total.
Akamai, for example, recently officially disclosed a new $11.6 billion contract with Anthropic to the SEC. The agreement covers dedicated cloud capacity and related support services.
This enormous need for computing capacity also explains why artificial intelligence is beginning to take Bitcoin miners’ best power connections. GPUs and models are not enough. Hundreds of megawatts, data centers, cooling systems and power connections are required—and these can sometimes take several years to build, a challenge already visible as AI data centers run into Africa’s electricity wall.
Anthropic spent $7.33 billion on computing capacity and infrastructure in 2025, nearly three times as much as in 2024. That was already more than its $4.59 billion in annual revenue. Its business has since grown to a different scale, but so has the bill.
The company had approximately $20.28 billion in cash and short-term investments, according to its 2025 figures. The $518 billion is obviously not due immediately. It represents commitments spread over several years and will have to be financed through future revenue, new funding rounds and potentially debt.
Musk is now selling computing capacity to his rivals
The agreement also reveals a profound transformation at SpaceX.
With xAI integrated into the group, it is no longer merely a rocket, satellite and Starlink company. It is gradually becoming a major provider of computing power.
Colossus was initially built to train Grok. The group can now lease excess capacity to other companies, including direct competitors of xAI.
Anthropic is probably the most striking example.
Claude and Grok are competing for developers, businesses and AI users. Yet when Anthropic runs certain workloads, some of the money can now end up with Grok’s parent company.
This model increasingly resembles that of the hyperscalers. Amazon has its own AI products but leases AWS to others. Google is developing Gemini while selling Google Cloud to competitors. Microsoft owns Copilot and provides Azure to numerous laboratories.
SpaceX/xAI is beginning to follow the same logic.
The group also states in its documents that it wants to sign additional computing contracts with third parties. It says it has sufficient capacity to power its own models while serving external customers.
For Anthropic, the calculation is different. The company simply cannot wait several years for its own data centers to be completed. Demand for Claude is growing much faster than the physical construction of the infrastructure. Buying capacity from Amazon, Google, Microsoft, Akamai or even a competitor makes it possible to save time.
It is also a risk.
Anthropic acknowledges in its prospectus that it is highly dependent on several groups that are simultaneously suppliers, investors or competitors. Hundreds of billions of dollars now rest on this architecture.
The AI rush is therefore entering a new phase. After the race for the best models, the real battle is also taking place behind data-center walls: GPUs, megawatts, networking and cooling.
And the amounts are moving to a different scale.
$84.5 billion for a single potential relationship with xAI. $518 billion in total infrastructure commitments. Anthropic is no longer betting solely on Claude’s success. It is building its financial future on the assumption that demand for artificial intelligence will continue growing fast enough to absorb an almost industrial-scale computing bill.
