Quick Summary
A SpaceX securities filing revealed on May 20, 2026 that Anthropic has agreed to pay SpaceX nearly $45 billion over three years for computing power to run its Claude AI models. The payments work out to roughly $1.25 billion a month through May 2029, and the compute comes from SpaceX’s Colossus data centers in Memphis, which were originally built to train Grok. The deal nearly doubles SpaceX’s annual revenue and lands as Anthropic’s own run-rate revenue has jumped past $30 billion.
Key Takeaways
- Nearly $45 billion over three years: Roughly $1.25 billion a month through May 2029, with reduced fees during a ramp-up in May and June 2026.
- The number came from SpaceX’s IPO paperwork: The deal was announced earlier in May without terms; the dollar figure only surfaced in SpaceX’s S-1 filing.
- It is mostly for inference, not training: The capacity primarily runs Claude for paying subscribers, not training new models.
- The compute lives at Colossus 1 and Colossus 2 in Memphis: Colossus 1 houses more than 220,000 Nvidia GPUs; Anthropic is scaling up GB200 capacity at Colossus 2 through June.
- SpaceX now sells AI compute: After merging with xAI earlier this year, SpaceX is leasing out the same hardware originally built to train Grok, and says it expects to sign more deals like this.
- Why it matters here: The bottleneck for AI right now is compute, not ideas, and the labs that secure capacity are the ones that can keep their products fast and available.
On May 20, 2026, a securities filing pulled back the curtain on one of the largest computing deals the AI industry has ever seen. Anthropic has agreed to pay Elon Musk’s SpaceX nearly $45 billion over the next three years for the raw computing power it needs to run Claude. The companies announced the partnership earlier in May but kept the terms quiet. The dollar figure only became public because SpaceX disclosed it in the paperwork for its initial public offering.
~$45 billion
Total Anthropic will pay SpaceX over three years.
$1.25 billion/month
Monthly payment running through May 2029.
220,000+ GPUs
Nvidia chips inside Colossus 1 in Memphis.
What the Deal Actually Says
In one line: Anthropic pays a fixed monthly fee for three years in exchange for guaranteed access to one of the largest GPU clusters on earth.
The structure is straightforward for a deal this size. Anthropic pays SpaceX about $1.25 billion every month through May 2029, with a reduced fee during the ramp-up period in May and June 2026 while the capacity comes online. Either side can walk away with 90 days’ notice, which is a notable escape hatch on a contract worth tens of billions of dollars. SpaceX’s filing frames the arrangement as a way to monetize unused compute capacity in its infrastructure, and the company says it expects to enter into more agreements like it with other firms.
- Roughly $1.25 billion per month, paid through May 2029.
- Lower fees in May and June 2026 while capacity ramps up.
- Either party can terminate with 90 days’ written notice.
Where the Compute Comes From
In one line: Anthropic is renting capacity from the Colossus supercomputers in Memphis, the same machines built to train Grok.
Earlier in May, Anthropic secured access to more than 300 megawatts of computing capacity at Colossus 1, a data center in Memphis, Tennessee that houses over 220,000 Nvidia GPUs spanning H100, H200, and the newer GB200 accelerators. It has since expanded the partnership to a second facility, Colossus 2. Co-founder and chief compute officer Tom Brown confirmed on X that Anthropic will scale up GB200 capacity at Colossus 2 throughout June, and noted the compute will go primarily toward inference workloads for Claude subscribers. There is a real irony here worth sitting with: Colossus was originally built to train Grok, the model meant to compete directly with Claude.
FAQ: Is this for training Claude or running it?
Mostly for running it. Tom Brown confirmed the Colossus compute goes primarily toward inference, which is the work of answering user prompts in real time, rather than training brand-new models. That detail matters because it tells you the bottleneck Anthropic is solving for: keeping Claude fast and available for paying users as demand climbs.
Why Anthropic Needs This Much Compute
In one line: Demand for Claude has outrun Anthropic’s capacity, and the company’s revenue has grown fast enough to afford the bill.
This deal is a direct response to a capacity problem. Anthropic’s run-rate revenue has climbed past $30 billion, up from roughly $9 billion at the end of 2025, driven largely by surging use of Claude for coding and other agentic work. That growth created a hard ceiling: the company simply did not have enough compute to serve everyone reliably. Securing 300-plus megawatts on a multi-year contract is the fast path to relief, faster than trying to build equivalent capacity from scratch. It also sits alongside Anthropic’s other recent compute deals with Amazon and Google, which together have let the company raise usage limits across Claude Code and the API. For brands working in answer engine optimization, a more available Claude simply means more chances for your content to be surfaced when people ask it questions.
- Run-rate revenue has passed $30 billion, up from about $9 billion at the end of 2025.
- Demand for Claude has consistently outpaced available compute.
- Renting capacity is faster than building new data centers from the ground up.
What SpaceX Gets Out of It
In one line: A deal that nearly doubles SpaceX’s annual revenue and turns its AI hardware into a rentable product.
For SpaceX, the math is striking. The company’s annual revenue currently sits around $18 billion, so an incoming $15 billion a year from Anthropic nearly doubles it. SpaceX merged with Musk’s xAI earlier this year, and since then it has been working to turn its computing infrastructure into a revenue stream by selling access to it. The Anthropic contract is the proof point. SpaceX stated in its filing that it has enough capacity to serve its own AI models and meet these obligations at the same time, and that it expects to sign similar deals with other companies going forward.
What this compute is for
Inference (running Claude)
Serving paid subscribers
Keeping Claude fast and available
What it is not for
Training new models
Replacing Amazon or Google deals
Those run in parallel
| The Deal at a Glance | Detail |
|---|---|
| Total value | Nearly $45 billion over three years |
| Payment schedule | About $1.25 billion per month through May 2029 |
| What it buys | 300-plus MW of compute, primarily for running Claude (inference) |
| Where | Colossus 1 and Colossus 2 data centers in Memphis, Tennessee |
| Hardware | 220,000-plus Nvidia GPUs (H100, H200, GB200) |
| How it surfaced | SpaceX’s S-1 IPO filing, disclosed May 20, 2026 |
| Exit clause | Either party can terminate with 90 days’ notice |
Scope Note
The reported figure comes from SpaceX’s IPO filing and Anthropic’s own confirmation, not from leaked or unverified sources. The roughly $45 billion total reflects the full three-year term if the contract runs to completion. Because either party can exit with 90 days’ notice, the final amount paid could be lower if the deal ends early.
The Bigger Picture
In one line: Compute has become a strategic asset, and access to it is now as decisive as the models themselves.
The headline number is eye-catching, but the real story is what it signals. For the leading AI labs, the competition is no longer only about who builds the best model. It is increasingly about who can secure enough computing power to actually run those models for millions of users without slowing down or hitting limits. Two direct rivals in the model race, Anthropic and the SpaceX-xAI side, just struck a deal that puts a hard price tag on that capacity. When competitors are willing to rent compute from each other at $15 billion a year, it tells you that compute procurement has become a competitive weapon in its own right. The model still matters. But increasingly, so does the electricity bill behind it. If you want to understand which of these models is most likely to surface your brand, our breakdown of how ChatGPT, Claude, Gemini, and Perplexity each decide which brands to cite is a useful next read.
Frequently Asked Questions
How much is Anthropic paying SpaceX?
Anthropic has agreed to pay SpaceX nearly $45 billion over three years, which works out to about $1.25 billion per month through May 2029. The fee is reduced during a ramp-up period in May and June 2026 while the computing capacity comes online.
What is Anthropic getting for the money?
Anthropic is getting access to more than 300 megawatts of computing capacity from SpaceX’s Colossus 1 and Colossus 2 data centers in Memphis, which together house over 220,000 Nvidia GPUs. The compute goes primarily toward inference, meaning running Claude for paying subscribers rather than training new models.
How did the price of the deal become public?
The two companies announced the partnership earlier in May without disclosing financial terms. The dollar figure only became public when SpaceX filed paperwork for its initial public offering on May 20, 2026, which laid out the monthly payments and the contract length.
Why is SpaceX selling AI computing power?
SpaceX merged with Elon Musk’s xAI earlier this year and has been working to generate revenue from its AI infrastructure by leasing access to it. The Colossus data centers were originally built to train Grok, and SpaceX now describes the Anthropic deal as a way to monetize unused capacity. It says it expects to sign similar agreements with other companies.
Does this mean Claude will get faster or have higher limits?
That is the goal. Because the capacity is aimed at inference and Anthropic has cited demand outpacing its available compute, more capacity is what allows the company to serve more users reliably and raise usage limits. Anthropic has already pointed to its recent compute deals as the reason it has been able to lift limits across Claude Code and the API. If you track AEO performance metrics, a more available Claude is one more engine worth watching.
