220,000 GPUs and 300 megawatts of Colossus 1 capacity online inside a month. The agentic economy is being built on a power and silicon supply chain most operators have never had to think about.

Eighteen months ago Elon Musk was suing Anthropic. This week he became its landlord. Anthropic announced a deal to take all of SpaceX's Colossus 1 supercomputer in Memphis: 220,000 Nvidia GPUs, including dense deployments of H100, H200, and the newer GB200 accelerators, and more than 300 megawatts of additional compute capacity coming online within the month. Buried in the same announcement is a sentence about partnering with SpaceX to develop "multiple gigawatts of orbital AI compute capacity." Not a typo. Space-based compute, on paper, ahead of an IPO reportedly slated for June.

The Colossus 1 deal is the smallest of Anthropic's recent compute commitments. The full stack is now more than 11 gigawatts of contracted capacity across SpaceX, Amazon, Google, Broadcom, Microsoft, NVIDIA, and Fluidstack, plus a $50 billion commitment to American AI infrastructure with Fluidstack and a $30 billion Azure agreement.

Colossus 1 is the supercomputer xAI built in Memphis with SpaceX. Anthropic is taking the full capacity, not a slice. The 220,000 GPUs are a mix of H100 and H200 (Nvidia's previous generation) plus dense GB200 deployments (the current Blackwell-class accelerators). More than 300 megawatts come online inside a month. That is enough to roughly double the available capacity behind Claude in the immediate term.

The use case Anthropic names is consumer-facing: directly improving capacity for Claude Pro and Claude Max subscribers, who have been hitting rate limits as agentic coding tools like Claude Code, Cursor, and Codex have moved from developer experiment to production infrastructure. Anthropic is also doubling rate limits for Claude Code and weekly Claude usage caps in the same release.

That is the headline. The deeper story is the orbital line. Anthropic's announcement says it has expressed interest in partnering with SpaceX to develop "multiple gigawatts of orbital AI compute capacity." SpaceX has been quiet about specifics, but Starlink-class infrastructure paired with low-earth-orbit data centers has been a thread in space-industry discussion for two years. Whether or not that materializes in any near term, the fact that it sits in an Anthropic announcement, alongside the most concrete commercial deal of the company's year, is a signal about where the compute frontier is heading.

Colossus 1 is one node in a much larger stack. Anthropic's own announcement inventories the recent commitments side by side: a $50 billion infrastructure investment with Fluidstack, a $30 billion Azure agreement with Microsoft and NVIDIA, an up-to-5-gigawatt Amazon agreement with nearly 1 gigawatt of new capacity arriving by the end of 2026, and a 5-gigawatt Google partnership with Broadcom beginning to come online in 2027. SpaceX's Colossus 1 sits in front of all of these on the timeline because its capacity is already built. The others are commitments and build-outs.

We have written about Anthropic's $40 billion Project Deal with Google and the broader AI compute shortage as the constraint underneath the agentic commerce build-out. Pay.sh, the agent stablecoin rail Solana and Google Cloud opened this week, runs on Google Cloud APIs that themselves run on capacity Google has been racing to add. The whole agent layer rests on the same constrained substrate.

What Anthropic is doing with these stacked deals is locking in priority access. Compute is being sold under multi-year commitments because the supply curve is short and the capex is heavy. Once the contracts are signed and the watts are reserved, the next tier of operators have to bid for what is left.

In January 2024, Elon Musk filed suit against OpenAI. Anthropic was not a defendant, but Musk's public commentary on the closed-source AI labs has been hostile and consistent for two years. xAI was founded as a competitive answer to both OpenAI and Anthropic. In the past month Musk has accused Anthropic of refusing to support a Pentagon vendor request and called for the company to be "broken up."

None of that survived contact with the compute crunch. SpaceX needed a customer for Colossus 1 capacity that xAI was not yet absorbing at the rate the build justified. Anthropic needed gigawatts on a timeline its existing partners could not deliver. The deal is transactional. It is also the second time in three months a frontier AI lab has taken capacity from infrastructure controlled by a competitor. The Pentagon's recent move to diversify frontier AI suppliers sits in the same context. Single-vendor dependency is no longer a posture any large buyer is willing to take.

For Mastercard, Visa, Stripe, and any other operator building agent infrastructure on top of these labs, the read is straightforward. Compute supply is volatile. Vendor neutrality at the model layer is becoming a structural concern, not a hedging gesture. The infrastructure underneath the agent is no longer a single-vendor question.

State of the Stack: Agentic Commerce 2026 tracked $690 billion in hyperscaler capex commitments over the next 24 months. Anthropic's stacked deals account for a meaningful chunk of that figure on the demand side. The contracts are between Anthropic and the suppliers, but the substrate is the same: data centers, GPUs, and the power to run them.

Two things follow.

First, the capex has to be paid for. The economic model for frontier AI labs is to recover that spend through subscription revenue (Claude Pro, Claude Max), enterprise API revenue (the developer platform), and per-token pricing on the inference side. Anthropic's price increases over the past 12 months are part of that recovery. Buyers should expect more, not less, especially as new commitments come online without immediate matching revenue.

Second, the watts are the constraint. GPUs can be manufactured. Power capacity has long lead times, planning approval bottlenecks, and grid-interconnection queues that run multiple years. Colossus 1 is online inside a month because the build is finished and the substation is energized. The Google-Broadcom 5-gigawatt build does not come online until 2027 because that capacity has to be built. State of the Stack flagged the watts question as the under-discussed constraint of the agentic economy. The Anthropic stack of deals is the clearest single illustration of why that matters.

For enterprises buying Claude or building on Anthropic's API, three near-term things change.

The rate limits ease. Claude Code and Claude weekly usage caps are doubling. Teams that were rationing access regain headroom, particularly for agentic workflows that require sustained inference rather than burst usage. That is a meaningful operational unlock for the next two quarters.

Pricing pressure remains. The $50 billion Fluidstack and $30 billion Azure commitments do not get repaid by giving capacity away. Subscription tiers have already moved up once this year. The next move is more likely to be tighter usage policies and per-task pricing options than across-the-board increases, but the direction is clear.

Vendor strategy changes. With the Pentagon moving on diversification and the major frontier labs cross-contracting capacity from each other's suppliers, model neutrality is becoming a default expectation in enterprise procurement. Buyers running agentic workloads on a single model provider should be planning for a multi-provider posture by the end of 2026.

The compute layer is the chokepoint. The networks built on top of it, including the agent rails Solana opened this week, depend on it staying ahead of demand. Anthropic just contracted to keep one frontier lab on the right side of that line for two more years.

At what point does compute stop being the silent partner of the agentic economy and start being the part of it that everyone is forced to talk about?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.