When SpaceX filed its S-1 registration document with the SEC in mid-May, the document revealed something unexpected buried in the business disclosures: Anthropic, one of OpenAI’s fiercest rivals, has agreed to pay xAI — Elon Musk’s AI company — $1.25 billion every month through May 2029 for access to the Colossus 1 data center in Memphis, Tennessee.

That is $15 billion a year. The deal could ultimately generate over $40 billion in revenue for xAI across the contract period.

The numbers are striking, but the story behind them is about more than one large transaction. It is a window into what frontier AI infrastructure actually costs, and why those costs matter to every business running AI workflows today.

How the Deal Came Together

Anthropic has been capacity-constrained for months. Earlier this year, the company secured access to all of Colossus 1’s compute — more than 220,000 Nvidia processors drawing 300 megawatts of power. The announcement in May focused on the access itself and what it meant for Claude users: rate limits lifted for paid subscribers, more headroom for Claude Code and API workloads.

What was not initially clear was the financial structure. SpaceX’s S-1 changed that.

The agreement includes a discounted rate for the first two months while xAI completes its infrastructure ramp-up. Either side can exit the contract with 90 days’ notice, which means neither party is locked in without an escape valve. Anthropic gets predictable compute capacity. xAI monetizes infrastructure it had sitting underused.

That underuse is telling. xAI built Colossus to run Grok, its flagship AI assistant. Grok’s adoption has not kept pace with xAI’s infrastructure investment, leaving significant spare capacity at the Memphis facility. Selling that capacity to a competitor is a pragmatic solution that turns a cost center into a revenue stream.

The Neocloud Model

The Anthropic-xAI arrangement is an example of what analysts are starting to call the “neocloud” model: AI companies with excess compute capacity acting as infrastructure providers to other AI companies, even direct competitors.

This is not how the hyperscalers envisioned the compute market working. Google, Microsoft, Amazon, and Oracle have spent billions building out data center capacity specifically to host AI workloads as a cloud service. But AI labs are now building their own infrastructure at a scale that sometimes overshoots their immediate consumption, and they are finding it more economical to lease that excess capacity to peers than to leave it idle.

The pricing signals something else: compute at this tier is genuinely expensive. Anthropic paying $1.25 billion a month means Colossus is not a rounding error on anyone’s budget. For reference, Anthropic reported approximately $10.9 billion in revenue in its most recent quarter — and is paying the equivalent of nearly half of that revenue in a single recurring compute contract.

Why This Matters If You Run AI Workloads

The headline number is dramatic, but the downstream implications for enterprise AI buyers are more practical.

Compute costs shape what AI products can offer. When the foundational infrastructure costs this much, every AI vendor — whether that’s Anthropic, OpenAI, Google, or a smaller developer — is managing an enormous cost structure. That pressure flows through to API pricing, rate limits, and tier structures. Understanding it helps business buyers negotiate smarter contracts and choose vendors whose cost models are sustainable.

Availability is now a competitive dimension. One of the main reasons Anthropic pursued the Colossus deal was to lift capacity constraints on Claude. Rate limits and slow response times are not just inconveniences — for businesses running agent workflows that depend on high-throughput AI, they break operations. More compute means more reliable throughput, and that is a genuine product differentiator.

The infrastructure race is consolidating. Three years ago, the assumption was that AWS, Azure, and Google Cloud would own enterprise AI compute. What we are seeing instead is a more complex supply chain: AI labs building their own infrastructure, leasing to each other, and in some cases building vertical relationships that bypass the traditional cloud layer entirely. Business leaders buying AI services should understand which layer of that stack their vendor sits on.

Competition between AI labs benefits buyers over time. Anthropic securing 220,000 Nvidia processors — whether from AWS, Google, or xAI — increases its ability to match OpenAI on availability and throughput. More capacity at Anthropic means better service levels for Claude Enterprise and Claude API customers. The same dynamic applies across the industry. Infrastructure competition creates availability competition, which ultimately creates pricing pressure.

The Musk Angle

There is an obvious irony here. Elon Musk has been publicly hostile toward Anthropic for most of 2025 and early 2026. The deal happening at all required a shift in that dynamic, which apparently came after Musk spent time with Anthropic’s team in late April and revised his view of the company.

For enterprise AI buyers, the politics matter less than the outcome. Anthropic gets the compute it needs to scale. xAI gets revenue from capacity it was not using. Both sides have a 90-day exit clause that keeps the arrangement honest.

The deal is a reminder that AI industry rivalries rarely map cleanly onto business interests. Companies that compete at the product layer can still find commercially rational reasons to partner at the infrastructure layer. That complexity is worth tracking, because it shapes the supply chain underpinning every AI product businesses deploy.

What This Means for Business

If your team is running AI workflows on Claude — code generation, agentic automation, document processing, customer interaction — the Colossus deal is net positive for service reliability. More compute supply typically means better uptime, higher rate limits, and more capacity for concurrent workloads.

More broadly, the $1.25 billion monthly figure is a useful calibration point for understanding the AI infrastructure economy. Compute is not cheap. The companies building on top of it are managing serious cost structures. When AI vendors price their enterprise tiers, they are working backward from numbers like this. Knowing that makes it easier to evaluate whether a given vendor’s pricing model is sustainable and whether their infrastructure commitments are real.

For any business evaluating AI vendors or building internal AI strategy, getting clarity on the infrastructure layer is as important as evaluating model capability. If you need help mapping the AI vendor landscape and making sense of which compute arrangements actually affect your operational choices, the Omni Advisory service is built for exactly that.