- Anthropic is paying xAI $1.25 billion per month for access to its Colossus data centers, according to a SpaceX SEC filing, a contract worth more than $40 billion if it runs to its May 2029 end date.
- xAI open-sourced its Grok Build coding assistant on July 16, 2026, under an Apache 2.0 licence, following a privacy scandal, not as a planned transparency initiative.
- xAI now plays simultaneous roles as model developer, infrastructure provider and revenue-generating supplier to a direct competitor, illustrating how GPU scarcity is redrawing competitive lines at the frontier.
A SpaceX SEC filing has revealed that Anthropic is paying xAI $1.25 billion a month for access to its Colossus data centers, a contract worth more than $40 billion if it runs to its May 2029 end date. Even a well-capitalised AI lab, it turns out, had to go to a direct competitor to secure enough GPUs. That same competitor, in the same week, released its coding tool as open source after a privacy scandal.
Anthropic Secures Massive Compute from xAI
The contract gives Anthropic access to Colossus 1 and Colossus 2, xAI’s Memphis data centers, with the filing citing roughly 330 megawatts of combined compute capacity across both sites; a separate, planned expansion phase at Colossus 2 is set to add at least 220,000 more GPUs and 400 additional megawatts, though the filing doesn’t specify how much of the current or expanded capacity Anthropic’s contract actually covers.
For xAI, the arrangement is straightforwardly good business. Colossus was built to train Grok, but the cluster carries more capacity than xAI currently consumes internally. Leasing that headroom to Anthropic at $1.25 billion a month turns sunk infrastructure cost into a significant revenue stream. For Anthropic, it solves a compute bottleneck that was limiting how fast it could scale Claude. The dependency is real: Anthropic is paying a competitor for the physical layer its own models run on.
The scale of the payment also reflects the actual cost of frontier AI development. Guaranteed access to tens of thousands of high-end GPUs, at this density and power draw, commands pricing that generic cloud providers cannot match, which is exactly why a dedicated AI infrastructure model is emerging alongside the hyperscalers. For more on the power demands this kind of infrastructure creates, see our coverage of AI workload power projections through 2030.
Grok Build Goes Open Source Following Privacy Concerns
xAI open-sourced Grok Build, its command-line AI coding assistant, after a serious privacy problem came to light.
Elon Musk responded by pledging to delete all previously uploaded user data and switching data retention off by default for all users, not just paid enterprise plans. The codebase was released under an Apache 2.0 licence. Previously, developers had to reverse-engineer the tool’s behaviour using external software; the release makes the full terminal renderer and coding agent infrastructure available for direct inspection. Developers can analyse the code, modify it, or run Grok Build locally against their own inference backend.
The open-sourcing was a direct response to a privacy failure, not a planned transparency initiative. That distinction matters when weighing xAI’s framing of the move as a trust-building measure.
xAI’s Broader Open-Source Philosophy
Grok Build’s release follows xAI’s earlier open-sourcing of Grok-1 on March 17, 2024. Grok-1 is a 314-billion-parameter Mixture-of-Experts model; xAI released its weights and architecture under Apache 2.0, giving researchers direct access to inspect and build on the model without proprietary restrictions. Musk framed that release as a commitment to open AI development, contrasting it with the closed approach taken by OpenAI and Anthropic.
Grok-1’s release was proactive. Grok Build’s was reactive. A company that releases models as a philosophical statement and releases tools after a data scandal is running two different programmes under the same banner. The Apache 2.0 licence is identical in both cases; the circumstances that produced each release are not.
Navigating the AI Compute Paradox
The two developments produce a sharp contradiction. xAI presents itself as a champion of open AI, releasing models and tools that proprietary labs keep closed. It is also now the critical infrastructure provider for Anthropic at $1.25 billion a month. Ideological positioning and commercial reality are pulling in opposite directions, and the commercial reality is winning.
That is not necessarily a criticism. Compute scarcity creates dependencies that cut across competitive lines. Anthropic needed GPUs; xAI had them. The business logic is clear, but it complicates the “open versus closed” framing Musk has used to differentiate xAI, because the company is now financially bound to the closed-model ecosystem it claims to oppose.
Implications for the AI Infrastructure Market
The Anthropic-xAI deal validates what is sometimes called the “Neocloud” model: purpose-built AI infrastructure providers that sit between traditional hyperscalers and the AI labs that need guaranteed GPU access at scale. The $1.25 billion monthly figure is the clearest public data point yet on what that access costs at the frontier. Generic cloud pricing does not get you there.
The contract also illustrates how AI competition now works. xAI is simultaneously a model developer, an infrastructure provider and a revenue-generating supplier to a direct competitor. Those roles used to belong to separate companies. The fact that one organisation now plays all three, and that a leading AI lab has no better option than to pay it more than $40 billion over four years, says more about the current GPU supply situation than about any particular company’s strategy. The companies that built large clusters first are now in a position to charge accordingly. For more coverage of AI chips and infrastructure, visit our AI Hardware section.
