Anthropic is quietly turning one of the most hotly debated AI compute deals of 2026 into something much bigger: a full-on expansion onto NVIDIA’s next-generation GB200 systems inside SpaceX’s Colossus 2 supercomputer campus in Memphis. And it is doing it fast, with capacity expected to ramp through June, right as demand for Claude is pushing existing GPU clusters to their limits.

If you missed the first chapter of this story, here is the short version. Earlier in May, Anthropic struck a deal with Elon Musk’s SpaceX to gain access to the Colossus supercomputer complex, originally built by xAI to train Grok. Colossus 1, the first phase of that build-out in Memphis, already ranks among the largest AI training sites on the planet, with more than 220,000 NVIDIA GPUs and over 300 megawatts of compute capacity dedicated to large-scale model training and inference workloads. That alone would be newsworthy for any AI lab, but what Tom Brown, Anthropic cofounder and chief compute officer, just confirmed publicly is that the partnership is now extending into Colossus 2 and, crucially, onto NVIDIA’s cutting-edge GB200 hardware.

On X, Brown said Anthropic is “expanding our partnership with SpaceX” and will be “scaling up on GB200 capacity in Colossus 2 throughout June,” thanking Musk and his team for “helping us find good homes for the Claudes.” That one-line update packs a lot in: it tells you Anthropic is not just renting leftover H100s somewhere, it is moving into the latest Blackwell-based racks, and it is doing so inside an already enormous campus that xAI has been aggressively growing to around 2 gigawatts of total capacity and roughly 555,000 GPUs at full build-out. In other words, Anthropic is literally colocating its frontier AI stack alongside one of its fiercest rivals, on infrastructure originally built to power Grok.

To understand why the GB200 mention matters, you have to zoom in a bit on the hardware. NVIDIA’s GB200 NVL72 is not just “the next GPU.” It is a rack-scale system that ties together 72 Blackwell GPUs and Grace CPUs into a single, tightly coupled AI supercomputer, delivering on the order of an exaflop of FP4 compute per rack, with tens of terabytes of unified memory and a blistering high-bandwidth NVLink fabric. The design is optimized specifically for trillion-parameter-scale training and large mixture-of-experts models, shrinking communication overhead so that model parallelism becomes much simpler and real-time inference on giant models becomes actually practical, rather than a marketing slide. For Anthropic, whose Claude models are increasingly being used in code assistants, agents, and enterprise workflows that demand low-latency, high-throughput inference, this kind of rack-scale coherence is a direct lever for better user experience: higher rate limits, fewer timeouts, and the headroom to roll out larger, more capable versions of Claude.

There is also a timing angle here. Anthropic has been on a compute-signing spree across the big cloud providers and chip vendors, and the SpaceX deal slots into that broader strategy rather than replacing it. In late 2025 and early 2026, the company announced massive expansions with Google, Broadcom, and Amazon, securing multiple gigawatts of future TPU and Trainium capacity to arrive between 2026 and 2027 and beyond. Those commitments will eventually bring online up to a million TPUs with Google and as much as 5 gigawatts of Trainium-based capacity with Amazon, but much of that comes in over a multi-year timeframe. The Colossus deal, by contrast, gives Anthropic immediate access to a huge pool of NVIDIA GPUs, with operational capacity coming online within weeks rather than quarters, which is critical when current clusters are saturated and enterprise demand is compounding month over month.

From SpaceX and xAI’s side, Anthropic moving into Colossus 2 is a sign of how fast the economics and politics of AI infrastructure are shifting. Musk has spent the last two years loudly criticizing Anthropic and other labs while simultaneously pouring billions into building out xAI’s own compute estate for Grok, centered on the Colossus campus in Memphis. Public filings and industry reports suggest that xAI’s GPU orders, combined with partners, have turned Colossus into one of the largest single-site AI installations on Earth, with the January 2026 expansion alone representing roughly $18 billion worth of NVIDIA hardware. Yet in practice, even that scale has slack periods and spare capacity, especially as new racks come online and demand ramps. Renting out slices of that capacity to Anthropic effectively turns what used to be a pure competitive asset into a revenue-generating AI “power grid” that can monetize idle cycles while spreading out the enormous capital cost of building these gigawatt-scale data centers.

For Anthropic, the upside is obvious. Claude has gone from a niche research model to a mainstream tool embedded in developer IDEs, enterprise copilots, and consumer-facing chat interfaces. Each new release pushes token throughput, context length, and model complexity higher, and that in turn pushes hard against whatever GPU capacity happens to be online that quarter. Users have been increasingly vocal about rate limits and occasional reliability issues, and some of the early commentary around the Colossus deal framed it explicitly as a way to raise caps for paid tiers and API customers once the new capacity stabilizes. With GB200 racks coming online in Colossus 2, Anthropic gets exactly the kind of high-performance, scale-out environment it needs to offer bigger quotas, more concurrent requests, and potentially “heavier” Claude variants that would have been prohibitively expensive to serve on earlier-generation hardware.

Behind the scenes, this also locks Anthropic even more tightly into the global race for AI compute, where the bottleneck is no longer clever architectures but physical resources: GPUs, power, cooling, and land. Industry observers have already started describing compute as the real moat, and this deal fits that narrative perfectly. Anthropic is now simultaneously tied into Amazon’s Trainium roadmap, Google’s TPU roadmap, and NVIDIA’s Blackwell roadmap via Colossus, effectively hedging its bets across three different silicon ecosystems while securing priority access to whichever one manages to scale fastest. In a world where fabs are oversubscribed and even hyperscalers struggle to get firm delivery schedules, having this kind of diversified compute portfolio may matter more than any one model architecture choice.

There is a more awkward dimension too: Anthropic is now running core workloads on infrastructure controlled by a direct competitor. Musk’s xAI is racing to catch up with OpenAI, Anthropic, and others, and Colossus exists primarily to fuel Grok’s training and deployment. In practical terms, that means Anthropic’s operational and security teams will need strict isolation guarantees, tight observability, and contractual safeguards to make sure data and models remain logically and cryptographically separated, especially for enterprise and government customers. None of the public announcements go into detail there, but in the current regulatory climate around AI, any hint of entanglement between competing labs’ training data or parameter sets would be a non-starter for large customers and regulators alike.

Still, the fact that this arrangement exists at all tells you how warped the supply-demand equation for AI compute has become. Two years ago, the idea that top-tier labs would be sharing GPU campuses would have sounded almost absurd. Today, with GPU lead times stretching out, power constraints biting in key regions, and model sizes still climbing, it is starting to look like the new normal: infrastructure mega-projects that behave more like shared utilities than single-tenant citadels. Anthropic’s move onto Colossus 2 GB200 racks is one of the clearest signals yet that, at the frontier, the line between “competitor” and “customer” is blurring in favor of whoever can keep their models fed with electrons and GPUs.