Internal Access Anthropic just signed a deal with SpaceX to use all 220,000 GPUs at the Colossus data centre in Memphis. Two years ago, this was unthinkable. The compute scarcity that defined AI in 2024 has forced even the most software-minded companies to play the hardware game.
Anthropic announced on Wednesday that they have signed a partnership with SpaceX to use all of the compute capacity at the Colossus 1 data centre in Memphis, Tennessee. The deal gives them access to more than 300 megawatts of power — roughly 220,000 NVIDIA GPUs — within the month.
They doubled Claude Code rate limits at the same time. Removed peak hour throttling for Pro and Max accounts. Raised API rate limits for Claude Opus models. All of it enabled by suddenly having access to infrastructure that, two years ago, would have been the kind of compute stockpile that defined who won the AI race.
It is a remarkable pivot for a company that famously chose not to play that game.
Around this time two years ago, the AI industry was quietly splitting into two camps. At OpenAI, Sam Altman was locking in every GPU, every data centre, every megawatt of power he could find. The Microsoft partnership was already feeding billions into Azure capacity. OpenAI was treating compute like a strategic resource — scarce, valuable, and worth hoarding.
At the same time, Dario Amodei was building Anthropic on a different conviction. The company would focus on model quality and safety research. Infrastructure would be someone else’s problem. They would rent what they needed, when they needed it, and let hyperscalers manage the capital expenditure. The philosophy was elegant: build the best intelligence, and the world will provide the compute.
It was a defensible position in 2024. The cloud had spare capacity, GPU supply was expanding, and the idea that compute would become a bottleneck seemed almost paranoid. Nobody was short of GPUs — yet.
By late 2024 and early 2025, the landscape had shifted. GPU demand outpaced supply. HBM memory shortages drove DRAM prices up by 50 per cent. NVIDIA’s CoWoS packaging capacity was sold out. Custom silicon — Google TPUs, AWS Trainium — could not absorb the overflow fast enough.
We wrote about this in detail back in February: the memory wall is real, and it means AI hardware gets more expensive before it gets cheaper. The physics of the problem do not care about business strategy. Every model that runs inference needs memory bandwidth, and the world has a finite number of fabs that can produce high-bandwidth memory.
The companies that had invested early — OpenAI with its data centres, Microsoft with its Azure commitments, Google with its TPU fabs — suddenly found themselves sitting on the most valuable resource in the industry. The companies that had not — Anthropic foremost among them — found themselves scrambling to secure capacity through every channel available.
What Anthropic announced on Wednesday is not an isolated deal. It sits alongside a cascade of other infrastructure commitments: up to 5 gigawatts with Amazon, 5 gigawatts with Google and Broadcom, $30 billion in Azure capacity with Microsoft and NVIDIA, and $50 billion invested in American AI infrastructure through Fluidstack.
They are also exploring orbital AI compute with SpaceX — literally running inference in space.
Taken together, these represent one of the largest compute procurement programmes in tech history. Anthropic is not just renting GPUs anymore. They are structurally changing their business to control infrastructure at a scale that Dario would have recognised as exactly the kind of heavy capital commitment the company was built to avoid.
The market voted on the Dario versus Sam bet. Compute is not a utility you can assume will be available. It is a strategic resource, and the companies that locked it in first are still sitting on the advantage.
None of this changes the fundamental equation for UK businesses. The global arms race for GPU capacity is happening at a scale that makes on-premises deployments for individual organisations look tiny by comparison — but that is precisely why on-premises AI is becoming more practical, not less.
The open-weight models that power our deployments run on hardware you own, in a building you control, with no dependency on anyone else’s infrastructure queue. When Anthropic is fighting for GPU capacity with SpaceX, Amazon, and Microsoft, the company deploying Qwen3.6-27B on a pair of RTX 4090s in a Leeds server room does not care. The model runs. The data stays. The capability belongs to them.
This is Capability Sovereignty in its most practical form. Not a theoretical concern about vendor lock-in — a material advantage that compounds every time a cloud provider throttles, a model gets deprecated, or a data centre runs out of power.
UK enterprises are already moving this way, quietly shifting workloads from cloud APIs to on-premises infrastructure. The Anthropic-SpaceX deal is not a reason to rush back to the cloud. It is a reason to remember why you left.
Investing in GPU and AI infrastructure early remains the smartest bet you can make. If you are a company building AI products, this means locking in hardware capacity before the next shortage. If you are a business using AI, this means deploying on hardware you own — because the global demand curve for artificial intelligence is going to grow exponentially over the next five years, and the companies that control their own compute will not have to beg anyone for capacity.
The Dario versus Sam bet is over. The next one is yours.
JD Fortress AI builds secure, on-premises AI infrastructure for UK businesses in regulated sectors. If you are looking at deploying AI and want full control over your hardware — no queue, no throttling, no vendor to negotiate with — get in touch for a confidential, no-pitch conversation.
If you're thinking about secure AI for your business, we'd love to have a conversation.
Get in Touch →