Anthropic has quietly assembled one of the most aggressive compute procurement strategies in Silicon Valley history, locking down vast tranches of GPU capacity through a complex web of agreements with hyperscalers, specialist neoclouds, and repurposed crypto infrastructure. The manoeuvre positions the Amazon-backed startup to train its next generation of frontier models without relying on a single provider, a strategic hedge that is simultaneously redrawing the economics of the AI cloud market.
The Great GPU Land Grab
For the better part of eighteen months, Anthropic’s infrastructure team has operated with the urgency of a wartime procurement office. The goal was never simply to rent servers; it was to secure guaranteed, long-term access to the H100 and upcoming Blackwell architectures that Nvidia cannot ship fast enough. Sources close to the negotiations describe a “deal spree” that saw the company sign contracts worth an estimated $2 billion across more than a dozen distinct providers.
This was not a standard cloud migration. Anthropic bypassed the traditional queue at Amazon Web Services—despite Amazon’s $4 billion investment—and instead constructed a distributed compute fabric. The portfolio now includes dedicated capacity from Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure, alongside specialist “neocloud” operators such as CoreWeave, Lambda Labs, and Crusoe Energy Systems.
The logic is ruthless pragmatism. By fragmenting its demand, Anthropic avoids the “noisy neighbour” problem where a single provider’s oversubscription degrades performance. It also creates competitive tension at the contract renewal table, a luxury few AI labs can afford.
From Hash Rate to Inference Rate
Perhaps the most unconventional pillar of this strategy involves former cryptocurrency miners. As the Ethereum merge rendered GPU mining obsolete, a generation of data centre operators found themselves sitting on high-voltage power connections, industrial cooling, and rack space—but no workload. Anthropic moved quickly to absorb this stranded asset class.
Deals with firms like Hut 8 and Applied Digital have seen former mining sites in Texas and North Dakota retrofitted for high-density AI training. These facilities often lack the sophisticated networking fabric of a tier-one cloud region, but they offer something hyperscalers cannot: speed of deployment. Where a new AWS availability zone takes years to permit and build, a converted mining shed can be racked and stacked in months.
One infrastructure engineer involved in the retrofits described the work as “plumbing for intelligence.” The challenge is less about the chips and more about the megawatts. Anthropic has effectively become a major player in the energy markets, negotiating power purchase agreements directly with utilities to ensure its contracted sites remain online during grid stress events.
The Amazon Paradox
The relationship with Amazon Web Services remains the most scrutinised aspect of this architecture. On paper, Anthropic is an AWS flagship customer, committed to using Trainium and Inferentia chips for future workloads. In practice, the startup has hedged heavily against silicon lock-in.
Nvidia GPUs still dominate the training clusters for Claude 3.5 Sonnet and the forthcoming Opus successor. Anthropic engineers have been candid internally about the performance gap between Nvidia’s CUDA ecosystem and Amazon’s homegrown Neuron SDK. The multi-cloud strategy buys the company time to port its stack to Trainium without stalling its roadmap.
It also serves as leverage. Amazon’s recent decision to double down on its own foundation model efforts—reportedly codenamed “Olympus”—creates a structural conflict of interest. By maintaining credible exit options on Google TPUs and Azure ND-series instances, Anthropic ensures it remains a partner, not a captive subsidiary.
The Neocloud Intermediaries
Specialist providers like CoreWeave have emerged as the critical middle layer in this stack. Backed by billions in debt financing secured against their GPU collateral, these neoclouds act as capacity aggregators. They absorb the complexity of managing thousands of GPUs across disparate physical sites, presenting Anthropic with a single API endpoint and a unified billing model.
This arrangement shifts capital expenditure off Anthropic’s balance sheet. The startup pays a premium per GPU-hour compared to a direct hyperscaler contract, but it preserves cash for research talent and avoids the multi-year capex commitments that could constrain a pivot to a new architecture.
The model is spreading. Competitors including Cohere and Mistral are now replicating the multi-cloud, neocloud-heavy approach, driving a secondary market for GPU-backed debt instruments that Wall Street is only beginning to understand.
Why it Matters
Anthropic’s compute offensive signals the end of the “single-cloud” era for frontier AI. The company has proven that the moat in this industry is not model architecture—it is the ability to secure, finance, and operate heterogeneous infrastructure at a scale that dwarfs traditional enterprise IT. As the industry pivots from training to inference, the same distributed fabric that built Claude will determine who can serve millions of token-generating users profitably. The winners will not be those with the smartest researchers, but those who mastered the plumbing first.
