The Cloud Is Being Rebuilt Around AI Demand

Anthropic has reportedly signed another $35 billion computing agreement, this time with Lambda (Reuters, 2026). On its own, that is a large cloud contract. Put beside the company’s other recent commitments, and it looks more like a map of where the infrastructure market is heading.

Anthropic has also agreed to spend $45 billion on capacity from Nscale in West Virginia (Financial Times, 2026), announced a $50 billion US infrastructure build-out with Fluidstack (Anthropic, 2025), and agreed to pay SpaceX for computing capacity (SpaceX Prospectus, 2026). That last agreement includes lower payments during its initial ramp-up and can be terminated with 90 days’ notice after the first three months.

That's approximately $175 billion across four deals. Naturally, the obvious question is why a model company needs so many suppliers. Part of the answer is simply scarcity. Training and running frontier models requires enormous clusters of high-end GPUs, plus power, cooling, and data centres to support them. Those facilities cannot be spun up overnight, and the companies building the largest models do not want their growth constrained by whichever provider happens to have capacity available.

So, they are locking it in early. The Lambda deal, for example, centres on roughly 350MW of capacity at a data centre in Nueces County, Texas, being developed by Hut 8. Nvidia is reportedly the leaseholder, while Lambda supplies the computing infrastructure. Well, that arrangement says quite a lot about the new cloud market.

A few years ago, enterprise cloud infrastructure mostly meant AWS, Microsoft Azure or Google Cloud. They remain dominant, but another layer is forming around them: specialist GPU clouds, data-center developers, chipmakers, energy providers and financiers all assembling pieces of the same infrastructure stack.

Lambda is one of the companies benefiting from that change. Its business is built around giving customers access to large quantities of Nvidia hardware without requiring them to build their own facilities. The company is now reportedly discussing a funding round of up to $3 billion at a valuation of $12 billion or more (Data Center Dynamics, 2026). It had already raised more than $1.5 billion in November 2025.

Microsoft has also signed a multi-year agreement with Lambda to deploy infrastructure using tens of thousands of Nvidia GPUs, including GB300 NVL72 systems (Lambda, 2025).

Nvidia keeps appearing on both sides of these deals for an obvious reason. It sells the chips at the center of the demand boom, but it has also increasingly used investment and commercial partnerships to help expand the infrastructure capable of running them. When more GPU cloud capacity is available more Nvidia hardware ultimately can be deployed. As a result, the cloud market is becoming much more capital-intensive and much less tidy.

Instead of buying computing capacity as needed, model developers are reserving vast amounts years in advance. Providers, in turn, need billions in financing before those customers ever run a workload. Lambda, for example, has been using debt to finance GPUs tied to specific customer deployments, including a $926 million loan for Nvidia GB300 infrastructure and another $1 billion debt package to acquire chips for Microsoft (TechCrunch, 2026).

Data-center developers need land and power commitments. Chip suppliers need visibility on future demand. And cloud providers increasingly need financing structures that look less like software and more like heavy industry.

Cloud computing used to sell flexibility: rent what you need, when you need it. But Frontier AI is pushing in almost the opposite direction. The largest customers are now behaving more like industrial buyers, securing critical infrastructure in advance because they cannot afford to discover later that somebody else booked it first.