How Cloud Companies Are Becoming AI Hardware Companies

For most of the cloud era, the big providers rented out other people’s hardware. They bought servers, networking gear, and chips from vendors like Intel, Nvidia, and Dell, and sold computing power on top. That model is changing. Amazon, Google, and Microsoft now design their own AI processors, sell access to them at enormous scale, and in some cases are being asked to supply chips to rivals. The line between a cloud company and a hardware company is getting thin. For more on the industry behind it, see our AI chip and semiconductor coverage.

From renting hardware to designing it

The change has been gradual. Google began building TPUs for its own services, then offered them through Google Cloud. Amazon started with networking and Graviton CPUs through Annapurna Labs before moving to AI accelerators. Microsoft is the newest entrant with Maia. We covered each strategy in why Google designs its own AI hardware, how Amazon is building its own AI silicon, and why Microsoft is building its own AI chips.

What’s new is the commercial scale. These chips are no longer internal cost-savers. They’re products that customers sign long-term contracts for, with gigawatt-sized capacity commitments attached.

Amazon: a chip business inside a cloud business

Amazon offers the clearest example. A May 2026 industry review quotes AWS CEO Matt Garman saying AWS had already deployed more than 1 million Trainium processors and was selling them as fast as it could produce them, while CEO Andy Jassy called it a multibillion-dollar business. A separate report said Amazon’s chip business had reached a $20 billion revenue run rate, a figure that suggests a standalone silicon company hidden inside AWS.

The customers are major AI labs. The same industry review says Project Rainier had roughly 500,000 Trainium2 chips running for Anthropic by October 2025, and that AWS also confirmed an OpenAI deal for 2 gigawatts of Trainium capacity. Trainium4 is also reported to support Nvidia’s NVLink Fusion, which would allow hybrid clusters mixing Trainium and Nvidia GPUs.

Google: selling TPUs beyond its own walls

Google has gone furthest in opening its chips to outsiders. Anthropic announced access to up to one million TPUs in October 2025, in a deal CNBC described as worth tens of billions of dollars. In April 2026, Anthropic announced a further agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity starting in 2027, and The Register reported the figure as 3.5 gigawatts.

Other customers are following. An industry review says Meta entered talks for multi-billion-dollar TPU deployments earlier this year, and FactSet noted that Alphabet’s first external TPU deals with Anthropic and Meta have helped it stand apart from the other hyperscalers. It is a striking turn: Meta has its own chip program and a stake in the custom-silicon race, yet it is still reported to be buying capacity on a rival’s chips. We covered Meta’s side in why Meta needs massive AI infrastructure.

Microsoft: catching up

Microsoft is earlier in the journey but heading in the same direction. Its Maia 200 is in production in Iowa and Arizona, according to Data Center Dynamics, and reports suggest Anthropic may rent Maia 200 servers to run its models. If that happens, Maia would move from an internal tool to an external product, following the path Trainium and TPU have already taken.

Why the model is attractive

Margins. Selling capacity on your own chips avoids paying Nvidia’s margin on every accelerator, and the saving can either improve profit or let the provider undercut rivals on price. This is the economics we described in why Big Tech is spending billions on AI infrastructure.

Customer stickiness. Chips come with software stacks, compilers, and tools. A customer that builds around Trainium or TPUs invests in that ecosystem, which can make it harder to leave.

Control over supply. Owning the design means a provider can set its own roadmap and tune the chip, the rack, and the data center together, which we examined in why Big Tech companies are building their own data centers.

Anchor customers. AI labs consume enormous amounts of compute. Landing one of them on your silicon fills a data center and funds the next chip generation.

The role of design partners

Becoming a hardware company doesn’t mean doing everything alone. The Register reports that Broadcom has a long-term agreement to develop and supply custom TPUs for Google’s future generations, and it quotes Broadcom CEO Hock Tan arguing that hyperscalers lack the skill to build custom accelerators on their own. That is a self-interested view from a supplier, but it reflects a real pattern: partners like Broadcom handle much of the physical design work, while the cloud companies define what they need.

Manufacturing is outsourced too. As we explained in why TSMC is so important to the global AI industry, these chips are fabricated at TSMC, so cloud companies building silicon compete for the same advanced nodes and packaging capacity as everyone else.

Nvidia isn’t being replaced

It would be wrong to read this as the end of Nvidia’s role. Anthropic, for instance, says it trains and runs Claude on AWS Trainium, Google TPUs, and NVIDIA GPUs so it can match workloads to the chips best suited for them, and Amazon remains its primary cloud provider and training partner. Big customers are using several chip types at once. Nvidia is responding by selling full systems, as we covered in why Nvidia is expanding beyond GPUs, and by working with rivals through standards like NVLink Fusion.

The risks

The strategy carries real risks. Chip design is expensive and unforgiving, and schedules can slip, as Microsoft’s Maia 200 reportedly did. Software maturity matters: developers are used to Nvidia’s tools, so each provider has to make its stack easy to adopt. Concentration is another issue, since a few large AI labs account for much of the demand, and Broadcom flagged in a filing that Anthropic’s ability to use additional compute depends on its ongoing commercial performance. Finally, running a chip business adds supply-chain exposure to the same bottlenecks we described in why semiconductor supply chains matter for AI.

The bigger picture

Cloud companies are becoming AI hardware companies because, in the AI era, the chip is the product. Owning the silicon lets a provider control costs, differentiate its service, and lock in the biggest customers, while the chip itself becomes a revenue line. The result is a market where Amazon, Google, and Microsoft compete as chip designers as well as cloud vendors, while still depending on Nvidia, Broadcom, and TSMC. For readers following the shift, the signals to watch are external customer wins, capacity deployed on custom chips, and how much of each provider’s AI revenue comes from its own silicon. Follow our AI hardware coverage as the line between cloud and chip company keeps blurring.