Nvidia sells the shovels in the AI gold rush, and its best customers have started forging their own. Google, Amazon, Microsoft and Meta all now run custom silicon in their data centers.

Why the Biggest Buyers Became Chipmakers
The first reason is arithmetic. When a company spends heavily with a single supplier, that supplier’s margins become a line item worth attacking. Nvidia’s data center GPUs are priced like the scarce, in-demand products they are, and a company with a large enough fleet eventually asks whether designing a part in-house would cost less over its lifetime. For a cloud giant serving enormous volumes of AI requests every day, a small efficiency gain per chip compounds quickly.
The second reason is specialization. A GPU is a general-purpose parallel processor that began life rendering graphics and was later adapted to neural networks. It is very good at many things, which also means it carries capability that a given workload may never touch. A chip built to serve one family of models, or to rank recommendations for a social feed, can drop what it doesn’t need and spend that silicon and power budget on what it does.
Supply is the third. Allocation of Nvidia’s top-end parts has been tight for long stretches, and being dependent on one vendor’s production schedule is an uncomfortable place for a company whose product roadmap depends on compute. Designing your own chip does not fully cure this, since most of these designs are still manufactured by TSMC and compete for the same advanced manufacturing capacity. The bottleneck moves rather than disappears.
Control rounds out the case. A company that owns its chip also owns its roadmap, and can tune hardware and software together.
What the Custom Chips Actually Look Like
Google got there first. Its Tensor Processing Units predate the current AI boom and have gone through several generations, first built to run Google’s own neural networks and later offered to cloud customers. Amazon followed with Inferentia for running models and Trainium for training them. Microsoft has its Maia accelerator, paired with a Cobalt processor for general cloud work, and Meta has built its MTIA chips, which started with the recommendation systems that power its apps. None of these is a clone of an Nvidia GPU, and that is the point: each is shaped around what its owner runs most.
These companies rarely go it alone. Chip design firms such as Broadcom and Marvell supply design services, high-speed interconnect technology and packaging expertise that help turn an architecture into something a foundry can build. OpenAI has reportedly been working on its own chip with Broadcom, which suggests the impulse is not limited to the cloud giants. Any AI lab with a large enough compute bill can see the same math.

The customers of these customers matter too. Anthropic has publicly said it runs its Claude models on Amazon’s Trainium and Google’s TPUs alongside Nvidia hardware. That is a meaningful signal: a leading AI developer that depends on cutting-edge compute is willing to spread its workloads across several architectures rather than commit to one.
The obstacle is software. Nvidia’s CUDA platform has had more than a decade to accumulate libraries, tools, tutorials and trained engineers, and that ecosystem is the real moat. A new chip needs compilers, framework support and a pile of optimized kernels before developers will trust it with production workloads. Frameworks like PyTorch and JAX narrow the gap by hiding hardware details, but performance tuning at scale still tends to be hardware-specific. This is why most large operators run mixed fleets rather than swapping one architecture for another.
Nvidia’s answer has been to sell more than chips. It now pitches entire rack-scale systems, with its own networking, memory configurations and software stacks, as a single product. A hyperscaler can design a capable accelerator, but matching the whole package of interconnect, memory bandwidth and developer tooling is a far larger project. And these same companies remain some of Nvidia’s largest buyers, even as they build alternatives. Both things are true at once.

The Awkward Part of the Relationship
The cloud providers are in an odd position. They rent Nvidia GPUs by the hour to customers who ask for them, while also promoting their own chips, often at lower prices, to the same customers. Every workload that moves from an Nvidia instance to an in-house chip improves the provider’s margins and weakens Nvidia’s hold. Nvidia, for its part, has reasons to keep a market alive where buyers have options beyond the big three clouds, which helps explain why specialized GPU cloud providers have become a visible part of the landscape.
The real test happens far from the chip designers, at the level of a developer renting compute. If a startup can run the same model on either architecture and one is meaningfully cheaper, how long does loyalty to CUDA last?









