Google and Amazon Try to Break Nvidia's GPU Grip Through PyTorch

Following Amazon, Google has released a tool that makes it easier for developers using PyTorch to move their applications from Nvidia GPUs to its own chips. The move shows the AI chip race shifting from hardware to the software layer. Nvidia's defense now rests less on processor speed than on the CUDA ecosystem.

What happened?

Both cloud giants are taking the same approach: running a model written in PyTorch on their own accelerators without having to rewrite the code from scratch. At Google, the target is TPUs; at Amazon, it is the Trainium and Inferentia chips. The goal of the tools is to reduce switching friction. In this move, a chip's on-paper performance takes a back seat.

Why does it matter?

Nvidia's market dominance is often explained by the superiority of its chips. In reality, the real lock-in is more than a decade of accumulated CUDA libraries, compiler tools and developer habits. If an engineer has optimized a model for CUDA, moving to another chip is not just a matter of moving files. It means rewriting custom kernels, changing debugging habits and relearning performance behavior. That cost often wipes out the rival chip's price advantage.

That is why the price/performance claims of TPU and Trainium have found only limited traction so far. Google has used TPUs at serious scale for its own internal workloads and a handful of large customers, but for the broader developer base the software layer has always been more laborious. Amazon ran into a similar obstacle with Trainium. The reason both companies are looking for a solution at the PyTorch layer is clear: PyTorch is the de facto standard in research and application development. If a developer already writes PyTorch, changing the hardware underneath should be a one-line decision.

AMD's ROCm effort struggled with the same problem for years. Even though the hardware was competitive, software support was inconsistent. Some libraries arrived late, certain operations ran slowly and documentation was incomplete. AMD was also a standalone chip vendor. It had no cloud platform to carry developers over, no billing relationship and no ready-made customer base.

The position of the hyperscale cloud companies is different here. They control the chip, the platform and the customer relationship all at once. They can tie a migration tool to managed services, credits and pricing incentives. They are also Nvidia's biggest customers. An alternative chip doubles as a bargaining chip in negotiations.

What is changing?

If the cost of switching falls, the decision logic changes. Today, a startup's GPU choice mostly rests on the reflex that "everyone uses CUDA." If migration tools mature, the choice could turn into a measurable criterion such as cost per token or per training hour. This could be especially effective on the inference side. Inference workloads are more predictable and less specialized, so they are easier to migrate than training.

Still, caution is warranted. The "migrate without changing code" promise works for simple models. When it comes to custom kernels, finely tuned performance optimizations or new architectures, compatibility gaps are likely to appear. Even if migration gets easier, the first release may bring performance losses or missing operator support. The success of these tools depends not on the demo on a marketing page but on stability in real production workloads.

Nvidia is not standing still either. It continues to strengthen its own software stack, libraries and cloud partnerships. It also has the advantage of always offering its newest chips to its own ecosystem first.

What's next?

There are three things to watch. First, how much performance loss independent developers experience with these tools in real projects. Second, whether other big players such as Microsoft and Meta will invest in similar layers. Third, whether the PyTorch community's work on hardware-agnostic compilers gains momentum.

Here is my assessment: Nvidia's leadership is not at risk in the near term, but its bargaining power is eroding. By lowering the cost of switching, cloud companies are making the GPU stop being the only option. That could create more lasting pressure on prices than chips alone could. For more information, see PyTorch's official website.