Google and Amazon are building their own AI chips so they don't have to depend on NVIDIA forever. They've invested billions, but for now NVIDIA holds over 80% of the market.
TPU (Tensor Processing Unit) is the chip Google has made in-house for over ten years, tailored specifically for its own AI models. Trainium is Amazon's answer, produced through AWS (its cloud division, which is actually where Amazon makes most of its money). Both do the same thing as NVIDIA's GPU-class chips: train and run AI models. The difference is that when Google trains Gemini on TPU-class chips, or Amazon offers Trainium to its cloud customers, the money stays inside the company instead of going to NVIDIA.
The practical example is Anthropic (the company behind Claude, the model I use every day). Anthropic deliberately trains on TPU, Trainium and NVIDIA chips at the same time, so it doesn't stay hostage to one supplier that could dictate price or a waiting line. A strategy only the big players can afford: building on three different architectures at once.
Why does it even come to this. Two players, one motive: Google and Amazon. NVIDIA sells chips at a markup because nobody can say no to it, demand outstrips supply. If you're Google or Amazon spending billions on AI every year, every percentage point saved through your own chip is direct profit. So the investment in TPU and Trainium isn't a technological whim, it's a plain business reflex: your own brand instead of forever paying the supplier.
So far the result is mixed. Amazon already claims its new generation of Trainium comes in at half the price of a comparable NVIDIA chip, and Google is producing its own TPUs at a scale it has never had before. And yet NVIDIA still holds over four-fifths of the market, because the software most AI companies build on was written for its chips first. Switching architecture is expensive and slow, even when the alternative is cheaper.
Here's where the direction shows
It limps exactly where you'd least expect: not in the chip, in everything around it. NVIDIA doesn't just sell a piece of silicon. It sells a software ecosystem an entire industry has written code on for a whole decade. Google and Amazon can produce a cheaper, even faster chip, but they can't recreate that ecosystem in two years.
So I don't see TPU and Trainium as a race with a winner. I see them as a slow shift in balance. Every percentage point of market share NVIDIA loses is hard-won, and it stays that way for years to come.