OpenAI is preparing to loosen its years-long dependence on Nvidia for AI hardware. To diversify its computing infrastructure, the company plans to buy Google's in-house TPU (Tensor Processing Unit) chips. At the same time, it is working to develop its own custom chip and position itself against rising infrastructure costs.

What happened?

OpenAI has moved to source Google's TPUs for training and running models. This marks the company's first serious break from its habit of relying on a single supplier — namely Nvidia — for its computing power. The decision coincides with OpenAI's simultaneous efforts to design its own chip and spread out its cloud resources.

Why does it matter? Why is Nvidia's monopoly cracking?

This is where the irony begins. Google is OpenAI's direct rival both in the cloud market and on the model side. Gemini goes head-to-head with ChatGPT; Google Cloud competes with Microsoft Azure, OpenAI's partner. In the midst of such rivalry, OpenAI buying chips from Google shows just how severe the computing shortage in the industry has become. Even competitors are turning into necessary allies in order to secure enough processing power.

The real issue is Nvidia's position. In the AI boom of recent years, training hardware has become almost synonymous with one thing: Nvidia's H100 and later GPUs. This de facto monopoly gave Nvidia extraordinary pricing power and high margins. The gap between the cost of a GPU and its selling price was by far one of the widest in the industry. When one of the largest buyers, like OpenAI, turns to an alternative, that bargaining balance begins to shift.

Google's TPU has been around for years; Google runs much of its own search, translation and Gemini workloads on these chips. But for a long time the TPU was seen as a "Google-only solution that wasn't really opened up to outsiders." OpenAI preparing to use this hardware at serious scale means the industry is effectively "validating" the TPU as a genuine alternative to Nvidia. That validation is more than a technical detail; it sends the market the message that "you can train large models without Nvidia."

Who is affected, and how?

The consequences play out in different directions for different players.

PartyImpact
NvidiaPressure on pricing power and margins; the monopoly perception weakens
GoogleA chance to open the TPU to the external market, with cloud revenue potential
OpenAISupplier diversity, leverage in cost negotiations
Other AI companiesLegitimization of a non-Nvidia option

For OpenAI, the logic is clear: being tied to a single supplier means both price and supply risk. A second serious source provides leverage at the negotiating table. Even just being able to say "I have an alternative" when sitting down with Nvidia works in its favor on the cost side.

For Nvidia, the picture is more complicated. Demand still far exceeds supply, so there is no sales problem in the short term. But the largest customers seeking alternatives raises questions about the long-term sustainability of those unique margins. The real question investors are watching isn't sales volume; it's how much of that high profit margin can be protected.

What's next?

There are a few things to watch. First, the actual scale of OpenAI's TPU use: is it an experimental deployment, or a serious share of production workloads? Second, when OpenAI's own custom chip project will bear fruit — if it succeeds, dependence on both Nvidia and Google decreases. Third, whether other major AI companies will follow this example.

This move looks like a sign that the unipolar era in AI hardware is slowly evolving into a multipolar structure. Nvidia is still the clear leader; but the industry is crossing a threshold where leadership is shifting from being "the only option" to "competing to be the best option." Carrying the competition over to the hardware side could lower costs in the long run — and that would be a positive development not just for the giants, but for everyone who uses AI.