Jensen Huang said TPUs only existed because of Anthropic. Then Oracle opened the door to every cloud customer who doesn't want to be Nvidia's tenant forever.
The Summary
- Google and Oracle announced TPUs will be available to Oracle cloud customers starting next year — the first time Google's custom AI chips have been sold through another hyperscaler's platform
- OpenAI is hiring engineers to optimize models for TPUs and reportedly plans to buy TPUs from Google, directly contradicting Nvidia CEO Jensen Huang's earlier claim that "without Anthropic, why would there be any TPU growth at all?"
- Anthropic's IPO filing revealed $114.5 billion in planned Google compute spending through 2033, signaling the scale of demand for alternatives to Nvidia's GPU monopoly
- Multi-chip strategies are becoming the hedge bet for companies that don't want to be locked into a single vendor's roadmap and pricing power
The Signal
Nvidia still owns the AI chip market. But the Oracle deal changes the distribution game. For the first time, Google's TPUs aren't just available through Google Cloud or to companies willing to co-locate in Google data centers. They're inside Oracle's infrastructure, which means any Oracle customer can now spin up TPU instances without migrating their entire stack to Google. That's a different threat surface than Nvidia faced six months ago.
The numbers tell the story of why customers want options. Anthropic plans to spend $114.5 billion on Google compute over seven years. Even if only half goes to TPUs, that's enough scale to justify Google treating chips as a real product line, not just internal infrastructure that happens to serve a few external customers. OpenAI quietly hiring TPU optimization engineers is the tell. When your biggest competitor is also your potential customer, the market structure is shifting.
"Jensen Huang dismissed TPUs as an Anthropic-only phenomenon. Now OpenAI is hiring engineers to run on them."
Google built TPUs a decade ago because they needed custom silicon for search and YouTube recommendations. The chips were optimized for inference, not training, which is where Nvidia made its fortune. But the agent economy runs on inference at scale. Every chatbot query, every API call to Claude or GPT, every autonomous action an AI takes is inference. Training happens once. Inference happens billions of times a day.
The hyperscaler move matters because it turns TPUs from captive infrastructure into a product with distribution. Oracle customers who want to run AI workloads now have three choices:
- Nvidia GPUs (dominant, expensive, supply-constrained)
- Google TPUs (optimized for inference, now available without switching clouds)
- AMD or other alternatives (still marginal in AI)
That's not a revolution. It's the start of optionality. And optionality erodes pricing power.
The Implication
Watch for more hyperscalers to cut deals like this. Microsoft has its own custom chips. Amazon has Trainium and Inferentia. The pattern is the same: cloud providers built custom silicon to reduce their own costs, then realized they could sell access to it and undercut Nvidia's margins while still making money. If you're building AI infrastructure, this is your signal to design for multi-chip compatibility from day one. If you're investing, the monopoly premium in Nvidia's stock price just got shakier.
The agent economy needs compute at scale. The question is whether that compute gets rented from one vendor or distributed across several. This Oracle deal is a vote for distributed.