OpenAI plans to buy Google TPUs: what it means for your AI bill
OpenAI plans to buy Google TPUs, per The Information and Bloomberg. No terms yet. Why AI chip supply shifts matter to your API costs and model routing.
OpenAI plans to buy Google's TPUs, the custom AI chips Google designs in-house. The Information reported it on October 27, Reuters picked it up, and Bloomberg reported the same plan on October 28. Nobody has signed anything public. But the biggest Nvidia customer in AI is shopping elsewhere, and that will reach the per-token price you pay.
What actually happened
Here is what the reports confirm, and what they do not:
- The plan: OpenAI intends to buy Google TPUs, citing a person familiar with the matter. Final terms of a deal have not been set.
- No numbers: No chip count, dollar value or delivery date has been disclosed.
- No confirmation: OpenAI declined to comment to Reuters. Google did not immediately respond.
- Not the first step: Reuters reported in June 2025 that OpenAI had started renting TPUs through Google Cloud. Buying them is a bigger commitment.
The context matters. OpenAI is one of the largest buyers of Nvidia GPUs. It also said in June that it designed its first custom chip with Broadcom and plans to deploy it by the end of this year. Last week its frontier models went live on Google Cloud. Google began selling TPUs directly to some customers this year, and Anthropic is already a big TPU buyer.
So OpenAI's compute now has three possible sources: Nvidia, Google and its own silicon.
Why it matters for your business
You will never buy a TPU. But your inference bill comes from these chip deals. When a model vendor gets cheaper compute, the price cut usually shows up later, and not always for every model.
What we tell clients:
- Don't price a product on today's token rate. Hardware supply is moving every quarter. Build your margins on a range, and re-check them each time a vendor changes its price list.
- Expect the same model to behave a bit differently on new hardware. Numerical changes across chips can shift outputs at the edges. Keep a small eval set and run it when a vendor announces infrastructure changes.
- Keep your model layer swappable. One interface, many providers. When a cheaper option arrives, you switch a config value, not a codebase.
- Ignore the hype until terms exist. A plan with no volume and no price is a signal, not a discount.
Key takeaways
- The Information and Bloomberg report OpenAI plans to buy Google TPUs; no volume, value or timing disclosed
- OpenAI declined to comment; Google did not immediately respond
- OpenAI already rents TPUs (since 2025), buys Nvidia GPUs and is deploying a Broadcom-built chip
- Chip supply changes reach you through API prices, so price your product on a range
- Run a small eval set whenever a vendor changes its serving hardware
Locked to one model vendor's price list? We build AI systems with a provider-agnostic model layer and an eval harness, so you can move to the cheaper option the day it ships. See how we build it, or model your AI costs first.
Sources: Reuters via Investing.com, Bloomberg.
- #openai
- #google-tpu
- #ai-chips
- #nvidia
- #llm-costs
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Apple adds MCP to Pages, Keynote, Numbers: AI agents edit docs
Apple's Pages, Keynote and Numbers get MCP support on macOS 27, so outside AI agents can read and edit your documents. What to automate and what to lock down.
Read itSAP TechEd 2026: AI gateway, Joule Studio free till December
SAP TechEd 2026 adds an Enterprise AI Gateway, Joule Studio spend tracking and MCP/A2A agent access. Runtime is free through December 2026. Meter it now.
Read it