OpenAI wants to buy Google's TPUs: the rivalry shifts from software to silicon

OpenAI wants to buy Google's TPUs: the rivalry shifts from software to silicon

🔎 Nvidia's biggest customer in the world turns to its direct rival

On October 27, 2026, The Information revealed that OpenAI — the world's largest buyer of Nvidia GPUs — plans to purchase Alphabet's TPUs to expand its computing capacity. The news, quickly picked up by Reuters and Channel News Asia, describes a historic shift: the generative AI champion is set to depend on the silicon of the company it competes with most directly on models.

Why now? Because every piece of the puzzle was already in place. Google has been selling its TPUs directly since this year, according to Reuters. Oracle has been offering them through its OCI cloud since October 27. And OpenAI's frontier models have been available on Google Cloud since the week before the announcement. Only one thing was missing: the move from renting to buying.

One clarification is needed right away: nothing has been signed. The report relies on a person close to the matter, the final terms — price, volume, timeline — have not been defined, and neither OpenAI nor Google has commented. But the direction of travel is no longer in doubt. The AI war is leaving the software layer and moving down into the silicon.

Key takeaways

• OpenAI plans to buy TPUs from Alphabet (The Information, October 27, 2026). Terms not finalized, both sides declining to comment.
• The dependence on Nvidia remains massive: Jensen Huang says OpenAI has committed to roughly $600 billion in Nvidia servers by 2030, after more than $100 billion in credit backing for a data center in Ohio.
• Google has already become a compute merchant: direct TPU sales since 2026, Anthropic among the buyers, and a promise from Alphabet's CFO — the bulk of TPU revenue in 2027.
• OpenAI is building its own chip, Jalapeño, with Broadcom — led by a former Google TPU engineer.
• The market has spoken: GOOGL stock jumped more than 1% in pre-market trading as soon as the news broke (Yahoo Finance, October 2026).

Recommended Tools

You won't be ordering TPUs one at a time, but the compute battle is reshaping the offering at every tier. Here's where to source your hardware, depending on your scale.

My advice: don't commit to a single silicon vendor until you've validated your workloads in real-world conditions. The market just handed you a bargaining chip — use it.

What the report says — and what it doesn't

The Information reports a purchase plan, not a signed contract. That's the entire nuance many headlines lost along the way.

The outlet cites, dated October 27, 2026, a person close to the matter. The final terms — price, volume, timeline — are not defined. OpenAI and Google decline to comment. This is news journalism about an ongoing negotiation, not a product announcement, and it should be read as such.

Around this revelation, however, verifiable facts keep piling up. The timeline summarizes the mechanics:

• OpenAI was already renting TPUs as of June 2025 (Reuters) — direct purchase changes the scale, not the nature of the relationship.
• OpenAI is actively recruiting engineers to optimize its frontier models on TPUs (Silicon Report, October 2026).
• Its frontier models — GPT-5.5 and GPT-5.4 Pro — have been accessible on Google Cloud since the week before the announcement.

Should you believe it?

Yes, more likely than not. An isolated rumor rarely checks out; a rumor backed by a rental started sixteen months earlier, targeted hiring, and an already-live commercial rollout looks like a trajectory, not a leak. My bet: official confirmation will arrive by mid-2027, at the exact moment Anat Ashkenazi promises Alphabet's TPU revenue will shift.

Why OpenAI is diversifying: Nvidia's weight in numbers

OpenAI isn't leaving Nvidia — it's protecting itself against a dependency that has become systemic. The numbers say it better than any editorial could.

First, the commitment: Jensen Huang states that OpenAI has committed to roughly $600 billion worth of Nvidia servers by 2030 (Silicon Report, October 2026). Next, the financing: Nvidia has extended more than $100 billion in credit support for a data center in Ohio. The supplier is thus lending its client the money that will, in part, be used to buy the supplier's own servers.

On the other side, demand is exploding: Amazon, Microsoft, and Alphabet plan up to $600 billion in AI spending this year (Reuters, October 2026). When a single supplier captures the bulk of this windfall and finances the demand itself, the rational customer looks for a second supplier. This is basic risk management, not a declaration of war.

The strategy is, moreover, openly acknowledged at OpenAI, which pursues a proudly claimed multi-sourcing policy (Bloomberg Línea, October 2026). Buying TPUs, developing Jalapeño with Broadcom, keeping Nvidia: three legs, one goal — never depending on a single silicon, whether for training or for inference.

Google is selling its TPUs to everyone — including its rivals

Google wasn't caught off guard: selling TPUs has become a strategic revenue stream, embraced at the highest financial levels.

The timeline proves it. Since April 2026, Google has been selling TPU hardware to a small group of customers (Reuters). Anthropic — OpenAI's direct rival — is among the buyers. And Google has sold so much TPU capacity that its own researchers have had to wait for access to the machines. When your researchers queue up behind your customers, you're no longer a lab: you're a foundry.

The financial trajectory is explicit. Anat Ashkenazi, Alphabet's CFO, expects a small share of TPU revenue in late 2026 and the vast majority in 2027 (Reuters, October 2026). Translation: the TPU is becoming a mass-market product, not a jealously guarded internal technology.

Anthropic perfectly illustrates this shift. OpenAI's rival has pledged $200 billion to Google Cloud — the largest cloud contract in history, locking in AI infrastructure — and is pushing that ambition all the way to targeting a $2 trillion IPO. Google is pocketing billions from both of its front-line rivals while they battle it out over models.

This compute ambition doesn't stop at Earth, either: with Project Suncatcher, Google wants to put its TPUs into orbit — solar-powered satellites for AI compute at 8x terrestrial power. When a company seriously contemplates selling space-based compute, it's because it has turned terrestrial compute into a full-fledged business.

Jalapeño: OpenAI's In-House Chip, Designed by Former Google Engineers

OpenAI isn't buying TPUs for lack of an alternative: it's building one in parallel. That's what makes the October 27 report so coherent.

According to Business Standard (October 2026), OpenAI is developing in-house chips dubbed Jalapeño with Broadcom, with the help of former Google engineers. A tasty detail: Richard Ho, who leads OpenAI's chip team, previously worked on... Google's TPUs. The silicon OpenAI is about to buy belongs to the same family as the one its own teams are learning to design.

Key takeaway: Jalapeño targets inference, where costs explode with hundreds of millions of users.

Why buy and build at the same time?

Because the two horizons don't overlap. A custom chip takes years to design, validate, and industrialize; Google's TPUs give OpenAI immediate inference capacity and a second credible supplier. The irony is dizzying — OpenAI is buying Google's silicon while former Google engineers design its own — but it's rational. Talent moves faster than patents, and buying and in-house development hedge each other.

Nvidia: the real underlying issue behind the controversy

The news doesn't weaken Nvidia in the short term, but it chips away at its narrative monopoly. And that matters more than a contract.

Nvidia remains indispensable: Oracle is expanding its agreement with Nvidia to a multi-gigawatt scale, with Cerebras and SambaNova chips as complements (Infobae, October 2026). OpenAI's commitment — roughly $600 billion in servers by 2030 — doesn't evaporate with a TPU purchase plan.

But look at the structure of the market. A supplier that lends more than $100 billion to its customers so they can buy its products is a pattern the IT industry has seen before — the telecom equipment makers around 2000, Cisco at the forefront. It usually ends when the customer finds a second source. That's exactly what OpenAI is doing.

My opinion: 2026 will remain the year when "buying compute" stopped automatically meaning "buying Nvidia." Not because TPUs would be objectively superior, but because Google opened up sales, Oracle is relaying them, and a buyer the size of OpenAI has shown the way. The alternatives — Cerebras, SambaNova, Jalapeño — no longer need to prove they can replace everything, only that they carry enough weight to make competition kick in.

The new compute map: everyone sells to everyone

In October 2026, compute alliances have ceased to be exclusive. Each giant is now simultaneously a customer, supplier, and competitor of the others. The table summarizes this unprecedented web of relationships:

Amin Vahdat, at Google, summed it up bluntly: the next generation of AI will require a broad range of compute resources (Bloomberg Línea, October 2026). This isn't press-release diplomacy — it's an engineer's assessment: no single architecture can cover the training, inference, and agent needs of a market that doubles every year.

Direct consequence: the boundary between competitors is becoming porous. OpenAI runs its frontier models on Google's cloud, which sells its chips to both camps, while Nvidia finances everyone. In this world, loyalty isn't a strategy — redundancy is.

What This Changes for Developers

Nothing changes for your API calls today — but the pricing trajectory and model portability are improving for the long term.

The first effect is mechanical: more silicon in competition means more total capacity and more pressure on prices. When Google sells TPUs to its own cloud, to Oracle, to Anthropic, and potentially to OpenAI, every inference dollar gets negotiated harder. Frontier models — GPT-5.5, Gemini 3.1 Pro, Claude Opus 4.7 — become accessible through multiple clouds: portability improves, and your dependence on a single vendor decreases.

The second effect is subtler: the battle shifts to the application layer. Google isn't content with just selling compute — it keeps rolling out product offensives with Antigravity 2.0, the agent-first suite that wants to kill Cursor and Claude Code and Gemini Spark, Google's 24/7 AI agent that wants to become your second brain. Compute is the foundation, but what you'll be buying tomorrow is the agent.

In the meantime, there's no need to over-engineer. To test frontier models without any commitment, go through the free AI APIs — Groq, Google, and OpenRouter cover most prototyping needs. And if your product boils down to consuming these APIs, simple hosting like Hostinger is more than enough: cutting-edge compute is the giants' problem, not yours.

❌ Common Mistakes

Mistake 1: "OpenAI is abandoning Nvidia"

This is the most widespread and most false reading. OpenAI remains committed to roughly $600 billion worth of Nvidia servers by 2030, according to Jensen Huang himself (Silicon Report, October 2026). The TPU purchase is a diversification, not a divorce. The fix: read this deal as an insurance policy, not a breakup.

Mistake 2: "Google and OpenAI are forming an alliance"

A commercial transaction is not an alliance. Neither company is commenting, the terms are not defined, and product competition remains head-on: Gemini 3.1 Pro and GPT-5.5 are battling for the top of the leaderboards. The fix: follow the contractual facts — signatures, volumes, billing — not the reconciliation narratives.

Mistake 3: "The TPU will replace the GPU, it's cheaper"

The TPU is optimized for Google's workloads and requires costly software porting. OpenAI is specifically recruiting engineers to optimize its frontier models on this hardware (Silicon Report, October 2026) — proof that the migration is anything but automatic. The fix: compare total cost of ownership, migration included, never a price per chip.

Mistake 4: "The deal is confirmed, we can plan around it"

No. Single anonymous source, undefined terms, refusal to comment from both parties. If you're making architecture decisions based on this information, make ones that would survive a cancellation: multi-cloud abstractions, short contracts, benchmarks on two silicon platforms. Prudence always costs less than a commitment made on a rumor.

❓ Frequently Asked Questions

Is OpenAI really going to buy Google TPUs?

The project is real according to The Information (October 27, 2026), but nothing is signed: price, volume, and timeline remain undefined, and both companies refuse to comment. OpenAI has already been renting TPUs since June 2025, its frontier models run on Google Cloud, and it is recruiting TPU engineers. The trajectory is credible; the signature is not yet secured.

Why is Google selling its chips to its own competitors?

Because TPUs have become a revenue machine. Alphabet's CFO, Anat Ashkenazi, expects a small share of TPU revenue by the end of 2026 and the vast majority in 2027 (Reuters, October 2026). Google has in fact sold so much capacity that its own researchers had to wait. Every chip sold to OpenAI or Anthropic funds the next generation of TPUs.

What exactly is a TPU?

A Tensor Processing Unit: a chip designed by Google for training and inference of AI models, used internally for nearly a decade before being sold. It is the most credible alternative to Nvidia's GPUs, known for being efficient on massive inference workloads — but with a software ecosystem that is more demanding to master than a standard GPU.

Is Nvidia threatened by this rapprochement?

Not in the short term: OpenAI remains committed to roughly $600 billion worth of Nvidia servers by 2030, and Oracle is expanding its Nvidia agreement to a multi-gigawatt scale (October 2026). But every TPU sold reduces Nvidia's share of future market growth — and the presence of Cerebras, SambaNova, and Jalapeño proves that the alternative now exists.

Will API prices drop thanks to this competition?

That's the logical medium-term consequence. More compute supply — TPUs at Google and Oracle, Nvidia GPUs, Cerebras, SambaNova — increases total capacity and intensifies competition on inference. OpenAI's frontier models are already multi-cloud. In the meantime, free APIs remain the best way to test without a budget.

✅ Conclusion

October 27, 2026 will go down as the day the AI war officially shifted from software to silicon: OpenAI wants to buy TPUs from its rival Google, Google is selling its chips to all of its competitors, and Nvidia is bankrolling its clients with hundreds of billions to keep its crown.

What comes next promises to be just as momentous: while OpenAI finalizes — or doesn't — its TPU purchase and fine-tunes Jalapeño, Google is pushing its ecosystem all the way into orbit. To make sure you don't miss any of this reshuffling, start by testing frontier models through free AI APIs: silicon wars almost always end with lower prices for developers.