OpenAI Reportedly Plans to Buy Google’s TPU Chips. Neither Company Has Confirmed It. Is This the End of Nvidia’s Lock?

By Thorsten Meyer | November 1, 2026 | ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

First, the honest status: this is a report, not an announcement. On Tuesday, 27 October, The Information reported that OpenAI plans to buy Google’s AI chips, the tensor processing units (TPUs) that Google designs for its own data centers. Bloomberg reported the same plan a day later. Both cite one person familiar with the matter. OpenAI declined to comment, Google did not comment either, and nobody has published a price, a number of chips or a delivery date.

So why does an unconfirmed report deserve a full article? Because the parts around it are confirmed, and they add up to a real shift. Google now sells its chips outside its own cloud. Oracle announced on the same Tuesday that it will rent them out from 2027. Anthropic’s IPO filing lists hundreds of billions of dollars in TPU commitments. And OpenAI, long known as one of Nvidia’s biggest customers, has spent the past year signing chip deals with almost everyone else. I checked each of those pieces at its source. Here is what is real, what is reported, and what it means for prices, for cloud customers and for the people who use ChatGPT.

Key takeaways

• Reported, not confirmed. The Information (27 October) and Bloomberg (28 October) report that OpenAI plans to buy Google’s AI chips, each citing one person familiar with the matter. Terms are unknown. OpenAI declined to comment; Google did not comment. Treat it as a strong report, not a deal.
• There is circumstantial evidence. OpenAI’s own careers page lists a “Software Engineer, TPU Compilers” job, posted on 20 October, to “lead the optimization of OpenAI’s state-of-the-art frontier models on TPUs”. That shows serious TPU work. It does not prove a purchase.
• Google really does sell TPUs now. Since the second quarter, Google has delivered TPU systems into customers’ own data centers. On its earnings call on 28 October it said nearly 100 customers use TPUs, named Anthropic and Apple, and announced partnerships with Oracle and Crux AI. Oracle will offer Google TPUs in its cloud from 2027.
• OpenAI’s “Nvidia lock” was already loosening on paper. Besides Nvidia (10 gigawatts in a 2025 letter of intent), OpenAI has announced deals with AMD (6 gigawatts), Broadcom (10 gigawatts of its own custom chips), Cerebras (750 megawatts) and Amazon (about 2 gigawatts of Trainium). Google would be supplier number six, and the first that is also a direct rival in AI.
• The real change is leverage, not liberation. Nvidia is still the default. But when the biggest buyers can credibly move work to other chips, they negotiate better prices and faster deliveries. Whether any of that reaches you as a cheaper or faster ChatGPT is not known.
• Nothing changes for users today. Most of these chips arrive in 2027 or later.

Products Worth Considering

Apple MacBook Air 13-inch with M5 chip, 16GB memory, 1TB SSD, and 12MP Center Stage camera (Midnight)

AmazonView Latest Price

Amazon Kindle Scribe Colorsoft 64GB (1st Generation)

AmazonView Latest Price

Jackery Explorer 500 v2 Portable Power Station, 512Wh, 500W

AmazonView Latest Price

As an affiliate, we earn on qualifying purchases.

What was actually reported

Here is the state of play as of Sunday, 1 November, sorted by how solid it is.

• Reported by The Information (27 October). Under the headline “OpenAI, a Longtime Nvidia Customer, Plans to Buy Google’s AI Chips”, the outlet reported the plan, citing a person familiar with the matter. The article is behind a paywall. Reuters, which relayed it the same evening, wrote that financial terms and other details could not be learned, that OpenAI declined to comment, and that Google did not immediately respond.
• Reported by Bloomberg (28 October). Bloomberg reported that OpenAI intends to buy AI computing chips from Google as part of its AI build-out, also citing one person, and that final terms are not yet known. I could not read Bloomberg’s paywalled article in full. The details here come from outlets that cite it, including The Asia Business Daily, and they match.
• Echoed, not new. A TechRadar article on 1 November framed it as OpenAI’s “first move away from depending solely on Nvidia”. As you will see below, that is not accurate. OpenAI signed large non-Nvidia chip deals more than a year ago.
• Primary evidence of TPU work. On OpenAI’s careers page, the “TPU Compilers” role was published on 20 October 2026. The posting asks for “practical experience optimizing machine learning workloads on TPUs”. An older posting for its accelerators team, from June 2025, already asked for experience with “TPUs, custom silicon, exploratory architectures”. Hiring for a chip is what you do before you use a lot of it. It is still not a purchase order.
• Not known. The number of chips, the TPU generation, the price, the delivery dates, and where the chips would run: in Google’s data centers, in OpenAI’s own sites, or at a partner such as Oracle.

Two earlier facts help read the report. In June 2025, Reuters reported that OpenAI had started renting Google’s TPUs through Google Cloud. Days later, OpenAI told Reuters it was only in early testing with some TPUs and had no active plans to deploy them at scale. And on 21 October 2026, a week before the report, OpenAI said it would make its frontier models available through Google Cloud. The two companies compete in search and chatbots, but they are increasingly doing business with each other.

Google turned its chips into a product

For a decade, TPUs were Google’s private advantage. You could rent them in Google Cloud, but you could not buy them. That changed this year, and Google’s own earnings calls document each step.

29 April 2026 (first-quarter call). Chief executive Sundar Pichai announced the change: “As TPU demand grows from AI labs, capital markets firms, and high performance computing applications, we will begin to deliver TPUs to a select group of customers in their own data centers in a hardware configuration.” Chief financial officer Anat Ashkenazi warned that this revenue “will fluctuate from quarter to quarter, depending on when TPUs are shipped”, with a small part expected in 2026 and the majority in 2027.

22 July 2026 (second quarter). Ashkenazi said Google “started delivering TPU systems to customer data centers in the second quarter” and expected “a relatively small portion of the revenues from our existing TPU system sales agreements this year, ramping as we exit 2026.”

28 October 2026 (third quarter). This is the call that came a day after the OpenAI report, and OpenAI was not mentioned in it. Google Cloud revenue rose 103% to $30.7 billion, and “TPU system sales were also a meaningful contributor,” Google said. Google Cloud chief Thomas Kurian said “nearly 100 customers use TPUs today”, named Anthropic, “both directly and via Broadcom”, and Apple, which is “scaling pre-training of their foundation models on the latest generation of TPUs”. He also announced partnerships “to deliver TPU systems with both Crux AI and Oracle, with more in the works.” Ashkenazi’s outlook: “we expect TPU system sales to continue to scale in Q4, followed by a meaningful revenue ramp in 2027.”

Two details from that call matter for the bigger picture:

• Google insists it sells systems, not chips. “A TPU system is not just a chip,” Ashkenazi said, but hardware, storage, networking and software assembled together. That is how Nvidia sells, too: complete racks, not loose processors.
• Google is paying up front. Ashkenazi said Google builds TPU inventory before the revenue arrives, and that this inventory build weighed on cash flow in the quarter. Selling chips is a new kind of business for Google, with new risks.

Google’s performance claims for its newest chips are its own: Kurian said the eighth-generation TPU 8t and 8i offer “up to 2.7 times better training performance and 80% better inference price performance respectively, versus our previous generation of TPUs.” Note the comparison point: Google’s previous TPU, not Nvidia.

Oracle (27 October). At its AI World event, Oracle announced that Google TPUs will be available in Oracle Cloud Infrastructure, “the first time Google TPUs will be offered on a hyperscale cloud platform outside Google”. The timing is “calendar year 2027, ramping over a multi-year period”. Oracle provides the power, cooling and networking. No prices were published.

For how investors judged Alphabet’s quarter, including the spending behind it, see Same night, opposite verdicts: why investors rewarded Microsoft and punished Alphabet.

Anthropic shows what a big TPU customer looks like

A disclosure first: Anthropic makes Claude, the AI assistant that helped research and draft this article. It is also one of OpenAI’s main competitors, and its filing shows that Google is an investor in it. I quote its numbers because they are the only detailed, legally filed picture of what TPU buying at this scale means.

Anthropic filed its registration statement for a stock market listing (Form S-1) with the US Securities and Exchange Commission on 26 October. In it:

• Google agreement. In March 2026, Anthropic agreed to spend “at least approximately $114.5 billion between April 2026 and July 2033” with Google, “regardless of actual usage”. The filing does not say which share of that is TPU capacity.
• More TPUs later. Anthropic “also committed to procure additional TPU compute from TPU providers, including Google, to be delivered between 2028 and 2031”, which it estimates at about $125 billion to $150 billion.
• TPU racks leased through Broadcom. Five-year equipment leases with Broadcom give Anthropic “access to TPU-based compute capacity”, with rent of about $161.2 billion excluding interest. Part of it came online in 2026, more from 2027.
• Four chip families. Anthropic says it runs on Amazon’s Trainium, Google’s TPUs, Nvidia GPUs and AMD GPUs, and moves work between them.

That last point is the template OpenAI appears to be following: no single supplier, and software that can move between chip types. For how Broadcom finances part of Anthropic’s chips, see Who really pays for Anthropic’s chips?.

OpenAI’s chip suppliers, checked one by one

Here is every major chip deal OpenAI has announced, with its status. The gigawatt figures describe the power of the planned data centers, not delivered hardware, and the deals have very different legal weight.

• Nvidia: at least 10 gigawatts (letter of intent, 22 September 2025). OpenAI and Nvidia announced plans to deploy “at least 10 gigawatts” of Nvidia systems, with Nvidia intending to invest “up to $100 billion” in OpenAI as each gigawatt comes online, and the first gigawatt on Nvidia’s Vera Rubin platform “in the second half of 2026”. A letter of intent is not a contract. As CNBC noted in January, Nvidia itself said in November 2025 that there was “no assurance” the deal would be completed as planned.
• AMD: 6 gigawatts (definitive agreement, 6 October 2025). A multi-year deal across several generations of AMD Instinct chips, starting with 1 gigawatt of MI450 chips in the second half of 2026. AMD gave OpenAI a warrant for up to 160 million AMD shares that vests with deployment milestones and share-price targets.
• Broadcom: 10 gigawatts of OpenAI’s own chips (announced 13 October 2025). OpenAI designs the accelerators, Broadcom builds them. The first one, called Jalapeño, was unveiled on 24 June 2026: an inference chip, meaning it runs models rather than trains them, with initial deployment planned by the end of 2026. Broadcom’s chief executive Hock Tan told CNBC it would “really ramp up in ’27”.
• Cerebras: 750 megawatts (14 January 2026). A multi-year agreement for Cerebras’ wafer-scale systems “to serve OpenAI customers”, rolling out in stages from 2026. CNBC reported a value of more than $10 billion.
• Amazon: about 2 gigawatts of Trainium (27 February 2026). Amazon is investing $50 billion in OpenAI, and OpenAI expanded its existing $38 billion AWS agreement “by $100 billion over 8 years”, committing to “approximately 2 gigawatts of Trainium capacity” across Amazon’s Trainium3 and Trainium4 chips. Trainium4 is expected from 2027.
• Google: reported, size unknown (27 and 28 October 2026). Not confirmed.

Add it up and the picture is clear. On paper, OpenAI has announced more non-Nvidia capacity (about 18.75 gigawatts across AMD, Broadcom, Cerebras and Amazon) than Nvidia capacity (10 gigawatts). Delivered reality is very different: most of the non-Nvidia capacity only arrives from late 2026 and 2027, and Nvidia chips still do most of today’s work across the industry.

So, is this the end of Nvidia’s lock?

No. But it may be the end of something smaller and still important: the idea that the biggest AI companies have no choice.

Nvidia’s real lock was never just its chips. It is the software around them: a programming platform called CUDA that much of today’s AI software is built on, plus years of tools and know-how. Moving a frontier model to another chip means rewriting and re-tuning a lot of that. That is exactly what OpenAI’s TPU compiler job describes. Google is attacking the same wall from its side: on the October call, Kurian said Google worked with Meta to deliver “Torch TPU”, so that PyTorch, one of the most widely used AI programming frameworks, runs on TPUs “using the tools they’re already familiar with”.

There is also a twist in Google’s own position. Google Cloud keeps selling Nvidia’s newest chips alongside its TPUs; Kurian listed “the latest NVIDIA GPUs” in the same breath as Google’s own systems. Google is Nvidia’s customer and its competitor at the same time. The market is not switching from one king to another. It is becoming a market with several suppliers, where the largest buyers spread their bets.

For OpenAI, the logic is simple. It needs more computing power than any single supplier can deliver, and every supplier it adds makes it less dependent on the others. Buying from Google, a direct rival in chatbots and search, shows how scarce computing power has become: you buy from your competitor when you cannot get enough elsewhere.

What this means for prices

• For the biggest buyers, competition already works as leverage. When OpenAI can credibly move part of its work to AMD, Broadcom, Amazon or Google chips, it can negotiate harder with all of them, including Nvidia. That is the main economic effect of a deal like this, even before a single TPU is delivered.
• The savings claims are vendor claims. Google says its newest inference TPU delivers 80% better price-performance than its previous TPU. Every chipmaker in this story makes similar claims about its own hardware. None of them has been independently tested on OpenAI’s real workloads.
• Cheaper chips do not automatically mean cheaper AI for you. If the cost of running AI falls, providers can cut prices, raise usage limits or simply keep the margin. All three have happened in this industry. Which one follows here is unknown.

What this means for companies

• Expect your AI provider to run on several chip types. That is mostly good for you: more capacity, fewer shortages. But behavior and speed can differ slightly between hardware. If quality matters, test your key tasks again after major provider updates.
• Keep your own code portable. If you train or run models yourself, prefer frameworks and serving tools that support more than one chip family. Google, for example, says its stack supports JAX, PyTorch, vLLM and SGLang. Lock-in costs more when prices are moving.
• More cloud choice arrives in 2027. Oracle will offer TPUs from 2027, and OpenAI’s models are coming to Google Cloud. If you are negotiating a multi-year cloud commitment now, ask what accelerators you will get, where, and at what price, and keep the right to change.
• Do not plan around a rumor. Until OpenAI or Google confirms the deal, treat it as a signal of direction, not as a fact for budgets or contracts.

What this means for users

• Nothing changes today. ChatGPT works the same tomorrow. Most of the chips in these deals arrive in 2027 or later.
• You will not see which chip answers you. The practical effects, if any, show up as speed, usage limits and prices, not as a label.
• More suppliers mean fewer bottlenecks. AI services have repeatedly rationed capacity with waitlists, usage limits and slow responses at peak times. A wider chip supply is one of the few things that can ease that.
• Be careful with the hype. Headlines like “breaking free from Nvidia” are ahead of the facts. Neither company has confirmed the purchase, and an unconfirmed report should not change what you think of either company’s products.

For the money side of this build-out, and why markets have started to price the AI bill, see The market just sent Big Tech its AI bill.

My reading

The headline story may turn out to be true, smaller than it sounds, or delayed. Two outlets with good track records report it, OpenAI is hiring TPU specialists, and Google is visibly building a chip-selling business. I would call it likely, not certain.

The bigger story is already confirmed. AI computing is turning from a one-supplier market into a several-supplier market, and the customers now big enough to shape it, OpenAI and Anthropic among them, are deliberately spreading their orders. That will not end Nvidia’s lead. It will make the lead more expensive to keep, and that pressure is good news for everyone who pays for AI, eventually.

Limits of this analysis

• The core claim rests on two paywalled reports, each citing one unnamed person. I could read The Information only up to the paywall and Bloomberg only through outlets citing it. If either company confirms or denies, the picture changes.
• Earnings-call quotes are taken from Alphabet’s official transcripts, which Alphabet itself says are provided for convenience; the webcast is authoritative.
• Gigawatt figures are announced plans with very different legal status (letter of intent, definitive agreement, commitment). They are not delivered capacity, and adding them up compares plans, not machines.
• Performance and price-performance figures are vendor claims.
• Anthropic’s numbers come from a registration statement that may still be amended before any listing.
• Nothing here is investment advice. I make no forecast for any company’s shares.

Sources

The Information, OpenAI, a Longtime Nvidia Customer, Plans to Buy Google’s AI Chips (27 October 2026, paywalled; reported).

Bloomberg, OpenAI Plans to Buy Google’s Chips as Part of Its AI Build-Out (28 October 2026, paywalled; read via citing outlets).

Reuters via The Star, OpenAI plans to buy Google’s AI chips, The Information reports (27 October 2026).

The Asia Business Daily, “Nvidia Chips Are Too Expensive,” OpenAI Also Turns to Google TPUs (28 October 2026).

OpenAI careers, Software Engineer, TPU Compilers (posted 20 October 2026) and Software Engineer, Accelerators.

TechRadar, OpenAI pledges to buy Google AI chips in its first move away from depending solely on Nvidia (1 November 2026).

Reuters via Yahoo Finance, OpenAI says it has no plan to use Google’s in-house chip (30 June 2025).

Alphabet, official earnings call transcripts: Q1 2026 (29 April 2026), Q2 2026 (22 July 2026), Q3 2026 (28 October 2026).

Oracle, Oracle Brings Google TPUs to Oracle Cloud Infrastructure (27 October 2026).

Anthropic, PBC, Form S-1 registration statement (filed 26 October 2026), sections on compute arrangements and chip platforms.

NVIDIA, OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems (22 September 2025).

AMD, AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs (6 October 2025).

CNBC, OpenAI chip deal with Cerebras adds to roster of Nvidia, AMD, Broadcom (16 January 2026), for the Broadcom timeline, the Cerebras value and Nvidia’s “no assurance” statement.

CNBC, OpenAI and Broadcom reveal Jalapeño, first AI chip in partnership (24 June 2026).

Cerebras, OpenAI Partners with Cerebras to Bring High-Speed Inference to the Mainstream (14 January 2026).

Amazon, OpenAI and Amazon Announce Strategic Partnership (27 February 2026).

Company statements are quoted as their claims. Media reports are marked as reported. Nothing here is investment advice.

Researched and drafted with AI assistance, including Anthropic’s Claude.