OpenAI and Anthropic will take about 26 million Google TPUs between them in 2028 and 2029, according to a new survey from TF International analyst Ming-Chi Kuo. That’s roughly 80 percent of all the TPUs Google is expected to ship in those two years, and it makes a rival one of their biggest hardware suppliers.
Kuo posted the findings on X Sunday night. His split: OpenAI takes about 12 million TPU 8t chips, designed with MediaTek and bought outright, while Anthropic takes about 14 million TPU 8i chips, designed with Broadcom, through leasing and direct purchases.
No company has confirmed a single number in the report.
Kuo’s figures are survey estimates, not disclosures, and no delivery schedule, price, or contract term has been published. Still, the direction of travel is hard to miss: the two labs that define the frontier of AI are now planning their compute around Google’s silicon.
What Kuo’s survey says about Google TPU orders
Kuo’s headline claim is visibility. “Strong new orders from OpenAI and Anthropic have extended Google TPU shipment visibility to 2029, ahead of Nvidia GPUs and other AI ASICs (mostly 4Q27–2Q28),” he wrote. Google can now see its chip orders stretching two years further out than its rivals can.
His shipment estimates, held back by supply chain limits, run 8 to 9 million TPUs in 2027, 14 to 15 million in 2028, and 16 to 18 million in 2029. Against that, OpenAI and Anthropic alone would absorb close to four-fifths of Google’s TPU output in 2028 and 2029.
There’s a neat logic to how Kuo assigns the chips. From 2028, Google’s newer v9 TPU becomes its main in-house compute platform, so v8-series orders can effectively be read as demand from OpenAI and Anthropic. Kuo also says the v10 line has been reorganized around a chip codenamed Quillfish, due for mass production in 2029, with MediaTek winning the design work.
Why OpenAI and Anthropic want Google’s chips
Both labs need far more compute than any one supplier can promise, and neither wants to depend on a single vendor. The Information reported in late October that OpenAI planned to buy Google TPUs, its first hardware purchase from outside Nvidia. Bloomberg reported the next day that final terms were not yet known and both companies declined to comment.
OpenAI has also committed roughly $600 billion to Nvidia servers through 2030, according to Nvidia’s own statements, so the TPU orders don’t replace Nvidia. They hedge it. Anthropic announced in October 2025 that it would expand its Google Cloud TPU use to as many as one million chips, and its amended IPO filing from November 10 puts at least $121.3 billion behind its Google hosting agreement, according to FourWeekMBA’s reading of the filing.
What this means for Nvidia
Nobody should write Nvidia off. Its GPUs still run most frontier training, and OpenAI’s $600 billion commitment through 2030 is real money. But Kuo’s note lands at an awkward moment. Jensen Huang resigned as CEO last week, and the board named CFO Colette Kress interim chief.
Kuo says that uncertainty is already helping Google. “The uncertainty around Nvidia CEO succession is helping Google discuss ‘post-2028 long-term plans’ with customers,” he wrote. Customers planning compute for 2029 want to know who they’ll be dealing with.
MediaTek and Broadcom split the work
MediaTek is the subplot Kuo is most sure about. “MediaTek will be the biggest winner in TPUs over the next few years,” he wrote, saying the Taiwanese chip designer will exclusively win the v9 and v10 design orders. Its share of TPU supply is projected at 45 to 50 percent in 2027, rising to 55 to 60 percent by 2029.
Broadcom takes the other half of the work. It designed the 8i chip Anthropic is buying, and its version carries the higher profit margin, Kuo said, though he added that Broadcom is shifting its focus toward non-TPU businesses.
Why this matters
Step back and the shape of the AI industry looks different than it did a year ago. The old map was simple: Nvidia sold the picks and shovels, and everyone else dug. Now the two biggest diggers are buying shovels from a third miner, one who also happens to be digging in the same field. OpenAI and Anthropic no longer have to take whatever Nvidia offers, whenever it offers it.
Kuo’s numbers could be wrong. The v9 chip was already delayed once, with its first tape-out pushed to January 2027 and shipments possibly slipping to early 2028. But the orders behind the estimates are the point: even if the exact counts shift, OpenAI and Anthropic have decided that Google’s silicon belongs in their plans. Nvidia’s decade of total control over AI hardware is ending, not with a rival chip, but with its best customers shopping elsewhere.
Frequently asked questions
How many Google TPUs will OpenAI and Anthropic order?
About 26 million combined across 2028 and 2029, per Ming-Chi Kuo: roughly 12 million TPU 8t chips for OpenAI and 14 million TPU 8i chips for Anthropic. None of the three companies has confirmed the counts.
Why are OpenAI and Anthropic buying Google TPUs instead of Nvidia GPUs?
Both labs want a second source of compute so they are not fully dependent on Nvidia, and Google now sells TPUs to outside customers instead of renting them only through Google Cloud.
Has Google confirmed the 26 million TPU orders?
No. The figures are estimates from Ming-Chi Kuo’s industry survey, published November 16, not company disclosures. Google, OpenAI, and Anthropic have not confirmed the numbers, the split, or any contract terms.
What does this mean for Nvidia?
Nvidia still dominates AI training, and OpenAI has committed hundreds of billions to its servers through 2030. But the TPU orders show its two biggest customers are now building a real alternative, and Kuo says the uncertainty around Nvidia’s CEO succession is helping Google pitch post-2028 plans.
Sources: Stocktwits/TradingView, FourWeekMBA, KOCPC, Coinotag, The Information, Bloomberg
