By Hamza Faisal

OpenAI and Anthropic are set to become Google's biggest chip customers by a wide margin, according to one of the supply chain's most closely watched analysts — ordering roughly 26 million of Google's custom AI chips across 2028 and 2029.

TF International analyst Ming-Chi Kuo said in a post on X on Sunday that his latest industry survey shows fresh orders from the two AI labs have extended visibility into Google TPU shipments through 2029, ahead of Nvidia GPUs and other AI accelerators.

Kuo estimates OpenAI will take about 12 million of the chips — the 8t variant, designed by MediaTek — and buy them outright. Anthropic, he says, will take about 14 million of the 8i variant, designed by Broadcom, through a mix of leasing and outright purchases. Both labs' demand is expected to grow year on year.

Two labs, most of Google's output

Set against Kuo's estimate of total TPU shipments — 8 to 9 million units in 2027, 14 to 15 million in 2028 and 16 to 18 million in 2029 — the two labs would account for roughly 79 to 87 percent of all TPUs Google ships in 2028 and 2029 combined.

Kuo also notes that from 2028, the v9 generation will become the core of Google's in-house computing capacity, so orders for the current v8 series can effectively be read as demand from OpenAI and Anthropic. Because of initial capacity constraints, the v8 series is still expected to make up most shipments through 2029. The v10 line has also been revised to a new design, codenamed Quillfish, with mass production expected in 2029.

MediaTek and Broadcom are the big winners

MediaTek emerges as the quiet winner. Kuo says the Taiwanese chip designer will be the sole design supplier for both the v9 generation — codenamed Humufish and Triggerfish — and the v10 Quillfish, accounting for roughly 45 to 50 percent of total TPU supply in 2027, rising to 55 to 60 percent by 2029. From v9 onward, all new TPU models will use Intel's EMIB-T advanced packaging.

The numbers line up with what Anthropic has already disclosed in dollars: its IPO filing says additional TPU compute is expected to be delivered between 2028 and 2031. OpenAI was reported in late October to be planning to buy Google's AI chips, and it agreed a separate 10-gigawatt custom-accelerator partnership with Broadcom in October 2025.

Nvidia's grip is loosening

The shift is the latest sign that Nvidia's once-total grip on AI computing is loosening. A JPMorgan Asset Management analysis cited this weekend put Nvidia GPUs at 39 percent of the computing infrastructure frontier AI labs have contracted to secure going forward, down from the 83 percent share of total AI chip revenue Nvidia held from 2023 to 2025. Broadcom ranked second at 24 percent, Google's TPUs third at 16 percent.

The timing is awkward for Nvidia. Founder Jensen Huang resigned as chief executive last week, with CFO Colette Kress as interim chief. Nvidia's stock fell 12.8 percent last week, its worst weekly performance since March 2025. Kuo says the succession uncertainty is helping Google discuss post-2028 long-term plans with customers.

Two caveats matter. Kuo's figures are estimates drawn from his own industry survey, not numbers disclosed by Google, OpenAI or Anthropic, though the market still expects AI capital spending to peak in 2028. The next hard data point arrives soon: Nvidia reports third-quarter results on Tuesday after the US market closes, with revenue the number to watch.