Kuo Estimates OpenAI and Anthropic Will Take 26 Million Google TPUs in 2028-29

 

Analyst Ming-Chi Kuo estimates that OpenAI and Anthropic will take about 26 million Google TPUs between them in 2028 and 2029 — roughly 12 million for OpenAI and 14 million for Anthropic. That is an estimate from an analyst's industry survey, posted on X on 16 November 2026, not a confirmed order announced by Google, OpenAI or Anthropic. No company has confirmed the counts, and no delivery schedule, price or contract term has been published.

The number matters anyway, because of what it implies if it is even roughly right. Kuo's figures would make two AI labs, not Google itself, the main buyers of Google's eighth-generation TPUs in those two years. It would also mean Google's custom-chip business has order visibility stretching to 2029, further than Kuo sees for Nvidia's GPUs and other AI accelerators.

What Kuo actually said

Kuo, an analyst at TF International Securities known for supply-chain reporting on Apple and Nvidia, posted what he called "key takeaways from my latest industry survey" at 01:02 UTC on 16 November 2026. His lead line was: "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)."

A TPU is Google's custom AI chip — a tensor processing unit designed for the maths that training and running large AI models requires — and "shipment visibility" means how far into the future orders can already be seen. Kuo's point is about the length of the runway: Google can already see TPU demand into 2029, while for Nvidia's GPUs and rival custom chips the visible order book mostly runs out between the fourth quarter of 2027 and the second quarter of 2028.

The numbers behind the headline are in the table below. Every figure in it is Kuo's estimate, and the totals and shares are arithmetic done on those estimates, not measurements.

ItemKuo's estimateDetail
OpenAI demand, 2028–2029About 12 million TPUsTPU 8t chips, designed by MediaTek; outright purchases
Anthropic demand, 2028–2029About 14 million TPUsTPU 8i chips, designed by Broadcom; leasing plus outright purchases
Combined demand, 2028–2029About 26 million TPUsSum of the two estimates; both growing year over year
Total TPU shipments, 20278–9 millionMidpoint about 8.5 million
Total TPU shipments, 202814–15 millionMidpoint about 14.5 million
Total TPU shipments, 202916–18 millionMidpoint about 17 million

Set the two labs' combined demand of about 26 million against Kuo's estimated total shipments of 30 to 33 million across 2028 and 2029, and the two labs come to roughly 79% to 87% of estimated output — a comparison of two estimates, not a measured share.

Close-up of a silicon AI chip on a circuit board in a laboratory, photorealistic

AI-generated illustration (generated 16 Nov 2026).

How Kuo reaches the numbers

Kuo's reasoning rests on a generational handover inside Google. From 2028, he says, the v9 series of TPUs becomes the core of Google's own in-house computing capacity. That means orders for the v8 series — the 8t and 8i variants and their design offshoots — can be read as demand from the two outside labs rather than from Google itself.

The shipment figures carry a second important qualifier: supply constraints. Kuo estimates total TPU shipments at 8 to 9 million in 2027, 14 to 15 million in 2028 and 16 to 18 million in 2029, and he presents those as reasonable figures under supply-chain constraints — what Google can actually deliver, not what customers would like to buy. If demand from the two labs alone is about 26 million chips across two years while total shipments are 30 to 33 million, there is very little slack for anyone else.

Kuo also looks beyond the v8 generation. He identifies MediaTek as the sole design provider for the v9 series and the v10 series, now named Quillfish, with mass production expected in 2029. He projects MediaTek's share of total TPU supply rising from about 45% to 50% in 2027 to 55% to 60% in 2029 — a projection that remains to be confirmed by actual shipment data.

What would support the estimate — and what would not

There is one public document that points in the same direction as Kuo's numbers, and one clear limit on what it proves. Anthropic filed for its IPO, and its filing commits enormous sums to compute. Our earlier reporting on the Anthropic IPO's $900.9 billion compute bill laid out the scale, and the filing text says additional TPU compute is expected to be delivered between 2028 and 2031. That is consistent with Anthropic taking millions of TPUs. But the filing states its commitments in dollars, and it gives no TPU unit count for 2028 or 2029. It supports the direction of Kuo's estimate. It does not confirm the 14 million figure, and anyone who cites the filing as proof of that number is overstating it.

That is the honest position for all three companies. Google has not disclosed TPU unit sales. OpenAI has not disclosed a TPU purchase. Anthropic has not disclosed a TPU count. Kuo's figures come from his industry survey, and supply-chain surveys can be wrong, early, or overtaken by events: orders can be cut, delayed, or renegotiated, and a customer can shift volume back to Nvidia.

The wider market context is worth stating plainly, because it frames the stakes without overstating them. The AI industry is spending staggering sums on compute, and the question of whether that spending is a bubble or a build-out is genuinely open — it is the question at the heart of our is the AI bubble bursting analysis. Kuo's note is evidence for the build-out side: if two labs really are lining up 26 million custom chips years out, they are planning as though demand keeps growing. But it is one analyst's survey, not signed contracts.

Wide shot of a data centre hall with rows of server racks and blinking lights, photorealistic

AI-generated illustration (generated 16 Nov 2026).

Why Nvidia is in this story at all

Kuo's note is being read as bad news for Nvidia, and the framing needs care. The comparison Kuo makes is about visibility, not sales. Nvidia's GPUs remain the default AI accelerator, and the company reports its own earnings on Tuesday 17 November — a date this article does not pre-empt and whose results it does not predict. What Kuo says is that Google's TPU order book is visible further ahead than Nvidia's GPU order book.

That is a real competitive signal, but it is not the same as Nvidia losing. Customers buy TPUs and GPUs for different reasons, and the chip market is growing fast enough that one company's longer visibility does not automatically mean another company's decline. The NVIDIA $3 billion Meta investment is a reminder of the same point from the other direction: Nvidia is also locking in demand through deep partnerships and equity, not just through quarterly chip sales.

Engineer in a cleanroom holding a silicon wafer, photorealistic

AI-generated illustration (generated 16 Nov 2026).

Key Takeaways

  • Ming-Chi Kuo estimates OpenAI will take about 12 million Google TPUs and Anthropic about 14 million across 2028 and 2029 — about 26 million combined. These are estimates from his industry survey, not confirmed orders.
  • OpenAI's share is the TPU 8t chip designed by MediaTek, bought outright; Anthropic's is the TPU 8i chip designed by Broadcom, leased and bought.
  • Kuo estimates total TPU shipments of 8–9 million in 2027, 14–15 million in 2028 and 16–18 million in 2029, so the two labs would be roughly 79% to 87% of estimated 2028–29 output — a comparison of two estimates, not a measured share.
  • His reasoning: from 2028 the v9 TPU becomes Google's core in-house platform, so v8-series orders can be read as demand from the two labs; supply constraints cap actual shipments.
  • Anthropic's IPO filing supports the direction — it commits dollars to TPU compute expected in 2028–2031 — but gives no unit count and does not confirm 14 million.
  • No company has confirmed the counts, the split, the payment mix, any delivery schedule, price or contract term; Kuo's visibility comparison puts TPU orders visible to 2029, ahead of Nvidia GPUs mostly visible to 4Q27–2Q28.

Conclusion

Kuo's estimate describes a remarkable shift if it is close to right: Google's custom chips becoming a product whose biggest customers are its two most important AI rivals, taking roughly four-fifths of output in 2028 and 2029. It fits with what Anthropic's IPO filing says about its compute appetite. But the operative word is "if": these are one analyst's survey estimates, unconfirmed by any of the companies named. Treat the 26 million as a signal worth watching, not as a deal that has been signed.

Sources

Comments