Two customers, most of the chips

Start with what the numbers are and what they are not. Kuo’s 12 million and 14 million are estimates from one analyst’s industry survey. They are not orders announced by OpenAI or Anthropic, not shipments reported by Google, and not figures taken from a filing. Anthropic is in the middle of a public listing process and its prospectus speaks in dollars and commitments; no 2028–29 TPU unit count from it was identified in the sources reviewed that readers could check these figures against.

Within those limits, the structure of the claim is specific enough to be tested later. OpenAI’s share is assigned to the 8t, a training-oriented variant designed by MediaTek, and described as outright purchases. Anthropic’s is assigned to the 8i, an inference-oriented variant designed by Broadcom, split between leasing and purchases. Both are said to grow year over year. Specific assignments like these either show up in later supply-chain reporting or they do not.

The share is a ratio of two estimates

The striking figure, that two outside labs would take roughly four-fifths or more of Google’s TPU output in 2028–29, is not something Kuo stated. It comes from dividing his demand estimate (26 million units for the two labs) by his shipment estimate (14–15 million in 2028 plus 16–18 million in 2029). Both sides of that division are his. The ratio rises if projected demand rises or projected shipments fall, and falls under the reverse changes. Readers should treat 79% to 87% as an illustration of what his numbers imply, not as a measured market share.

The underlying logic is at least stated openly. Kuo argues that from 2028 Google’s own expansion runs mainly on the v9 series, so v8-series volume can be attributed to external customers. That attribution is an inference, and it does a lot of work in the headline figures: Google retaining v8 capacity for itself, or serving other customers from it, would challenge Kuo’s attribution, and the effect on the ratio would depend on which of his estimates changed.

Evidence that would strengthen or challenge it

Three things would firm this up. A unit figure in an Anthropic filing or an OpenAI disclosure; a dollar commitment alone would not settle units without price, lease and service terms behind it. Corroboration from a second supply-chain source with its own survey, rather than coverage repeating Kuo’s post. And MediaTek or Broadcom commentary that matches the design split he describes, since both companies’ TPU roles are named precisely.

Until one of those appears, the durable part of the story is narrower than the headline: a well-followed supply-chain analyst believes Google’s TPU order book now runs to 2029, against mostly Q4 2027 to Q2 2028 visibility for rival accelerators in his estimate, and he attributes the extended visibility partly to OpenAI and Anthropic orders while describing Google’s own transition toward the v9 series. That is a claim about direction. The 26 million is a claim about scale, and it rests on one survey.

Sources

  1. [1] Ming-Chi Kuo, X post, 16 November 2026 (primary source for all unit estimates, the v8 attribution, the MediaTek design role and the shipment-visibility claim), as reproduced in full by ChainCatcher and corroborated by the Stocktwits relay below
  2. [2] ChainCatcher, "MingChi Kuo: Google TPU shipment visibility extended to 2029," November 16, 2026Read source
  3. [3] Stocktwits via TradingView, "More Pain For NVDA? Top Supply Chain Analyst Says New OpenAI, Anthropic Chip Orders Boost Google’s TPU Outlook," November 16, 2026Read source