OpenAI and Anthropic between them will take about 26 million Google TPUs in 2028 and 2029, analyst Ming-Chi Kuo of TF International wrote on X on 16 November, citing his latest industry survey. He put OpenAI at about 12 million TPU 8t chips, designed by MediaTek, and Anthropic at about 14 million TPU 8i chips, designed by Broadcom; the 26 million total is our arithmetic.
Kuo also estimated total TPU shipments of 14 million to 15 million in 2028 and 16 million to 18 million in 2029. Set against those 30 million to 33 million units, the two labs’ demand comes to about 79% to 87%, by our arithmetic. All of these are Kuo’s estimates, not figures disclosed by Google, OpenAI or Anthropic.
Business Pill · Vertical integration
An evergreen Business Engineer explainer on vertical integration. It teaches the general idea only and does not discuss Google, OpenAI, Anthropic or TPUs.
The key insight: As we read it, Kuo’s estimates make two AI labs, not Google itself, the main buyers of Google’s eighth-generation TPUs in 2028 and 2029: about 26 million chips, or about 79% to 87% of his shipment estimate for those years, by our arithmetic. Anthropic’s filing points the same way in dollars, with additional TPU compute it says is expected to be delivered between 2028 and 2031.
What Kuo Wrote
Kuo’s post says that “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 that from 2028, Google’s own buildout moves to the next generation, v9, “so v8-series orders (8t / 8i or their design variants) can be viewed as coming from OpenAI and Anthropic.”
On how the two labs pay, Kuo wrote: “OpenAI buys outright, while Anthropic uses a mix of leasing and outright purchases.” He said demand from both grows year over year.
His shipment estimates run from 8 million to 9 million TPUs in 2027, a total of 38 million to 42 million over the three years, by our arithmetic. Kuo wrote that “The market broadly expects AI capex to peak in 2028, yet this plan shows TPU shipments still growing in 2029, driven by direct orders from AI service leaders OpenAI and Anthropic.”

What Was Posted on X
Ming-Chi Kuo, an analyst at TF International, posted his Google TPU update at 01:02 UTC on 16 November, describing it as key takeaways from his latest industry survey. It is the source of the OpenAI and Anthropic unit estimates, the 2027-2029 shipment ranges and the MediaTek and Broadcom supplier shares in this piece; all are his estimates, not company disclosures.
Jukan quoted the two paragraphs of Kuo’s post with the 2028-2029 demand figures for OpenAI and Anthropic and the 2027-2029 shipment estimates, adding only a two-word reaction. It repeats Kuo’s numbers and adds none of its own.
What Anthropic’s Filing Says
Anthropic’s amended S-1, filed with the SEC on 10 November, gives dollars rather than chip counts. It says Anthropic committed to spend at least approximately $121.3 billion under its Google hosting agreement between April 2026 and July 2033, as covered in our piece on the $121.3 billion floor.
The filing says Anthropic also committed to procure additional TPU compute through one of several routes, including a CSP hosting agreement with Google, “direct purchase from Google with five-year financing, or direct purchase and/or lease from a third-party TPU provider with five-year financing.” It says that compute “is expected to be delivered between 2028 and 2031” and estimates the payments at approximately $125.0 billion to $150.0 billion.
Separately, the filing lists five-year equipment leases with Broadcom whose rent comes to approximately $161.2 billion excluding interest. It says the leases give Anthropic “access to TPU-based compute capacity”. Kuo’s mix of leasing and purchases matches the delivery routes the filing describes.
In October 2025, Anthropic announced plans to use “up to one million TPUs”. Kuo’s 14 million for 2028 and 2029 is 14 times that figure, by our arithmetic, across different years and chip generations.

What Is on Record for OpenAI
The Information reported on 27 October, and Bloomberg the next day, that OpenAI plans to buy Google’s AI chips, as set out in our piece on OpenAI’s planned TPU purchase. Bloomberg reported that final terms of any deal were not yet known, according to a person familiar with the matter, and that OpenAI and Google declined to comment.
Kuo’s post puts a unit count on that plan: about 12 million TPU 8t chips over 2028 and 2029, bought outright, by his estimate. Anthropic’s 14 million is about 54% of the two labs’ combined 26 million, by our arithmetic.
MediaTek and Broadcom
Kuo wrote that MediaTek won the design service order for v10, now called Quillfish, with mass production expected in 2029. He wrote that MediaTek will be the sole supplier for v9 and v10, “accounting for about 45–50%, 50–55%, and 55–60% of total TPU supply in 2027–2029, respectively.”
On Broadcom, Kuo wrote that its focus will be non-TPU custom chips and AI networking, “with non-TPU revenue expected to surpass TPU revenue as early as 2027.” Broadcom’s AI revenue already equals 18.8% of Nvidia’s Data Center line, as shown in our piece on Broadcom’s AI revenue.
A J.P. Morgan Asset Management chart in Michael Cembalest’s note of 14 November gives Google TPUs 16% of frontier labs’ future compute agreements, as covered in our piece on the frontier-lab compute deals.
The Structural Read
As we read the post, the labs’ split matches their suppliers: Kuo assigns OpenAI’s 12 million to the MediaTek-designed 8t and Anthropic’s 14 million to the Broadcom-designed 8i, so each design house rides one lab’s orders.
As we read Anthropic’s S-1/A, its stated TPU-linked commitments already come to $282.5 billion, the $121.3 billion Google floor plus $161.2 billion of Broadcom lease rent, by our arithmetic, before the $125.0 billion to $150.0 billion it estimates for additional TPU compute.
Ming-Chi Kuo, TF International, on X, 16 November 2026
“OpenAI buys outright, while Anthropic uses a mix of leasing and outright purchases.”
Three Implications
FOR MEDIATEK Kuo wrote that it will supply about 45–50%, 50–55%, and 55–60% of total TPU supply in 2027 to 2029, as the sole supplier for v9 and v10.
FOR BROADCOM Kuo wrote that its non-TPU revenue is expected to surpass TPU revenue as early as 2027.
RELATED Anthropic’s Google commitment is in our piece on the $121.3 billion floor.
The Business Engineer Lens
This story maps onto the Business Engineer framework The AI Supercycle.
The framework splits the silicon layer into three categories, “generalist substrate (NVIDIA), hyperscaler custom (TPU, Trainium, MTIA), merchant custom challengers (Cerebras, Groq, SambaNova)”, and argues: “The within-layer split is what makes silicon strategically interesting again.”
As we read it, Kuo’s post adds a twist to the hyperscaler-custom category: by his estimates, Google’s v8 TPUs in 2028 and 2029 go mostly to two outside labs. The framework predates the post and does not discuss these orders.
What Is Not Established
The unit counts are Kuo’s estimates from his industry survey. Anthropic’s S-1/A states commitments in dollars, and in the filing text we read it gives no 2028 or 2029 TPU unit count. Kuo’s demand figures and his shipment figures are separate estimates, so the 79% to 87% share is a comparison of two estimates, not a measured share.
We did not contact Google, OpenAI, Anthropic, MediaTek or Broadcom for this piece.
The Bottom Line
Kuo puts OpenAI at about 12 million Google TPUs and Anthropic at about 14 million across 2028 and 2029, against his estimate of 30 million to 33 million TPU shipments in those two years, about 79% to 87%, by our arithmetic. Anthropic’s own filing says its additional TPU compute is expected to be delivered between 2028 and 2031.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece rests on Ming-Chi Kuo’s post on X of 16 November 2026, which we read in full, and which he describes as drawn from his latest industry survey and research conclusions. Every unit count, supplier share and shipment figure is Kuo’s estimate, not a disclosure by Google, OpenAI, Anthropic, MediaTek or Broadcom. We checked his claims about Anthropic against its amended S-1 filed with the SEC on 10 November and its announcement of 23 October 2025; the reports by The Information and Bloomberg on OpenAI’s planned chip purchase are as we covered them on 28 October. The totals, shares and ratios are our arithmetic. We did not contact the companies named.
Sources: Ming-Chi Kuo (TF International), post on X, 16 November 2026 · Jukan, quote-post on X, 16 November 2026 · Anthropic PBC, Form S-1/A Amendment No. 2 (SEC), 10 November 2026 · Anthropic, Expanding our use of Google Cloud TPUs and Services, 23 October 2025









