W.137Kuo estimates OpenAI and Anthropic need about 26 million Google TPUs in 2028–2029
Analyst Ming-Chi Kuo said on November 16 that new orders from OpenAI and Anthropic extend shipment visibility for Google’s TPUs to 2029, according to press reports of his post. He estimates demand of about 12 million chips from OpenAI and about 14 million from Anthropic over 2028–2029, and total TPU shipments of 8–9 million in 2027, 14–15 million in 2028 and 16–18 million in 2029. On his own figures the two labs’ 26 million is 79% to 87% of the two years’ shipments. None of the companies has confirmed the figures. MediaTek, which Kuo calls the biggest winner, closed 7.89% lower in Taipei.
What Kuo said
The claim comes from one analyst’s supply-chain survey. Ming-Chi Kuo of TF International Securities posted on X at 01:02 UTC on November 16. We could not open the post. What follows is taken from press reports of it, chiefly a near-complete Chinese rendering by Kocpc and reports by ETtoday, cnyes, TokenPost and Stocktwits, which agree on the figures each carries; cnyes gives only the lab figures. The opening line, shown in an embed, says new OpenAI and Anthropic orders extend shipment visibility to 2029.
He says orders exist and gives his own estimates of their size. According to those reports, OpenAI’s demand is about 12 million units of a chip called 8t, designed with MediaTek, which it would buy outright. Anthropic’s is about 14 million units of 8i, designed with Broadcom, through a mix of leasing and purchase. Both figures cover 2028 and 2029 together, with no split between the years. Stocktwits presents them as estimates; GuruFocus uses both “orders” and “projected to require”. The units are chips.
He estimates total shipments for three years. The reports give 8–9 million TPUs in 2027, 14–15 million in 2028 and 16–18 million in 2029, which Kuo describes as more reasonable estimates given supply-chain constraints. He gave no 2026 figure. Kocpc and ChainCatcher report him saying that from 2028 the v9 generation is the main platform for Google’s own computing, and that v8-series chips remain the bulk of shipments through 2029. He attributes the v8-series orders to the two labs on that reasoning.
He sets this against shorter visibility elsewhere. Stocktwits quotes him placing TPU visibility “ahead of Nvidia GPUs and other AI ASICs (mostly 4Q27–2Q28)”. GuruFocus paraphrases that as late 2027 or early 2028. ETtoday reports him saying uncertainty over who will lead Nvidia may help Google discuss plans beyond 2028 with customers.
No company has confirmed any of it. We found no statement from Google, OpenAI, Anthropic, MediaTek, Broadcom or Intel on the figures.
Kuo’s 2029 estimate is 5.6 to 6.3 times the 2025 shipments in one broker’s model
Google TPU shipments, millions of chips a year. Faint grey: a GF Securities model for 2024, 2025 and 2026, as an aggregator reported it on August 25, 2026; the 2026 figure was an estimate when made. Pink: Ming-Chi Kuo’s estimates of November 16, 2026 for 2027 to 2029; the solid part is the low end of each range and the dashed part runs to the high end. The two sets come from different analysts with different methods and are joined here only for scale. Kuo gave no figure for 2026 or earlier.
Kuo’s estimates by year, and what they imply for MediaTek
| Year | TPU shipments, millions | Change on the prior year | MediaTek share, Kuo | Implied MediaTek units, millions |
|---|---|---|---|---|
| 2027 | 8–9 | no 2026 figure | 45–50% | 3.6–4.5 |
| 2028 | 14–15 | +56% to +88% | 50–55% | 7.0–8.25 |
| 2029 | 16–18 | +7% to +29% | 55–60% | 8.8–10.8 |
| 2027–2029 | 38–42 |
What the figures imply
On Kuo’s own numbers the two labs would take most of two years’ TPUs. About 12 million and about 14 million make 26 million. His shipment ranges for 2028 and 2029 add to 30–33 million. The labs’ demand is 79% to 87% of that; Kocpc’s write-up puts it at about 80%. That leaves 4 to 7 million chips over two years for Google’s own use and every other customer. The two figures are not the same kind: the 26 million is demand, and the shipment ranges are what Kuo thinks the supply chain can deliver. If shipments fall short of demand, the labs would receive less than 26 million in those two years.
Growth slows in 2029 in his ranges. From 2027 to 2028 shipments rise between 56% and 88%. From 2028 to 2029 they rise between 7% and 29%. The reports have him noting that the market expects AI capital spending to peak in 2028 while TPU shipments still grow in 2029. Kocpc reports that the v9 and later chips use Intel’s EMIB-T packaging, whose initial capacity is limited.
Kuo’s figure for Anthropic alone is 14 times the largest chip count Anthropic has announced. In October 2025 Anthropic said it would have access to up to one million TPUs, with well over a gigawatt of capacity online in 2026. Kuo’s about 14 million is a two-year demand estimate, and Anthropic’s figure was a ceiling tied to 2026. In April 2026 it announced an agreement with Google and Broadcom for “multiple gigawatts of next-generation TPU capacity” from 2027, with no chip count or dollar value. Broadcom said in December 2025 that Anthropic was the customer behind orders of $10 billion and $11 billion. OpenAI has announced no TPU purchase. The Information reported on October 27 that it was in talks to buy TPUs, and OpenAI declined to comment. Tech Times says Reuters reported in June 2025 that OpenAI had begun renting Google’s chips.
The two labs’ 26 million is 79% to 87% of the TPUs Kuo expects to ship in 2028 and 2029
Millions of TPU chips over 2028 and 2029 combined, all Kuo’s estimates as reported. Strong grey: OpenAI’s demand for the 8t chip. Grey: Anthropic’s demand for the 8i chip. Third row: the two stacked. Pink: total TPU shipments for the two years, solid to the sum of the low ends (14 + 16) and outlined to the sum of the high ends (15 + 18). Faint grey: the total minus 26 million, our subtraction. The labs’ rows are demand and the total is a supply-constrained shipment estimate, so the last row is what the figures leave over, not a forecast Kuo made.
Against earlier estimates
Kuo’s 2027 range is in the upper half of the estimates we found. Morgan Stanley put 2027 at 5 million in July; an August report says that was raised from 3 million. A Commercial Times supply-chain report relayed by TrendForce in January gave production volume of 5 million in 2027 and 7 million in 2028. GF Securities’ model had 8.84 million and UBS 9.87 million, both reported in August. Kuo’s 8–9 million is 1.6 to 1.8 times Morgan Stanley’s figure and brackets GF’s. UBS’s is 10% to 23% above his range, and a Wccftech article of June 23 gave an unattributed 10–11 million. We read the brokers’ numbers in aggregators and not in the notes themselves.
For 2028 he is at about twice the January supply-chain figure. His 14–15 million is 2.0 to 2.1 times the Commercial Times report’s 7 million. Fubon Research said in a note of July 30 that Google planned to have 12–15 million v9 TPUs in 2028, according to Xenospectrum; that is a supply plan and not a shipment forecast. We found no earlier estimate for 2029.
One detail conflicts with an earlier account. Every report of the November 16 post pairs 8t with MediaTek and 8i with Broadcom. A Wccftech article of June 23, built around a Kuo post of June 22, gave the reverse in its own words. We could not resolve it, and it does not change the unit figures.
Kuo’s 2027 estimate is 1.6 to 1.8 times Morgan Stanley’s, brackets GF’s and is below UBS’s
Estimates of Google TPU shipments in 2027, millions of chips, with the date each was reported. Faint grey: estimates published before November 16. Pink: Kuo’s range, solid to 8 million and outlined to 9 million. The Morgan Stanley, GF Securities and UBS figures are as aggregators reported them; the Commercial Times figure is production volume, as TrendForce relayed it.
The suppliers, and the shares
Kuo names MediaTek as the main beneficiary. The reports have him saying MediaTek is the sole design supplier for the v9 chips and for v10, now a chip called Quillfish that replaces earlier plans and is due for mass production in 2029, and that its share of TPU supply rises from 45–50% in 2027 to 55–60% in 2029. Applied to his shipment ranges that is 3.6–4.5 million chips in 2027 and 8.8–10.8 million in 2029, our multiplication. He gave no share for Broadcom.
MediaTek fell more than the Taiwan market on the day. ETtoday reported the stock down NT$130, or 3.16%, in the first five minutes, and said the fall narrowed after Kuo’s report. It closed at NT$3,795, down NT$325 or 7.89%, Storm Media reported. The TAIEX closed down 6.03% and TSMC down 5.74%. MediaTek’s fall was 1.9 points larger than the index’s. W.134 covers the session.
US chip stocks were lower before the open. Reuters reported Nasdaq 100 futures down 0.97% at about 10:08 UTC, Nvidia down about 2.2% and Broadcom among chip stocks down between 3.5% and 5%. Nvidia reports results on November 17.
What this changes on inthesingularity.ai
- Will AI compute run short or long through 2030? Lead: an analyst series for TPU shipments to 2029 (8–9M in 2027, 14–15M in 2028, 16–18M in 2029). It is one analyst’s estimate, unconfirmed, and would need a watts-per-chip assumption to convert to gigawatts.
- How much compute does AI demand need by 2027? Lead: about 12 million TPUs for OpenAI and about 14 million for Anthropic over 2028–2029, the first unit figures we have seen for either lab beyond 2027. OpenAI has announced no TPU purchase.
- Lead for a tracker: estimates of TPU shipments by year and by source. For 2027 they run from 5 million (Morgan Stanley, July) to 9.87 million (UBS, August) among named brokers. The brokers’ notes should replace the aggregator figures before anything is built on them.
- How long can Nvidia run without Jensen Huang? Lead: Kuo puts visibility for Nvidia GPUs and other AI ASICs mostly at the fourth quarter of 2027 to the second quarter of 2028, and says the succession may help Google with customers planning beyond 2028. Nvidia reports results on November 17.
Sources and method
| Item | Figure | Source |
|---|---|---|
| Kuo’s post | Posted November 16, 2026, 01:02 UTC. We could not open it; its content here is from the reports below | Ming-Chi Kuo on X |
| Kuo’s post | TPU shipment visibility to 2029, ahead of Nvidia GPUs and other AI ASICs, which are mostly at the fourth quarter of 2027 to the second quarter of 2028; lab unit figures framed as estimates | Stocktwits via Yahoo Finance, Nov 16 |
| Kuo’s post | OpenAI about 12 million 8t (MediaTek) bought outright and Anthropic about 14 million 8i (Broadcom) leased and bought, over 2028–2029; total shipments 8–9 million in 2027, 14–15 million in 2028, 16–18 million in 2029; MediaTek share 45–50% (2027), 50–55% (2028), 55–60% (2029); v10 lineup replaced by Quillfish, 2029, MediaTek design service; v9 onward uses Intel EMIB-T packaging | Kocpc, Nov 16 (Chinese rendering of the post) |
| Kuo’s post | Same unit figures and shipment ranges; MediaTek called the biggest winner; Nvidia’s succession may help Google discuss plans beyond 2028 with customers | ETtoday, Nov 16 |
| Kuo’s post | Same lab unit figures; no shipment ranges | cnyes, Nov 16 |
| Kuo’s post | Same unit figures and shipment ranges; earliest report with an unambiguous timestamp (02:02 UTC) | TokenPost, Nov 16 |
| Kuo’s post | v8-series chips remain the bulk of shipments through 2029; from 2028 the v9 is the main platform for Google’s own computing | ChainCatcher, Nov 16 |
| Kuo’s post | Gives Nvidia’s visibility as late 2027 or early 2028; mixes “orders” and “projected to require” | GuruFocus, Nov 16 |
| Kuo’s post | Truncated embed of the post’s opening line | Semiconductorsx on Substack, Nov 16 |
| Earlier estimates | GF Securities model: 2.76 million (2024), 2.854 million (2025), 4.504 million (2026), 8.84 million (2027); UBS: 9.87 million in 2027; Morgan Stanley’s 2027 figure raised from 3 million | Unbiased Headlines, Aug 25 |
| Earlier estimates | Morgan Stanley: 3.2 million in 2026 and 5 million in 2027 | KuCoin news flash, Jul 15 |
| Earlier estimates | Commercial Times supply-chain report: production volume expected to reach 5 million in 2027 and 7 million in 2028 | TrendForce, Jan 5 |
| Earlier estimates | An article built around Kuo’s June 22, 2026 post that gives, in its own words, the reverse designer mapping: MediaTek for v8i, Broadcom for v8t; an unattributed 10–11 million TPUs for 2027 | Wccftech, Jun 23 |
| Earlier estimates | Fubon Research, note of July 30, 2026: Google plans to have 12–15 million v9 TPUs in 2028 | Xenospectrum, 2026 |
| Agreements | Anthropic: up to one million TPUs, well over a gigawatt online in 2026, worth tens of billions of dollars | Anthropic release, Oct 23, 2025 |
| Agreements | Anthropic: agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity from 2027; no dollar value | Anthropic release, Apr 6, 2026 |
| Agreements | Broadcom named Anthropic as the customer behind a $10 billion order and a further $11 billion order | Techstrong, Dec 2025 |
| Agreements | OpenAI in talks to buy Google TPUs, reported by The Information on October 27, 2026; scale undisclosed; OpenAI declined to comment; Reuters reported in June 2025 that OpenAI had begun renting Google’s chips | Tech Times, Oct 29 |
| Shares | MediaTek closed at NT$3,795, down NT$325; TAIEX closed at 41,040.08, down 2,635.04 points (6.03%); TSMC NT$2,215, down 5.74% | Storm Media, Nov 16 |
| Shares | MediaTek down NT$130 to NT$3,990 in the first five minutes; the fall narrowed after Kuo’s report | ETtoday, Nov 16 |
| Shares | US premarket at about 10:08 UTC: Nasdaq 100 futures down 0.97%; Nvidia down about 2.2%; Broadcom among chip stocks down between 3.5% and 5% | Reuters via Yahoo Finance, Nov 16 |
What we read, and what we did not. We could not open Kuo’s post on X. Its time, 01:02 UTC on November 16, is from the embed and from the post’s ID. The content is from press reports that agree with each other on every figure they share; where they differ in framing we say so. We did not read the Morgan Stanley, GF Securities or UBS notes; their figures are from aggregators. The labs against shipments. 12 + 14 = 26 million. 14 + 16 = 30 million and 15 + 18 = 33 million. 26 ÷ 33 = 78.8%; 26 ÷ 30 = 86.7%. 30 − 26 = 4 million; 33 − 26 = 7 million. 8 + 30 = 38 million and 9 + 33 = 42 million for 2027–2029. Growth. The low bound of each change divides the next year’s low end by the prior year’s high end, and the high bound the reverse: 14 ÷ 9 − 1 = 55.6%; 15 ÷ 8 − 1 = 87.5%; 16 ÷ 15 − 1 = 6.7%; 18 ÷ 14 − 1 = 28.6%. Comparisons. 8 ÷ 5 = 1.60; 9 ÷ 5 = 1.80. 16 ÷ 2.854 = 5.61; 18 ÷ 2.854 = 6.31. 14 million ÷ 1 million = 14; Anthropic’s figure was “up to” one million. 14 ÷ 7 = 2.00; 15 ÷ 7 = 2.14. 9.87 ÷ 9 − 1 = 9.7%; 9.87 ÷ 8 − 1 = 23.4%. $10 billion + $11 billion = $21 billion of Broadcom orders attributed to Anthropic. MediaTek. Implied units are the low share times the low shipment figure and the high share times the high: 8 × 45% = 3.60 and 9 × 50% = 4.50 million (2027); 14 × 50% = 7.00 and 15 × 55% = 8.25 million (2028); 16 × 55% = 8.80 and 18 × 60% = 10.80 million (2029). Previous close: 3,795 + 325 = NT$4,120. 325 ÷ 4,120 = 7.89%; 130 ÷ 4,120 = 3.16%. TAIEX: 2,635.04 ÷ 43,675.12 = 6.03%. 7.89% − 6.03% = 1.86 points. Conflicts between sources. Orders or estimates: Kuo’s opening line refers to new orders; Stocktwits frames the unit figures as what the labs may seek; GuruFocus uses both framings. We describe them as his estimates of demand. Nvidia’s visibility: mostly the fourth quarter of 2027 to the second quarter of 2028 in Stocktwits, ChainCatcher and Kocpc; late 2027 or early 2028 in GuruFocus. Designer mapping: 8t with MediaTek and 8i with Broadcom in all November 16 reports, the reverse in a June 23 Wccftech article. One Substack summary attributes the annual shipment ranges to the two labs and says “through 2030”; we treated that as an error. Some translated wires misspell Kuo’s name and firm. Not used. A Morgan Stanley rule of thumb relating TPUs to gigawatts and dollars, which we saw only in an aggregator and which predates these chip generations; estimates of Nvidia’s GPU unit shipments, which are not like for like; press reports that Anthropic’s April agreement is about 3.5 GW and that it will spend $200 billion with Google Cloud, neither of which the companies have stated; Alphabet’s capital spending guidance, which we could not check against its filing in time; and MediaTek’s own revenue targets, which do not name Google.
Wire reports are written within hours of the news, on a six-hour schedule. Each lists its sources and shows the arithmetic behind its figures. Not investment advice.