AI News

OpenAI and Anthropic Google TPU Orders Total 26 Million Chips

OpenAI and Anthropic Google TPU orders total about 26 million chips for 2028 and 2029, according to TF International analyst Ming-Chi Kuo. The combined total is derived from Kuo’s estimates of about 12 million units for OpenAI and about 14 million for Anthropic. Kuo reported the findings from an industry survey in a post on X on 16 November 2026 at 01:02 UTC.

This article reflects reporting available as of 16 November 2026.

Key Takeaways

  • Combined estimated demand is about 26 million Google TPUs in 2028 and 2029.
  • OpenAI is estimated at about 12 million units and Anthropic at about 14 million.
  • Shipments are estimated at 8–9 million in 2027, 14–15 million in 2028 and 16–18 million in 2029.
  • v10 is now Quillfish, due in 2029; Icefish is now v11, due in 2030.
  • All figures are Kuo’s estimates, not company disclosures.

The Short Answer

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). The Ming-Chi Kuo TPU survey puts OpenAI demand at about 12 million units and Anthropic demand at about 14 million across 2028 and 2029. The OpenAI TPU purchase is outright, while Anthropic combines leasing with outright purchases.

TPUs are custom chips for AI training and inference. An AI accelerator is specialised hardware that speeds up AI workloads.

How Large Are the Google TPU Orders?

The figure of about 26 million comes from adding Kuo’s two estimates: about 12 million for OpenAI and about 14 million for Anthropic. Kuo did not state a combined total; it is arithmetic derived from the estimates. Both buyers are expected to increase demand year over year.

OpenAI is expected to take TPU 8t chips, designed by MediaTek, while Anthropic is expected to take TPU 8i chips, designed by Broadcom. The Anthropic TPU lease arrangement mixes leasing with outright purchases, while OpenAI buys outright.

BuyerEstimated demand, 2028–2029Chip and designerPurchase model
OpenAIAbout 12 million unitsTPU 8t, designed by MediaTekOutright purchase
AnthropicAbout 14 million unitsTPU 8i, designed by BroadcomLeasing and outright purchases
Combined total (derived)About 26 million unitsTPU 8t and TPU 8iPurchases and leases

FourWeekMBA quotes the full X post from Kuo. TradingView also reports the estimates. ChainCatcher and KuCoin also covered the survey.

Why Do v8 Chips Remain Central Through 2029?

From 2028, Google’s v9 becomes the mainstay of Google’s own compute buildout, according to Kuo. As a result, v8-series orders, including the 8t and 8i chips or design variants, can be viewed as coming from OpenAI and Anthropic. The v8 series should remain the bulk of TPU shipments through 2029.

What Changed in the v10 and v11 Roadmap?

The v10 lineup changed from Icefish and Merope, also known as Bladerunner, to Quillfish, with mass production expected in 2029, and MediaTek won the design service order. Icefish is now v11, with mass production expected in 2030, and the design order also went to MediaTek. Merope has lower visibility and may be cancelled, according to Kuo.

Who Designs and Packages the New TPUs?

MediaTek will be the sole design supplier for v9 (Humufish and Triggerfish) and v10 (Quillfish). Kuo estimates MediaTek at about 45–50% of total TPU supply in 2027, about 50–55% in 2028 and about 55–60% in 2029.

  1. MediaTek designs v9 and v10; Broadcom designs the TPU 8i.
  2. From v9 onward, all new TPUs use Intel’s EMIB-T advanced packaging.
  3. The v8 series remains the bulk of shipments through 2029.

What Do the Shipment Estimates Show?

Kuo’s TPU shipment estimates are 8–9 million units in 2027, 14–15 million in 2028 and 16–18 million in 2029. Recent AI chips news covers the wider competition among AI chip suppliers, including Nvidia GPUs and other AI ASICs.

What do the figures imply for Google TPU shipments 2029?

The Google TPU shipments 2029 estimate of 16–18 million units is the largest annual figure in Kuo’s survey. Combined demand of about 26 million units across 2028 and 2029 represents about 79–87% of the combined 2028–2029 shipment estimates of 30–33 million units. That share is arithmetic derived from Kuo’s estimates, not a figure Kuo stated.

What Does Anthropic’s S-1 Filing Add?

Anthropic’s amended S-1, filed with the SEC on 10 November 2026, provides separate evidence for the scale of its Google commitment. Anthropic’s IPO S-1 filing sets out the Google hosting and equipment commitments. Anthropic committed to spend at least about $121.3 billion between April 2026 and July 2033. Additional TPU compute expected between 2028 and 2031 is estimated at about $125.0–150.0 billion.

The filing also records five-year equipment leases with Broadcom, with rent of about $161.2 billion excluding interest, giving access to TPU-based compute capacity. In October 2025, Anthropic announced plans to use up to one million TPUs. The Information reported on 27 October 2026, and Bloomberg the next day, that OpenAI plans to buy Google’s AI chips; terms were not known, and OpenAI and Google declined to comment.

How Does Nvidia Succession Uncertainty Fit In?

Kuo states that uncertainty over Nvidia’s CEO succession helps Google discuss post-2028 long-term plans with customers. Jensen Huang resigned as Nvidia CEO on 12 November 2026, and CFO Colette Kress is interim CEO. Jensen Huang’s resignation as Nvidia CEO is the background.

Key Terms at a Glance

  • TPU: A tensor processing unit, a custom Google chip for AI training and inference.
  • ASIC: An application-specific integrated circuit, a chip designed for one task.
  • EMIB-T: Intel’s embedded multi-die interconnect bridge technology for packaging.
  • Design service order: A contract for engineering that turns a chip architecture into a manufacturable design.
  • Shipment visibility: The period over which future shipments can be forecast from orders and demand.

Frequently Asked Questions

How many TPUs will OpenAI and Anthropic take in 2028 and 2029?

OpenAI is estimated to take about 12 million TPUs and Anthropic about 14 million, for a combined derived total of about 26 million. The OpenAI TPU purchase is outright, while the Anthropic TPU lease model mixes leasing and outright purchases.

Are the 26 million units a figure stated by Kuo?

No, the 26 million total is arithmetic derived from Kuo’s estimates of about 12 million for OpenAI and about 14 million for Anthropic.

What did the Ming-Chi Kuo TPU survey estimate for 2027?

Shipments were estimated at 8–9 million units in 2027, rising to 14–15 million in 2028 and 16–18 million in 2029.

When will the v10 and v11 TPUs reach mass production?

Mass production of v10, now Quillfish, is expected in 2029, and of v11, now Icefish, in 2030. MediaTek won both design orders.

Related Reading on Future AI Wire

Sources