OpenAI and Anthropic are expected to significantly increase their use of Google’s artificial intelligence chips, potentially extending demand for the company’s Tensor Processing Units (TPUs) through 2029. According to analyst Ming-Chi Kuo of TF International Securities, the two AI companies could account for approximately 26 million Google TPU units across 2028 and 2029.
The forecast highlights Google’s growing position in the AI infrastructure market, where demand for specialised processors is increasing as companies build and operate increasingly powerful AI models. Kuo estimates that OpenAI will account for around 12 million TPU units, while Anthropic could require approximately 14 million. These figures are analyst estimates rather than confirmed purchase disclosures from the companies.
OpenAI and Anthropic Could Order 26 Million Google TPUs
Kuo’s latest industry assessment estimates that OpenAI and Anthropic will become major customers for Google’s eighth-generation TPU processors.
The projected demand is divided between two different chip variants. OpenAI is expected to use Google’s TPU 8t, designed with MediaTek, while Anthropic is associated with the TPU 8i, developed with Broadcom.
| Company | Expected TPU variant | Estimated demand in 2028–2029 | Design partner |
|---|---|---|---|
| OpenAI | TPU 8t | 12 million units | MediaTek |
| Anthropic | TPU 8i | 14 million units | Broadcom |
| Combined | TPU 8 series | 26 million units | — |
According to Kuo, OpenAI is expected to purchase its chips outright. Anthropic, meanwhile, is likely to combine direct purchases with equipment leasing arrangements.
Both companies’ demand is expected to increase year over year, potentially giving Google greater visibility into its AI chip shipments beyond the next few years.
The estimates do not represent publicly confirmed orders. They reflect Kuo’s assessment of the supply chain and expected customer demand.
Google TPU Shipments Could Continue Growing Through 2029
Kuo expects Google’s total TPU shipments to increase over the next three years, even if broader spending on AI infrastructure begins to slow.
His estimates put total shipments at approximately 8 million to 9 million units in 2027, 14 million to 15 million in 2028, and 16 million to 18 million in 2029.
| Year | Estimated Google TPU shipments |
|---|---|
| 2027 | 8 million–9 million |
| 2028 | 14 million–15 million |
| 2029 | 16 million–18 million |
The forecast suggests that shipments could continue expanding in 2029, despite expectations that the wider AI capital expenditure cycle may peak in 2028.
OpenAI and Anthropic’s projected orders are an important part of this outlook. Their demand could help Google maintain a strong shipment pipeline while other AI infrastructure customers reassess their spending.
However, the shipment figures are forecasts, not confirmed production or delivery totals.
Why OpenAI Is Turning to Google’s AI Chips
OpenAI has historically relied heavily on Nvidia graphics processing units (GPUs) to train and operate its AI models. However, the company is increasingly exploring alternative computing platforms as the cost and scale of AI infrastructure continue to rise.
Google’s TPUs are application-specific processors developed to handle demanding machine-learning workloads. Unlike general-purpose GPUs, they are designed specifically for AI-related calculations and can be integrated into Google’s wider data-centre infrastructure.
Reuters reported in October 2026 that OpenAI planned to purchase Google’s AI chips, citing reporting by The Information. The potential arrangement would deepen the commercial relationship between two companies that compete in AI products and search.
Using more than one chip supplier could also help OpenAI diversify its computing infrastructure, reduce dependence on a single hardware ecosystem and match different workloads with suitable processors.
The expected 12 million TPU 8t units in Kuo’s forecast would represent a substantial expansion of this strategy if the estimate materialises.
Anthropic Expands Its Partnership With Google and Broadcom
Anthropic has already established a significant relationship with Google and Broadcom to secure computing capacity for its Claude AI models.
In April 2026, Anthropic announced an expanded agreement with the two companies for multiple gigawatts of next-generation TPU-based computing capacity, expected to begin coming online in 2027.
The arrangement is intended to support the growing demand for Claude and Anthropic’s expanding customer base.
Anthropic also uses hardware from multiple providers, including Google’s TPUs, Nvidia GPUs and Amazon’s Trainium processors. This diversified approach allows the company to distribute workloads across different computing platforms.
Kuo’s estimate of 14 million TPU 8i units in 2028 and 2029 suggests that Google’s role in Anthropic’s infrastructure could become even more significant.
The projected demand also reflects the importance of Broadcom, which is involved in designing Google’s custom AI processors and supplying related infrastructure components.
MediaTek and Broadcom Could Benefit From Google’s TPU Expansion
The projected orders could create opportunities for companies involved in designing and supplying Google’s custom AI chips.
MediaTek is expected to play a major role in Google’s upcoming TPU generations. Kuo identifies the company as the design partner for the TPU 8t variant and Google’s v9 and v10 platforms.
He estimates MediaTek’s share of total TPU supply could rise from approximately 45%–50% in 2027 to 50%–55% in 2028 and 55%–60% in 2029.
Broadcom is associated with the TPU 8i design expected to serve Anthropic’s requirements. The company also has a broader role in Google’s custom silicon and AI networking infrastructure.
Kuo expects Broadcom’s non-TPU businesses, including custom AI accelerators and networking chips, to become increasingly important to its revenue.
The potential growth in TPU shipments could therefore benefit a wider semiconductor supply chain rather than Google alone.
Google’s Next-Generation TPU Roadmap
Google’s TPU development roadmap is another important part of the forecast.
Kuo expects Google’s v9 platform to become the primary system for its internal computing requirements from 2028. This could leave more of the available v8-series production capacity for external customers, including OpenAI and Anthropic.
The analyst also expects changes to Google’s future TPU product lineup. The v10 platform, identified as Quillfish, is expected to enter mass production in 2029, with MediaTek providing design services.
Another platform, Icefish, has reportedly moved to the v11 generation, with mass production expected in 2030. The Merope project has less visibility, and its future remains uncertain.
Supply constraints could also influence the transition between generations. Kuo expects the newer platforms to use Intel’s EMIB-T advanced packaging technology, but initial production limitations could keep v8-series processors central to shipments through 2029.
These are supply-chain projections and have not been independently confirmed by Google.
The Bigger Picture
The projected orders underline a broader change in the AI semiconductor market. Companies developing large AI models are increasingly looking beyond traditional GPU suppliers to secure enough computing capacity for training and inference.
Google’s TPUs give the company an opportunity to compete in this market through both its cloud services and its custom-designed hardware. Selling or providing access to these processors also strengthens Google Cloud’s position as an infrastructure provider for AI companies.
The development is particularly significant because OpenAI and Anthropic compete directly with Google’s own AI products. Their potential use of Google-designed chips demonstrates how commercial relationships in the AI infrastructure market can cross competitive boundaries.
At the same time, the forecasts highlight the importance of MediaTek, Broadcom and advanced semiconductor packaging suppliers. The growth of custom AI processors could create additional opportunities for companies beyond the traditional GPU market.
However, large infrastructure commitments also carry financial risks. AI companies must generate sufficient revenue and maintain customer demand to justify the cost of purchasing or leasing millions of processors and securing the electricity and data-centre capacity needed to operate them.
Looking Ahead
OpenAI and Anthropic’s projected demand could make Google a more important supplier in the global AI computing market through the end of the decade. If Kuo’s estimates prove accurate, the two companies could account for a substantial portion of Google’s TPU shipments in 2028 and 2029. MediaTek and Broadcom could also benefit from the growing demand for custom AI processors.
The key question will be whether these projected orders translate into actual purchases, production and deployment. Google has not publicly confirmed the estimated 26 million-unit demand, and the final figures will depend on supply availability, AI infrastructure spending and the commercial growth of OpenAI and Anthropic. For now, the forecast points to continued demand for specialised AI chips even if the wider infrastructure investment cycle begins to mature.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.



