The earlier ASIC gate noted that the big cloud companies buy NVIDIA but also build their own AI chips. The one with the longest track record is Google’s TPU.
This is the Google TPU deep-dive for Gate 1, “AI chips,” of The AI Hardware Supply Chain, End to End: what a TPU is, how it differs from a GPU, who uses it, and who helps Google make it.
What Is a TPU
TPU stands for Tensor Processing Unit. It is Google’s in-house chip built for machine learning, a kind of ASIC (a chip built for one specific task).
A GPU is a more general-purpose parallel chip. It can run almost anything, and its software ecosystem (CUDA) is mature. A TPU is designed together with its matrix units, HBM memory, chip-to-chip links, and compiler. It focuses on AI training and inference, with the goal of being cheaper on cost, power, and scale. Think of a GPU as a Swiss army knife that can cut anything, and a TPU as a chef’s knife Google made for its own few dishes.
Google still uses NVIDIA, though. Google Cloud offers both TPUs and NVIDIA GPUs as two product lines.
A Brief History: From the First Generation to Ironwood
| Generation | Key points |
|---|---|
| First gen | Used inside Google’s data centers from 2015, running search ranking, Street View, and AlphaGo |
| Second to fourth gen | Moved step by step from inference to large-scale training |
| Fifth and sixth gen | v5e (lower cost), v5p (higher performance), then Trillium (v6e) |
| Seventh gen Ironwood (TPU7x) | Generally available; 192GB of HBM per chip, 9,216 chips in a single pod |
| Eighth gen TPU 8t / 8i | Announced April 2026; 8t for training, 8i for inference; as of October 2026 the product page still says Coming soon |
All specs are from Google. Think of the eighth gen as “announced, not yet launched,” not as something you can rent today.
TPU vs GPU: Will It Replace NVIDIA
The answer depends on the scale: inside Google, or the whole market.
Inside Google, Gemini’s needs can shape the chip design directly, so the TPU has a clear edge. Research firm TrendForce estimates that TPUs will be about 78% of Google’s own AI-server shipments in 2026, and that Google is the only major cloud provider whose ASIC shipments exceed its GPU shipments.
Worldwide, the picture is different. In April 2026, TrendForce estimated that systems using ASICs (including TPUs) are about 27% of AI servers, so GPUs are still the majority. NVIDIA’s moat is the CUDA ecosystem and the fact that it runs on every cloud, and outside developers face switching costs to move to TPUs. So for now it is coexistence and a division of labor, not replacement.
Who Uses TPUs
Google’s own Search and Gemini run on TPUs. Outside users have mostly rented them through Google Cloud. On the July 2026 earnings call, Alphabet’s CFO said Google delivered TPU systems directly into customers’ own data centers for the first time in the second quarter, and most revenue from these deals will be recognized in 2027.
The most watched outside customer is Anthropic. In October 2025 it announced expanded use of Google Cloud TPUs, up to a million chips and over 1GW of capacity in 2026. In April 2026 it signed with Google and Broadcom to get about 3.5GW of next-generation TPU compute through Broadcom starting in 2027.
But Anthropic doesn’t rely on TPUs alone. It also uses AWS Trainium and NVIDIA GPUs, and in August 2026 AMD announced that Anthropic will deploy up to 2GW of MI450. Using several vendors is normal for large AI companies.
Who Helps Google Build TPUs, and Taiwan’s Role
Google controls the TPU’s architecture and software. Partners handle detailed design, manufacturing, and packaging.
The clearest partner is Broadcom. In an April 2026 SEC filing, Broadcom confirmed it will do custom design and supply for future generations of Google’s TPU, and provide networking parts for AI racks, with the deal running through 2031.
On manufacturing, there are only media reports so far. According to TrendForce in August 2026, citing Commercial Times, the eighth-gen TPU uses TSMC’s N3P process and CoWoS-S packaging, and MediaTek takes part in the chip design. Google has not disclosed the process node and has not confirmed MediaTek’s role.
The “TPU stocks” often discussed in Taiwan’s market (design services, testing, packaging and test, substrates, cooling and power, and so on) mostly come from brokers’ supply-chain guesses, not from Google or Broadcom announcements. Being on a list does not mean a company has orders. This article only describes industry roles and is not investment advice.
Penchan’s summary
The TPU shows what happens when your own demand is big enough to justify your own chip. Google uses vertical integration to get better cost and energy efficiency, and now it is even putting TPUs into customers’ own data centers.
When you read about TPU gains, keep two lines clear: Google’s internal mix versus the global market, and official announcements versus supply-chain reports. The TPU is an important path besides NVIDIA, but it is not a replacement yet.
To see which kind of chip a TPU is, read What Is an ASIC. For the GPU side, read What Is a GPU and Blackwell. To see how these chips get their data, read HBM. To see the whole chain again, go back to the supply-chain overview.