TPU (Tensor Processing Unit)

A TPU (Tensor Processing Unit) is the custom AI accelerator chip Google designs for its own machine-learning workloads, co-designed with Broadcom and manufactured by TSMC. Unlike a general-purpose GPU, a TPU is an application-specific chip tuned for the matrix math behind neural networks, which makes it Google’s answer to running AI at data-center scale.

What is a TPU?

A TPU is an application-specific integrated circuit (ASIC): a chip built to do one job well rather than anything a program asks. That job is the dense matrix multiplication at the core of training and running neural networks. Google designs the chip, but it does not build it alone. Broadcom has co-designed Google’s TPUs since 2016, and the relationship now runs through 2031, covering the networking and packaging silicon around the accelerator as well (Capacity Media). The silicon is fabricated by TSMC; the seventh-generation Ironwood TPU is built on TSMC’s 3nm (N3P) process with CoWoS advanced packaging (Tom’s Hardware).

Each generation pushes scale. Google describes Ironwood, announced in 2025, as its seventh-generation TPU and the first designed specifically for inference, scaling up to 9,216 liquid-cooled chips in a single pod (Google).

How fast is a TPU?

The numbers explain why Google keeps building its own silicon instead of just buying GPUs. Each Ironwood chip hits a peak of 4,614 teraflops, and Google says perf-per-watt is double that of Trillium, the prior generation (Google). Wired together, the scale is the real story:

“When scaled to 9,216 chips per pod for a total of 42.5 Exaflops, Ironwood supports more than 24x the compute power of the world’s largest supercomputer.” — Google, Ironwood: The first Google TPU for the age of inference

Performance per watt is the metric that matters at this scale, because a data center is power-limited long before it is space-limited. A chip narrowed to one job, neural-network matrix math, can spend its transistor budget on that job instead of the general-purpose flexibility a GPU carries, and at a fleet of thousands of accelerators those efficiency gains turn directly into lower cost per token served. That is the economic case for an ASIC, and it is why Google’s tight loop of in-house chip design plus its own models plus its own cloud is hard for a rival to copy.

Why does the TPU matter for investors?

The TPU is the anchor of the Google (Alphabet) AI value chain: it is what makes Google’s supply chain distinct from a generic semiconductor basket, and it routes spending to specific suppliers. Broadcom’s AI semiconductor revenue, much of it custom accelerators like the TPU, rose 74% year over year to $6.5 billion in fiscal Q4 2025 (The Motley Fool). Demand is also widening beyond Google: in October 2025 Anthropic agreed to expand its use of Google Cloud TPUs to up to one million chips and more than a gigawatt of capacity coming online in 2026 (Data Center Dynamics).

Related terms & concepts

FAQ

Who designs and makes Google's TPU?

Google designs the TPU but co-designs it with Broadcom, a partnership running since 2016 and extended through 2031 (Capacity Media). The physical chip is fabricated by TSMC; the seventh-generation Ironwood TPU uses TSMC's 3nm (N3P) process with CoWoS packaging (Tom's Hardware).

How is a TPU different from a GPU?

A GPU is a general-purpose parallel processor; a TPU is an application-specific chip (ASIC) tuned narrowly for the matrix multiplication that neural networks rely on. That focus can deliver better performance per watt for AI workloads, which is why hyperscalers like Google build their own. Google says Ironwood, its seventh-generation TPU, was the first designed specifically for inference (Google).

Why do TPUs matter to investors?

TPUs are the anchor of Google's AI supply chain and a major revenue driver for its suppliers. Broadcom's AI semiconductor revenue, much of it custom accelerators like the TPU, rose 74% year over year to $6.5 billion in fiscal Q4 2025 (The Motley Fool). Demand is also expanding beyond Google: Anthropic agreed to use up to one million TPUs by 2026 (Data Center Dynamics).

Sources & references

  1. Ironwood: The first Google TPU for the age of inference · Google, 2025-04-09
  2. Broadcom locks in long-term Google TPU deal through 2031 · Capacity Media, 2025-10-24
  3. The custom AI ASIC state of play (May 2026) · Broadcom deals, Google TPUs, Meta MTIA & beyond · Tom's Hardware, 2026-05-01
  4. Broadcom (AVGO) Q4 2025 Earnings Call Transcript · The Motley Fool, 2025-12-12
  5. Google and Anthropic confirm massive 1GW+ cloud deal with up to one million Google TPUs · Data Center Dynamics, 2025-10-23