What Is a TPU? Google’s Tensor Processing Unit, Explained

EXPLAINED · Mechanism · Last reviewed: 07 Oct 2026 · Next review: after Alphabet Q3 2026 results (28 Oct 2026)

A TPU, or tensor processing unit, is a chip Google designs for one job: the large matrix math that trains and runs AI models. Google Cloud calls it a custom application-specific integrated circuit, or ASIC.

Why an investor cares: TPU money now shows up in several companies’ results, at different times. Broadcom books it when chips ship. Google books it when hours are used or systems are delivered. A lab like Anthropic pays over years. Knowing which step comes first tells you whose numbers move first. For the wider map of chip roles, see who makes AI chips.

What is a TPU?

A TPU is a chip built for matrix math. Training and running an AI model is mostly multiplying large grids of numbers. A TPU does that one job. It is an ASIC, an application-specific chip, rather than a chip built for every kind of work.

Google’s documentation says TPUs train models using hardware designed for “large matrix operations often found in machine learning algorithms.” They have high-bandwidth memory on the chip package, and they link together in groups called slices.

How does a TPU work?

Three parts do the work: matrix units, memory and links.

The matrix unit, or MXU, is the core. Google Cloud describes it as a 128 by 128 systolic array. That is a grid of small multipliers that pass results to their neighbours, so numbers flow through the grid instead of going back to memory each step.

Memory feeds the grid. Each Ironwood chip has 192 GiB of HBM, moving 7,380 GB/s. For how that memory is built, see HBM, explained. For how it sits next to the chip, see CoWoS, explained.

Links join the chips. Ironwood chips connect in a 3D torus, and a single pod scales to 9,216 chips. A full pod holds about 1.77 million GiB of HBM (9,216 × 192 GiB = 1,769,472 GiB, our arithmetic).

Source: Google Cloud documentation, TPU7x (Ironwood), last updated 2 Oct 2026.

Ironwood is the seventh generation. Against v5p, it has about 2.0 times the memory per chip, 2.7 times the memory bandwidth and 10 times the peak FP8 compute (192 ÷ 95; 7,380 ÷ 2,765; 4,614 ÷ 459, our arithmetic). Compute grew faster than memory. Google’s own docs warn that “HBM can still be a bottleneck” for memory-bound work.

TPU vs GPU: what is the difference?

A GPU is a general-purpose parallel chip. A TPU is Google’s own chip, built for matrix math. Both put HBM next to the compute. Google Cloud’s own guide sets out which work fits which.

Who makes Google’s TPU?

Google designs it. Broadcom builds it. Broadcom’s Form 8-K of 6 April 2026 says it has a Long Term Agreement to develop and supply custom TPUs for Google’s future generations. A second agreement covers networking and other parts for Google’s next AI racks through up to 2031.

On its 2 September 2026 call, Broadcom said it delivered Ironwood “in high volume to both Anthropic and Google” in the quarter. It also began production shipments of the next generation, TPU v8i, for Google, which it said has more memory and bandwidth than Ironwood.

The chip needs three scarce inputs. Broadcom named them on the same call: leading-edge wafers, substrates and HBM memory. Each one, it said, can become the bottleneck at different times.

Who gets paid when a TPU is built?

Broadcom is paid when the chip ships. Its AI semiconductor revenue was $16.7 billion in the quarter to 2 August 2026, up 221% on a year earlier. Custom accelerators, which it calls XPUs, were 73% of that, about $12.2 billion (0.73 × $16.7 billion, our arithmetic). That pool covers all six of its XPU customers, not only TPUs. It guides $21.7 billion of AI revenue for the next quarter.

Memory makers are paid for the HBM on every chip. Broadcom said rising memory content in its XPUs is pulling its consolidated gross margin down, to a guided 73% in Q4. Micron said more than 75% of its fiscal 2027 output is already committed, the Korea Herald reported on 1 October 2026. See who makes HBM and Micron’s 2027 commitments.

Google is paid two ways. It sells TPU hours on Google Cloud, which grew revenue 82% to $24.8 billion in Q2 2026. And it now sells whole TPU systems. Alphabet said it delivered TPU systems to customer data centres for the first time in Q2. It expects a “relatively small portion” of that revenue this year and the “vast majority” in 2027. Google also pays first: its capex was $44.9 billion in Q2, mostly on technical infrastructure, about 60% of that on servers. See what a hyperscaler is.

The lab pays to use them. Anthropic said in October 2025 it plans to use up to one million TPUs, worth tens of billions of dollars, with well over a gigawatt online in 2026. Its IPO prospectus shows a $125.2 billion five-year TPU lease commitment, Reuters reported on 1 October 2026. Broadcom agreed to lend up to $42 billion, about a third of it (42 ÷ 125.2 = 33.5%, our arithmetic). See Broadcom’s chip financing and who finances AI chip clusters.

Where is the stuck step for TPUs?

The hall. A TPU can be made, shipped and billed by Broadcom before it does any paid work. It earns for Google and the lab only once it is racked, powered and cooled in a finished building.

Broadcom said so on 2 September 2026. Its chief executive said land, power and shell “dictates specific timing of when this capacity gets deployed.” Of his own outlook, he asked: “even as we ship the chips, are they going to be deployed on a timely basis?”

Alphabet shows the same gap from the other side. It said it is “supply constrained,” will use third-party capacity in Q3 as a bridge, and expects most TPU-system revenue in 2027, not 2026.

Memory is the second pinch. It sits on the chip itself. Each new TPU carries more HBM, and memory makers have already committed most of next year’s output. For why memory binds briefly and power binds for years, see where the stuck step moves next.

Contracts carry the gap. Broadcom’s 8-K says Anthropic’s use of the extra capacity “is dependent on Anthropic’s continued commercial success.” Anthropic’s prospectus warns that some defaults could make a large part of its lease payments due at once, Reuters reported.

How big is the TPU build?

Big, and measured in gigawatts. Broadcom said on 2 September that, after 1 GW of Ironwood for Anthropic in 2026, it expects Anthropic to deploy another 5 GW of TPU v8i in 2027. Its April 8-K put Anthropic’s access through Broadcom at about 3.5 GW from 2027. The later figure is an expectation on a call, not a filed agreement.

For Google itself, Broadcom said it plans to deliver “multi-tens of billions of dollars of TPUs annually over the next several years.” It sees AI chip revenue of about $115 billion in fiscal 2027 and $230 billion in fiscal 2028, and says it has secured the supply for both. It put its own content at $20 billion to $30 billion per gigawatt.

What to check next

Three dated checks. 28 October 2026 (US time): Alphabet’s Q3 results, for TPU-system revenue and capex against its $195–205 billion 2026 guide. 9 December 2026: Broadcom’s Q4 results, against its $21.7 billion AI revenue guide. Through 2027: whether Anthropic’s 3.5 GW via Broadcom starts on time. That is the real test of whether the halls keep up with the chips.

Common misconceptions

• People assume a TPU is just Google’s GPU. Actually it is an ASIC built for matrix math. Google’s own guide sends models with many custom operations to GPUs.
• People assume only Google uses TPUs. Actually Anthropic plans up to one million, and Alphabet began delivering TPU systems to customer data centres in Q2 2026.
• People assume Google makes TPUs alone. Actually Broadcom develops and supplies them, under a long-term agreement filed on 6 April 2026.
• People assume a shipped chip is earning. Actually it earns only once the hall is powered. Alphabet expects most TPU-system revenue in 2027.

FAQ

Related terms

• HBM — the stacked memory on every TPU.
• CoWoS — how memory and compute are packaged side by side.
• GPU as a service — how rented accelerator hours are sold.
• Who makes AI chips — designers, foundries and packagers by role.

Stack doors

Layers: Semiconductors and Hyperscalers. Markets: where AI chip packaging shows up in Taiwan’s monthly sales.

Get the Wire — short daily notes when chip shipments, capex or compute contracts move.

Sources