TPU vs GPU: Google’s Chip, Nvidia’s Chip, and Who Gets Paid

EXPLAINED · Versus · Last reviewed: 08 Oct 2026 · Next review: after Alphabet Q3 2026 results (28 Oct 2026)

TPU vs GPU: a TPU is a chip Google designs for AI matrix math only; a GPU is a general-purpose parallel chip that runs AI and much else.

Why an investor cares: the chip choice decides who books the money, and at what margin. A GPU dollar goes to Nvidia. A TPU dollar is split: Broadcom is paid to build it, and Google keeps the design. For how a TPU works inside, see what a TPU is.

What is the difference between a TPU and a GPU?

Focus. A TPU does one job, the large matrix math that trains and runs AI models. A GPU does many kinds of parallel work, and AI is now its biggest use.

Google Cloud calls the TPU a custom application-specific integrated circuit, or ASIC. Google designs it. Broadcom develops and supplies it. A GPU is designed by a chip company such as Nvidia, which sells the chip, or a full rack with its own CPUs and networking.

Sources: Google Cloud documentation (2 and 6 Oct 2026); Broadcom and Nvidia results calls. GPU memory per chip = machine HBM ÷ GPU count; 192 GiB ≈ 206 GB, our arithmetic.

Is a TPU better than a GPU?

It depends on the work. Google Cloud’s guide sends models dominated by matrix math, with large batches and long training runs, to TPUs. It sends models with many custom operations that must run partly on CPUs to GPUs.

Each side also makes its own claim. Broadcom said on 2 September 2026 that the new TPU v8i is comparable to, if not better than, Nvidia’s Vera Rubin GPU, and that custom chips run at “less than half the cost” of a GPU. Nvidia said on 26 August 2026 that such chips are built “for one cloud or one service,” while its GPUs run every model and are “in every cloud.” These are company claims. Our sources hold no independent test.

Is a TPU cheaper than a GPU?

Per chip, on Google Cloud’s price list, yes. On a 3-year commitment an Ironwood TPU costs $5.40 an hour. An Nvidia B200 costs about $7.09 an hour, worked out from an 8-GPU machine.

Source: Google Cloud TPU pricing and accelerator-optimized VM pricing, checked 8 Oct 2026. The B200 price is for an 8-GPU a4-highgpu-8g machine, which includes 224 vCPUs, 3,968 GB of memory and local SSD; per-GPU figures are our arithmetic. A price per chip is not a price per unit of work.

A chip-hour is not a unit of work, and the GPU machine price also covers CPUs, memory and disks. The fair test is cost per finished job.

Who gets paid: TPU vs GPU

Different companies are paid at each step.

Sources: Broadcom Q3 FY2026 call (2 Sep 2026); NVIDIA CFO commentary and call (26 Aug 2026); Alphabet Q2 2026 call (22 Jul 2026).

Nvidia books the most per chip. Its Data Center revenue was $89.0 billion in the quarter to 26 July 2026, with $48.7 billion, about 55%, from hyperscalers (48.7 ÷ 89.0, our arithmetic). Broadcom’s AI chip revenue was $16.7 billion in the quarter to 2 August 2026. Custom chips like the TPU were 73% of that, about $12.2 billion (our arithmetic), across six customers, not only Google. For the full map of roles, see who makes AI chips.

Per gigawatt of data centre, Nvidia said it takes about $40 billion with Vera Rubin. Broadcom said its content is $20 billion to $30 billion per gigawatt. Both are statements on results calls, not audited figures.

Why does Google sell both TPUs and GPUs?

Because its customers ask for both. Alphabet said on 22 July 2026 that it offers “the industry’s broadest range of accelerators from Google and NVIDIA,” including Nvidia’s Vera Rubin and its own TPU 8t and 8i. Its software runs across both, and Google uses both to serve its own models.

That makes Google both a buyer and a seller. It buys GPUs from Nvidia and pays Broadcom for TPUs. Its capex was $44.9 billion in Q2 2026, mostly on technical infrastructure, and about 60% of that on servers. It earns back when hours are used: Google Cloud revenue rose 82% to $24.8 billion. Since Q2 it also sells whole TPU systems into customers’ own data centres, with most of that revenue expected in 2027. See what a hyperscaler is and how accelerator hours are sold.

Where is the stuck step for TPUs and GPUs?

Memory. Both chips carry high-bandwidth memory (HBM) on the package, and the amounts are close. On Google Cloud, an Nvidia B200 has 180 GB and a GB300 has 279 GB. An Ironwood TPU has 192 GiB, about 206 GB.

Both chip makers say memory is the pinch. Nvidia’s supply commitments rose from $119 billion to $279 billion in one quarter, “primarily related to the procurement of memory.” It called memory prices “extreme” and said they are “headed even higher into next year.” Broadcom said the rising memory on its custom chips is diluting its gross margin, and named HBM among the inputs that can each become the bottleneck. For who makes it, see who makes HBM and HBM, explained.

So the memory makers are paid either way. Picking a TPU over a GPU changes who books the chip revenue. It does not shorten the memory queue. For what memory prices look like now, see Samsung’s 2027 HBM4 ask. After memory, the next wait is the powered hall; see where the stuck step moves next.

What to check next

Three dated checks. 28 October 2026 (US time): Alphabet’s Q3 results, for TPU-system revenue and capex. 17 November 2026: Nvidia’s Q3 fiscal 2027 results, the date it gave on its last call, against its $108 billion revenue guide and 74% margin. 9 December 2026: Broadcom’s Q4 results, against its $21.7 billion AI revenue guide. In each, watch what they say about memory costs.

Common misconceptions

• People assume a TPU is Google’s version of a GPU. Actually it is an ASIC built for matrix math. Google’s own guide still sends some work to GPUs.
• People assume Google uses only its own chips. Actually Google Cloud sells Nvidia GPU hours too, including Vera Rubin, and uses both chips itself.
• People assume a cheaper chip-hour means cheaper AI. Actually the fair test is cost per finished job. A chip-hour price says nothing about speed.
• People assume choosing TPUs avoids the memory squeeze. Actually both chips carry HBM. Broadcom says memory is diluting its margin too.

FAQ

Related terms

• TPU — how Google’s chip works and who builds it.
• HBM — the stacked memory on both chips.
• GPU as a service — how accelerator hours are sold.
• Neocloud vs hyperscaler — who else sells GPU hours.

Stack doors

Layers: Hyperscalers and Semiconductors. Wire: Micron’s 2027 memory commitments.

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

Sources