Glossary · from SEC filings
TPU vs GPU: what is a TPU?
A TPU, or Tensor Processing Unit, is Google’s own chip for AI. A GPU, or graphics processing unit, is the general-purpose accelerator Nvidia and AMD sell to everyone. The comparison matters to investors because custom chips like the TPU are the hyperscalers’ main way to reduce what they pay Nvidia, and in 2026 Alphabet started selling TPU systems to other companies.
What a TPU is
Google designed the TPU for the mathematics at the heart of neural networks: multiplying very large grids of numbers, the tensors in its name. It strips out much of what a general-purpose chip needs and spends the space on that one job, which makes it efficient for training and running Google’s own models. Alphabet’s 10-K describes its seventh generation, Ironwood, and says it offers customers both GPUs and its own TPUs through Google Cloud.
A GPU started as a graphics chip and became the standard AI accelerator because it was already built to do many calculations at once. Nvidia’s advantage is not only the chip but the software around it, CUDA, which a generation of developers has written their code for, and the networking and complete systems it now sells alongside.
Alphabet: Alphabet on its accelerators
“Our technical infrastructure allows us to use and offer our customers a range of AI accelerator options, including specialized Graphics Processing Units (GPUs) and our own custom-built Tensor Processing Units (TPUs), such as Ironwood, our seventh-generation TPU.”
Alphabet 10-K, filed February 5, 2026
TPU vs GPU: the trade-offs
A GPU is flexible: it runs nearly any model, on any cloud, with the software most developers already use. A TPU is specialized: very efficient for the workloads it was designed for, but available mainly through Google and tied to its software. For a buyer, the GPU is the safe choice and the TPU the cheaper one at scale, if its models fit.
The economics are what matter for investors. Nvidia keeps an unusually large share of each dollar it sells: a gross margin of 75% over its latest four quarters. A hyperscaler that designs its own chip pays a design partner and a foundry instead, and keeps the difference. That is why Google, Amazon, Microsoft and Meta all design chips of their own, even while buying large numbers of Nvidia GPUs.
Alphabet now sells TPU systems
Until recently TPUs were available only as a service inside Google Cloud. Alphabet’s second-quarter 2026 10-Q says that in that quarter it began recognizing revenue from the sale of TPU systems, which it describes as hardware, software, installation, support and extended warranty, and that its inventory now consists primarily of TPU-system hardware for sale to enterprise customers, alongside its Pixel devices.
That turns Google from a TPU user into a TPU seller, competing directly with Nvidia for some customers. Google Cloud brought in $93.2B of revenue over its latest four quarters, +76%; Alphabet doesn’t yet report TPU sales separately, so the size of the business is not yet visible in the filings. Watch for it in future 10-Qs.
Alphabet: Alphabet begins selling TPU systems
“In addition, in the second quarter of 2026, we began recognizing revenue from the sale of TPU systems.”
Alphabet 10-Q, filed July 23, 2026
Alphabet: What Google Cloud’s product sales are
“Our Google Cloud product sales generally consist of the sale of TPU systems comprising hardware, software, installation, support, and extended warranty services.”
Alphabet 10-Q, filed October 29, 2026
Alphabet: Alphabet’s inventory
“Inventory consists primarily of hardware related to TPU systems for sale to enterprise customers and consumer devices, which primarily include the Pixel family of products.”
Alphabet 10-Q, filed October 29, 2026
Who designs and makes custom AI chips
Most custom AI chips are designed with a partner. Broadcom is the largest: its 10-K describes custom silicon that integrates logic, high-bandwidth memory and other components using advanced packaging, and it reported $89.1B of revenue over its latest four quarters, +49%. Marvell is the other large custom-silicon designer. Whatever the design, the chips are manufactured mostly by TSMC, packaged with HBM, and assembled into systems, so custom chips shift money between layers of the stack more than they remove it.
Amazon designs its own Trainium and Inferentia chips, and its filings describe customer commitments that include obligations tied to the performance of AWS’s chips. Microsoft and Meta have their own designs too. See the chips layer, Broadcom and Marvell.
Broadcom: Broadcom’s custom silicon
“Our custom silicon solutions provide the platform to integrate embedded logic, high-bandwidth memory, serializer/deserializer technology, IP cores and processor cores using advanced packaging technologies.”
Broadcom 10-K, filed December 18, 2025
Amazon: $38B AWS commitment, expanded by $100B over 8 years
“In Q1 2026, AWS and OpenAI Group PBC ("OpenAI") announced an expansion of the existing $38.0 billion multi-year commitment and commercial arrangement with OpenAI by $100.0 billion over 8.0 years, which includes contractual obligations related to the performance of AWS chips.”
Amazon 10-Q, filed July 31, 2026
How to compare AI chips as an investor
Chip specifications make headlines, but three other things decide who wins. Cost per unit of useful work: what a buyer pays, including power, to train or run a model, which is where custom chips usually aim to win. Software: how much work it takes to move a model onto the chip, which is where Nvidia’s CUDA gives it an advantage that specifications don’t show. And availability: a chip a buyer can get this quarter beats a better one it can get next year, which is why supply commitments matter so much.
Watch the buyers’ own words too. Alphabet now describes offering customers a choice of GPUs and TPUs, and AWS contracts with its largest AI customers include obligations tied to its own chips. When a hyperscaler’s largest customers commit to its in-house silicon, that silicon has moved from an experiment to a product line, with consequences for every merchant chip supplier.
For investors the useful question is not which chip is best but who captures the value. When Nvidia sells a GPU, it keeps most of the gross profit. When a hyperscaler deploys its own chip, the profit is split among the hyperscaler, its design partner, TSMC and the memory and packaging suppliers. Custom chips therefore move money from Nvidia toward Broadcom, Marvell and the hyperscalers themselves, more than they shrink the total spent.
GPUs from AMD, and the other alternatives
Not every alternative to Nvidia is a custom chip. AMD sells GPUs to the same customers on the same terms as Nvidia, and has used customer warrants to win very large commitments: OpenAI and Meta each intend to deploy up to 6 gigawatts of AMD data-center GPUs. That makes AMD the main merchant alternative, competing with Nvidia on price and supply rather than asking customers to design their own chips. See the AMD company page.
AMD: AMD’s OpenAI and Meta agreements
“In October 2025 and February 2026, we entered into multi-year agreements with OpenAI OpCo, LLC (OpenAI) and Meta Platforms, Inc. (Meta), respectively, under which each customer intends to deploy up to 6 gigawatts of AMD data center GPUs, with the first gigawatt of each deployment powered by our AMD Instinct MI450 series products.”
AMD 10-Q, filed August 5, 2026
What it means for Nvidia
Nvidia’s data-center revenue was $278B over its latest four quarters, +90%: custom chips have not slowed it yet. Alphabet’s own 10-K says supply capacity for AI accelerators, GPUs and TPUs alike, is highly competitive and rapidly evolving, which is another way of saying demand still exceeds what anyone can make.
The risk for Nvidia is gradual rather than sudden. Each workload a hyperscaler moves to its own chip is one it no longer buys from Nvidia, and the most common, stable workloads are the easiest to move. The first place that would show is Nvidia’s gross margin, if buyers with credible alternatives start negotiating harder. See Nvidia’s customers and suppliers.
Alphabet: Alphabet on accelerator supply
“For example, industry supply capacity for AI accelerators, including GPUs as well as our custom-built TPUs, is highly competitive and rapidly evolving.”
Alphabet 10-K, filed February 5, 2026
Related terms
More from the glossary: Hyperscaler, Neocloud, High-bandwidth memory (HBM), Remaining performance obligations (RPO), AI data center, CoWoS (advanced packaging), Vendor financing, and every other term you need to read AI companies’ filings.
Free · quarterly · from SEC filings
Get the AI Stack Report every quarter
After each Big Tech earnings season we update the map from the new filings and publish what changed: who spent, who got paid, who kept it and what the filings reveal about the loop. Free, in Rob May’s Investing in AI newsletter.
Form not showing? Subscribe on Substack. No spam; unsubscribe any time.
Questions about TPUs and GPUs
What is a TPU?
A Tensor Processing Unit: Google’s custom chip for AI, designed for the matrix calculations at the core of neural networks. Alphabet’s 10-K describes its seventh generation, Ironwood.
What is the difference between a TPU and a GPU?
A GPU is a general-purpose accelerator that runs almost any model with widely used software; a TPU is specialized for neural-network math and available mainly through Google. GPUs are more flexible; TPUs can be cheaper at scale for workloads that fit.
Is a TPU better than a GPU?
Neither is better in general. TPUs can be more efficient for the workloads Google designed them for; GPUs run a wider range of models and software. Most large buyers use both.
Does Google sell TPUs?
Yes, since 2026. Alphabet’s second-quarter 10-Q says it began recognizing revenue from the sale of TPU systems in that quarter, and that its inventory now consists primarily of TPU-system hardware for enterprise customers.
Who makes Google’s TPUs?
Google designs them, typically with a design partner, and they are manufactured by a foundry. Broadcom is the largest custom AI chip design partner, and TSMC manufactures most leading AI chips.
Are TPUs a threat to Nvidia?
They are the main long-term competitive question for Nvidia, along with other hyperscalers’ custom chips. So far Nvidia’s data-center revenue has kept growing fast; the first sign of pressure would be its gross margin.
What is the difference between an ASIC and a GPU?
An ASIC (application-specific integrated circuit) is designed for one set of tasks, as a TPU is for neural-network math. A GPU is programmable for many kinds of parallel work. ASICs can be more efficient for their target; GPUs are more flexible.
Can you buy TPUs outside Google Cloud?
Since 2026, yes: Alphabet’s 10-Q says it began selling TPU systems, including hardware, software, installation and support, to enterprise customers.
Why do hyperscalers design their own chips?
To lower the cost of AI computing and reduce dependence on a single supplier. A custom chip lets a hyperscaler keep part of the margin Nvidia would otherwise earn, at the cost of design work and less flexibility.
What is Ironwood?
Google’s seventh-generation TPU, named in Alphabet’s 10-K as one of the AI accelerators it uses and offers to Google Cloud customers.
What other custom AI chips are there?
Amazon’s Trainium and Inferentia, and chips designed by Microsoft and Meta, usually with partners such as Broadcom or Marvell.
Research and education, not investment advice: nothing here is a recommendation to buy or sell any security. Figures come from companies’ SEC filings as reported; companies named were chosen because they show how the stack works, not as a view on their shares.