Custom AI Chips Explained: Google TPU, Amazon Trainium and ASICs

Most AI computing runs on Nvidia’s graphics processors, or GPUs. But the largest cloud companies also design chips of their own for AI: Google has its Tensor Processing Units (TPUs), Amazon its Trainium chips, Meta its MTIA chips and Microsoft its Maia accelerators. These are custom chips, known in the industry as ASICs, short for application-specific integrated circuits. The cloud companies design them, but other listed companies help design and manufacture them, and those companies are on Pacific Tape’s SOX map and Taiwan map.

What is a custom AI chip?

A chip built for one kind of work rather than many. Google’s own definition: “Tensor Processing Units (TPUs) are Google’s custom-developed, application-specific integrated circuits (ASICs) used to accelerate machine learning workloads.” At the heart of a TPU is a matrix unit, “a 128x128 systolic array,” that does the multiplications AI models are made of.

Broadcom, which builds such chips for customers, describes the category in its annual report: “ASICs are custom products built to our individual customers’ specifications, such as our custom accelerators or XPUs, for hyperscalers, companies with AI frontier models and system integrators.”

How is that different from an Nvidia GPU?

The difference is flexibility. Nvidia describes its business as built on “a unified underlying programmable architecture,” with “the foundational NVIDIA CUDA development platform that runs on all NVIDIA GPUs,” so the same chips serve gaming, data centers, cars and other markets. Amazon puts the contrast this way: “Unlike GPUs—which are more flexible for a range of tasks—AWS Trainium chips are designed from the ground up specifically for AI workloads.”

Google’s documentation says TPUs suit “models dominated by matrix computations,” while GPUs suit models “with a significant number of custom PyTorch/JAX operations.” In short, a custom chip gives up some flexibility to do its owner’s main AI work more cheaply or efficiently. Nvidia lists “large cloud services companies with internal teams designing hardware and software that incorporate accelerated or AI computing functionality” among its competitors.

Which companies make their own AI chips?

Company	Chip	What the company has said
Google	TPU	Ironwood, its “seventh-generation” TPU (April 2025), has 192 GB of high-bandwidth memory per chip; Google says TPUs have powered its own AI work “for more than a decade”
Amazon	Trainium (and Inferentia)	Trainium3 (December 2025) is “AWS’s first 3nm AI chip”; Amazon designs its chips through Annapurna Labs, which it bought in 2015
Meta	MTIA	Developed since 2023, “a family of custom-built silicon chips to power our AI workloads efficiently”
Microsoft	Maia	Maia 200, announced in January 2026, is an accelerator for running AI models

These chips are a large business. Amazon said its chips business, which also counts its Graviton processors and Nitro hardware, passed “a $25 billion annual revenue run rate” in the second quarter of 2026. Alphabet said Google Cloud revenue rose 103% to $30.7 billion in the third quarter of 2026 and that “TPU systems sales were also a meaningful contributor,” and it has begun supplying TPU systems to some customers’ own data centers. For scale, Nvidia’s data center revenue was $89.0 billion in its quarter ended July 26, 2026.

Who helps design them?

The cloud companies set what the chip must do, and specialist chip designers help turn that into a chip that can be manufactured. Five listed companies say they do this work for AI accelerators:

Broadcom is the largest. In an April 6, 2026 filing it said it and Google “have entered into a Long Term Agreement for Broadcom to develop and supply custom Tensor Processing Units (‘TPUs’) for Google’s future generations of TPUs,” running up to 2031. Meta said in April 2026 that it is “partnering with Broadcom to co-develop multiple generations of custom silicon.” Broadcom’s AI semiconductor revenue was $16.7 billion in its quarter ended August 2, 2026, up 221% from a year earlier, and it forecast $21.7 billion for the following quarter.

Marvell develops “custom semiconductor solutions tailored to individual customer specifications,” including for AI, and in December 2024 signed a “five-year, multi-generational agreement” with Amazon Web Services that includes “the supply of custom AI products.” Data center products were 79% of its revenue in its quarter ended August 1, 2026. It does not name its custom chip customers.

MediaTek, best known for smartphone chips, said on October 30, 2026 that it began shipping its first AI accelerator for a US cloud customer in September and is “confident of delivering 2.5 billion US dollars revenue in the fourth quarter.” It does not name the customer.

Global Unichip (GUC) and Alchip, both listed in Taiwan, design custom chips for other companies. GUC said its “Cloud” category, mainly “CPU, AI accelerators, and BMC projects,” was 86% of its third-quarter 2026 revenue, which rose 118% from a year earlier. Alchip said shipments of a 3-nanometer accelerator to a North American cloud customer started in May 2026. Taiwan’s Chip Stocks: Who Makes What covers both companies.

Who manufactures them?

Mostly TSMC. Broadcom says that in fiscal 2025 “approximately 95% of the wafers manufactured by our CMs were produced by TSMC,” and that its wafer needs are “a meaningful portion of TSMC’s total production capacity.” TSMC owns about 34.8% of GUC, and TSMC’s annual report lists AI ASICs alongside AI GPUs among the high-performance computing products that were 58% of its 2025 revenue. Google and Amazon do not say where their chips are made.

Custom accelerators also use high-bandwidth memory (HBM), like Nvidia’s GPUs: each Ironwood TPU carries 192 GB. That links them to the memory makers covered in How Samsung and SK hynix Compare in HBM, and the chips are often packaged with TSMC’s CoWoS, explained in What Is CoWoS?

What does this mean for reading chip stocks?

That AI chip demand is not only Nvidia’s story. Part of the cloud companies’ spending goes to chips they design themselves, and that shows up in the results of the companies that help build them: Broadcom’s AI revenue, Marvell’s data center sales, and the custom-chip lines of MediaTek, GUC and Alchip. It also reaches TSMC, which manufactures for nearly all of them. Most of these companies do not name the customers behind those numbers; Broadcom’s Google filing is an exception. For how Nvidia’s own results reach Asian chip stocks, see Why Nvidia’s Earnings Move Memory Stocks, and for who designs and who manufactures in general, see Foundry, Fabless and IDM.