The earlier Blackwell gate covered NVIDIA’s GPUs, the mainstream of AI chips. There is another thread: cloud giants like Google, Amazon, Meta, and Microsoft. They buy GPUs from NVIDIA, but they also build their own chips, and most of these in-house AI chips are ASICs.
This piece is the deep-dive offshoot of Gate 1, “AI chips,” in The AI Hardware Supply Chain, End to End.
What Is an ASIC? How Is It Different from a GPU
ASIC stands for Application-Specific Integrated Circuit. In plain terms, it is “a chip built for one specific task.”
A GPU is a general-purpose parallel-computing chip. It can run many kinds of models and workloads, so it is flexible, but it also pays for features a given job never uses. An ASIC is the reverse: it targets one class of work (say, a certain kind of inference) and gives up flexibility to win on cost, power, and efficiency for that job. Google itself defines its TPU as an AI-purpose ASIC.
An analogy: a GPU is a Swiss army knife that can cut anything; an ASIC is a chef’s knife made for one dish. It does one thing, but does it fast and cheap.
Why Cloud Giants Build Their Own Chips
They can buy NVIDIA GPUs, so why spend big to build their own? Four main reasons:
- Cost: their data centers use so many chips that building in-house saves a lot on purchasing and running costs.
- Optimization: chips designed around their own models, recommendation systems, and inference services.
- Less dependence: they don’t want to bet everything on one supplier.
- Differentiation: tying the chip to their own cloud creates a combination others can’t copy.
The Four Big Clouds’ Custom Chips
| Chip | Status (as of October 2026) |
|---|---|
| Google TPU | Seventh-gen Ironwood is generally available, with 192GB of HBM per chip and up to 9,216 chips in one pod; eighth-gen 8t/8i was unveiled in April 2026, and Google Cloud still marks it “coming soon” |
| AWS Trainium | Trainium3 is AWS’s first 3nm AI chip; it began shipping in early 2026 and is nearly sold out; Trainium4 is expected to be broadly available only around 2027 |
| Meta MTIA | Inference-focused; in March 2026 Meta said it had deployed hundreds of thousands of chips, MTIA 300 is in mass production, and three more generations are coming within two years |
| Microsoft Maia | Maia 200 was announced in January 2026; built on TSMC’s 3nm process and focused on inference |
Custom chips are also starting to be sold outside. Alphabet’s CFO said that starting in Q2 2026, Google for the first time delivered TPU systems directly into customers’ own data centers. Amazon said in July 2026 that its chips business had passed $25 billion in annualized revenue, and that both Anthropic and OpenAI had made multi-year, multi-GW commitments to Trainium.
OpenAI is also working with Broadcom. In June 2026 they unveiled OpenAI’s first inference chip, Jalapeño. Engineering samples are already running models in the lab, and the target is initial deployment before the end of 2026. Broadcom CEO Hock Tan said volume ramps only in 2027, with full speed in the first half of 2028, so it is not yet a mass-produced chip.
Broadcom, Marvell, and Taiwan: Who’s Helping Them Build
The cloud giants have the ideas, but detailed chip design, packaging, and interconnect still need specialist partners. They fall into two groups.
ASIC co-designers: Broadcom and Marvell don’t just sell parts. They design custom chips together with the cloud giants and supply networking interconnect. Broadcom has publicly disclosed work with Google TPU, Meta MTIA, and OpenAI. In fiscal 2026 Q3 (ended August 2, 2026), its AI semiconductor revenue was $16.7 billion, up 221% year over year. Marvell also disclosed an expanded custom chip collaboration with Google in an August 2026 SEC filing.
Taiwan’s ASIC design-service firms: Alchip and GUC (Global Unichip) offer a full turnkey service from design to contracted manufacturing. Their actual customers are mostly confidential, and when the market pairs them with a given cloud giant, that is often analyst speculation, not a company announcement. This article only describes industry roles; it does not list beneficiary stocks and is not investment advice.
“ASIC concept stocks” are a hot theme in Taiwan’s stock market. The market usually groups design services, silicon IP, advanced packaging, foundry, and test interfaces together. It’s fine to understand why the group is hot, but don’t mistake a hot theme for a stock recommendation.
Will ASICs Replace NVIDIA?
The answer depends on separating “growing” from “how big a share.” In April 2026, research firm TrendForce trimmed its estimate for the share of ASIC-based AI servers in 2026 from nearly 28% to about 27%. Most are still GPU-based.
NVIDIA’s moat is also not just the chip itself. It also has the CUDA software ecosystem that most developers use and the NVLink high-speed interconnect, and moving the whole stack away is very costly.
Even AI companies use several vendors at once. Anthropic uses both Google TPU and AWS Trainium, and AMD’s August 2026 earnings report said Anthropic will deploy up to 2GW of AMD GPUs. Custom ASICs will slowly take some demand that used to go to GPUs, but everyone is spreading their bets, not switching sides completely.
Penchan’s summary
An ASIC is a custom chip that gives up flexibility for efficiency and cost. Behind Google TPU, AWS Trainium, Meta MTIA, and Microsoft Maia, you often find Broadcom, Marvell, and Taiwan’s design-service firms.
For the GPU side, read What Is Blackwell. To see how these chips get packaged and fed data, see CoWoS and HBM. For all eight gates of the chain, head back to the supply-chain overview.
Custom ASICs are the cloud giants’ long-term way to save money and stand out. GPUs are still the mainstream; in the short term the two coexist, and one does not replace the other.