The earlier Blackwell gate covered NVIDIA’s GPUs, which are the mainstay of AI chips. But the story has another thread: cloud giants like Google, Amazon, Meta, and Microsoft don’t just buy GPUs from NVIDIA — they also “build their own chips.” Most of these in-house AI chips are a thing called an ASIC.
This piece spells out the ASIC. First what it is and how it differs from a GPU, then why the giants build their own, the role of Broadcom and Taiwan’s firms, and whether it will replace NVIDIA. This 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, which in plain terms is “a chip tailored to one specific task.”
Compare it with a GPU. A GPU is a more general-purpose parallel-computing chip: it covers a broad range, can run many models and workloads, and is highly flexible — but precisely because it has to be general, it carries some costs you “pay for but don’t use.” An ASIC is the reverse: it locks onto one class of work (say a certain kind of inference) and designs the chip to be just enough and just as lean as needed, trading away general flexibility for cost, power, and efficiency advantages on that one thing. Google flatly defines its own TPU as an AI-purpose ASIC.
Here’s an analogy: a GPU is like a Swiss army knife that can cut anything; an ASIC is like a chef’s knife custom-made for one dish — it does only that, but does it fast and cheap.
Why Cloud Giants Build Their Own Chips
Cloud giants can clearly buy NVIDIA’s GPUs, so why spend big to build their own? Mainly four reasons.
One is cost: the sheer number of chips their data centers need is so large that building in-house saves substantial procurement and operating expense. Two is optimization: a chip tailored to their own models, recommendation systems, and inference services runs more efficiently than a general-purpose GPU. Three is reducing reliance: they don’t want to bet their lifeline entirely on a single supplier. Four is differentiation: tying the chip to their own cloud creates a combination others can’t copy. Add up these four, and it’s enough to keep the deep-pocketed giants investing for the long haul.
Core-Data Snapshot
Below we capture the players and the scale of custom ASICs. Growth rates and shares are research-firm estimates.
| Topic | Data | Timing / Nature |
|---|---|---|
| Google TPU (seventh-gen Ironwood already in production; eighth-gen 8t/8i unveiled 2026-04) | 192GB HBM3E per chip; a superpod (a giant cluster of thousands of chips strung together) up to 9,216 chips | 2025-11 GA |
| AWS Trainium3 | First 3nm AWS AI chip; UltraServer up to 144 chips | Late 2025 |
| Microsoft Maia 200 | TSMC 3nm, 216GB HBM3e, inference-focused | Announced 2026-01 |
| CSP custom-ASIC shipment growth (2026) | About 44.6% (GPU about 16.1%) | TrendForce estimate |
| AI server GPU vs ASIC share (2026) | GPU about 69.7%, ASIC about 27.8% | TrendForce estimate |
The Four Big Clouds’ Custom Chips
Let’s lay out the main players:
Google TPU: the most veteran custom ASIC. The one in production is the seventh-gen Ironwood, formally commercialized at the end of 2025, where a single cluster (a superpod, a giant compute cluster stringing thousands of chips together) can scale to over nine thousand chips; Google also unveiled its eighth-gen TPU (8t/8i) in April 2026. Even Anthropic has expanded its use of Google TPUs to run Claude. AWS Trainium/Inferentia: Trainium handles training and Inferentia handles inference, with the latest Trainium3 being AWS’s first 3nm AI chip. Meta MTIA: Meta officially says it has deployed “hundreds of thousands of” MTIA chips for inference, with several more generations to come over the next two years. Microsoft Maia: the Maia 200, announced in early 2026, uses TSMC’s 3nm process, focuses on inference, and serves Microsoft’s own Copilot and OpenAI models.
Separately, OpenAI is also co-designing its own accelerator with Broadcom and planning a massive deployment, but the target is to start landing only from the second half of 2026 — it’s still in the design and build-out phase, so it shouldn’t be treated as a mass-produced fact.
Broadcom, Marvell, and Taiwan: Who’s Helping Them Build
Cloud giants have the ideas, but the chip’s detailed design, packaging, and interconnect still need specialist partners. That brings out two kinds of roles.
The first is the ASIC co-designers, with Broadcom and Marvell as the representatives. They don’t just sell parts — they design the custom chips together with the cloud giants and supply the networking interconnect. Broadcom has publicly disclosed long-term collaborations with Google TPU, Meta MTIA, and OpenAI.
The second is Taiwan’s ASIC design-service firms, with Alchip and GUC (Global Unichip) as the representatives, offering the full turnkey stack from design to contracted manufacturing (turnkey, meaning they handle the whole package from design through farming out the foundry work); some 3nm-related designs have already entered customer products or volume production, while 2nm is mostly still at the design-platform and customer-development stage. To be clear: the actual customers of these design-service firms are mostly confidential, and when the market pairs them to a given cloud giant, that’s often analyst speculation or market chatter, not a company announcement. This article only describes industry roles; it does not compile a list of beneficiary stocks, nor does it constitute investment advice.
“ASIC chip stocks” are a hot theme in Taiwan’s stock market, and the market usually lumps the whole supply chain — design services, silicon IP, advanced packaging, foundry, test interfaces — into the discussion. It’s fine to understand what’s driving the buzz in this group, but don’t mistake a hot theme for a stock recommendation.
Will ASICs Replace NVIDIA?
This is the most frequently asked question, and the answer has to keep “growth rate” and “total volume” separate.
In 2026, the shipment growth rate of cloud giants’ custom ASICs (about 44.6%) is indeed faster than GPUs (about 16.1%). But growing fast doesn’t equal being large in scale: research firms estimate that in 2026 AI servers are still GPU-led, at around 70%, with ASICs at close to 30%. More crucially, NVIDIA’s moat isn’t only in the chip itself — there’s also the CUDA software ecosystem everyone uses (the programming environment that locks in developers) and the NVLink high-speed interconnect, and moving the whole stack off them is very costly.
So the more pragmatic view is this: custom ASICs are the cloud giants’ long-term route to “save themselves money and differentiate,” gradually taking over a portion of the demand that used to belong to GPUs, but in the short term it’s not a matter of one replacing the other. One telling signal is that even Anthropic runs its models simultaneously on Google TPU, AWS Trainium, and NVIDIA GPU — using multiple vendors and spreading its bets is the norm right now.
Key Takeaways for This Gate
After looking at the ASIC, first remember its positioning: it’s a custom chip tailored to one specific task, trading away general flexibility for efficiency and cost.
Cloud giants build their own ASICs to save cost, optimize workloads, reduce reliance, and differentiate; Google TPU, AWS Trainium, Meta MTIA, and Microsoft Maia are the representatives, often with Broadcom, Marvell, and Taiwan’s design-service firms in the background. In 2026 ASICs are growing faster than GPUs, but GPUs still hold roughly 70% and NVIDIA’s ecosystem moat is still in place — for the short term it’s coexistence, not replacement.
To see the GPU-mainstay side, go back and read What Is Blackwell; to see how these chips get packaged and fed data, see CoWoS and HBM; to see all eight gates of the chain, head back to the supply-chain overview.