Custom AI chips (ASICs) — why Google builds its own instead of buying Nvidia
A custom AI chip, or ASIC, is a processor designed for the specific workload of one company rather than sold as a general-purpose part. The Google TPU, Amazon Trainium, Microsoft Maia and Meta MTIA all qualify. The co-design market that turns a hyperscaler's specification into manufacturable silicon is estimated to be about 95 percent Broadcom and Marvell, with Broadcom above 70 percent. In 2026 custom ASIC shipment growth is estimated to have passed merchant GPU growth for the first time — roughly 44.6 percent against 16.1 percent
The three lines
- Definition — silicon designed for one company's one workload, not sold as a general part
- Why — at hyperscale, generality is waste; the metrics are now cost per token and power
- Market — co-design is ~95% Broadcom and Marvell; ASIC growth passed GPU growth in 2026
Key questions
- How is a custom AI chip different from a GPU?
- **They are designed from opposite starting assumptions.** A GPU is built for a world where **you do not know what will run on it** — model architectures change, research shifts, and it has to work for every buyer. Generality is mandatory, and it costs silicon area and power. A custom ASIC starts from the opposite premise: **the workload is already known**. Google knows which operations its models use and in what proportion, and lays out circuits to match. Anything unused is simply not built. The result is **the same work at lower power and lower cost** — with one trade: if the model architecture changes substantially, the chip becomes an awkward fit. Hence the division of labour: **GPUs for research and variety, ASICs for fixed workloads at enormous scale**.
- Why do Google and Amazon build their own chips?
- **Past a certain scale, building is cheaper than buying.** A custom chip costs hundreds of millions of dollars and two to three years before it exists. Without volume to spread that over, it is a loss. Hyperscalers run **the same computation billions of times a day**, and there a few percent of power efficiency becomes hundreds of millions of dollars a year. The purchasing criteria themselves have moved: from **peak throughput and bandwidth** to **cost per token, power, cooling, utilisation and total cost of ownership** — every one of which favours purpose-built silicon.
- Who actually builds these custom chips?
- **The hyperscaler writes the specification; two companies turn it into silicon.** Broadcom and Marvell hold an estimated **95 percent** of that co-design market between them, with **Broadcom above 70 percent** of custom accelerators. Broadcom reported AI revenue of **$8.4 billion** in its most recent quarter, up **106 percent** year over year; Marvell has guided to as much as **$11 billion** of AI ASIC revenue for 2026. Their work is physical design, verification, packaging and the path to foundry production. That makes them less Nvidia's competitors than **the subcontractors for customers trying to route around Nvidia**.
Nvidia GPUs sell well. And the companies buying the most of them are simultaneously building their own chips.
Google, Amazon, Microsoft, Meta and OpenAI all do. That apparent contradiction is what this piece is about.
1. What "custom" means
ASIC stands for Application-Specific Integrated Circuit. The name is the definition.
| Property | GPU (general purpose) | Custom AI chip (ASIC) |
|---|---|---|
| Design premise | you don't know what will run | you do know what will run |
| Audience | sold to everyone | runs one company's workload |
| Circuitry | includes what you may never use | only what gets used |
| If the architecture changes | still fine | becomes an awkward fit |
| Time and cost | you buy it | 2–3 years, hundreds of millions |
The last two rows are the trade: give up flexibility, get efficiency.
2. What exists today
| Company | Chip | Primary use |
|---|---|---|
| TPU | training and inference (the oldest example) | |
| Amazon (AWS) | Trainium · Inferentia | training / inference split |
| Microsoft | Maia | its own cloud and Copilot |
| Meta | MTIA | recommendation and ranking workloads |
| OpenAI | Jalapeño | its own models |
Different names, one logic: the workload is fixed and the volume is there.
3. Why now — the criteria moved
| Old criteria | Current criteria |
|---|---|
| peak throughput | cost per token |
| memory bandwidth | power (W) |
| top-line performance | cooling load |
| — | utilisation |
| — | total cost of ownership |
Every item on the right favours purpose-built silicon. A general-purpose chip is designed to be ready for anything, and that readiness shows up as watts and area.
Data-center power has become a real constraint on top of that, which raised the pressure further.
4. The transition, in numbers
| Metric | 2026 estimate |
|---|---|
| Custom ASIC shipment growth | ~44.6% |
| Merchant GPU shipment growth | ~16.1% |
That is estimated to be the first year ASIC growth passed GPU growth. But it is growth, not size. In absolute units and dollars GPUs remain far larger. The accurate reading is a small number compounding fast.
5. Who designs, who builds
| Role | Who |
|---|---|
| Specification | Google, Amazon, Microsoft, Meta (hyperscalers) |
| Co-design → manufacturable silicon | Broadcom · Marvell |
| Fabrication | foundries (TSMC and others) |
| Memory | HBM suppliers |
The middle row is the bottleneck. Only two companies effectively carry a specification through physical design, verification and packaging.
| Company | Position |
|---|---|
| Broadcom | 70%+ of custom accelerators; recent quarter AI revenue $8.4bn (+106%) |
| Marvell | Works on Amazon Trainium and Microsoft Maia; guides to up to $11bn AI ASIC revenue in 2026 |
| Combined | roughly 95% of co-design |
Which means the more hyperscalers route around Nvidia, the bigger Broadcom and Marvell get. Marvell's 8 percent drop on August 28, 2026 — over a Google contract worth up to $120 billion through fiscal 2033 — happened inside this structure, and the problem was timing, not size.
6. Why GPUs do not disappear
| Situation | Right chip |
|---|---|
| Research with changing architectures | GPU |
| Startups and enterprises with varied workloads | GPU |
| Domains where no standard has settled | GPU |
| Fixed work repeated at enormous scale | ASIC |
| Shaving inference cost to the cent | ASIC |
An ASIC is the choice that comes after a workload settles, not before.
7. Frequently asked
Is Nvidia in trouble? Losing on growth rate is not the same as losing revenue. ASICs take mainly the hyperscalers' fixed workloads; everything else still buys general-purpose parts. The structural risk is that the largest customers defect first.
Could a normal company build one? Volume decides. With two to three years and hundreds of millions of dollars up front, unless you repeat the same computation enormously, renting GPUs is cheaper.
Is a TPU faster than a GPU? The question does not resolve. It is usually more efficient on its own workload and may not be on others — and most published numbers are vendor claims with little independent verification.
How do chip costs show up in the accounts? Depreciation schedules change reported profit materially, which is covered in "What GPU depreciation is."
Do ASICs raise HBM demand? Yes. Custom chips hit the same bandwidth wall. Memory suppliers benefit regardless of which side of this contest wins.
8. What we could not confirm
- Growth rates — 44.6 versus 16.1 percent are estimates that shift with units versus dollars.
- Share — the ~95 percent combined figure is an estimate, and the market definition varies.
- Company figures — Broadcom's $8.4bn and Marvell's up-to-$11bn are company-reported.
- Performance — independent verification of chip-level numbers is scarce.
- Development cost — the hundreds-of-millions, two-to-three-year rule of thumb varies widely.
Sources
- Tom's Hardware — The custom AI ASIC state of play: Broadcom deals, Google TPUs, Meta MTIA & beyond
- Hashrate Index — Inside the Custom AI Chip Race: Google, AWS, Microsoft, Meta, OpenAI
- TechTimes — Custom AI Chips Outpace Nvidia GPU Growth in 2026
- The Next Web — Google in talks with Marvell Technology to build new AI inference chips alongside Broadcom TPU programme
- Silicon Analysts — AI Data Center Value Chain: Every Layer from Chips to Cloud (2026)
- Oplexa — Custom ASIC Market 2026: Why Hyperscalers Are Ditching NVIDIA