Custom Silicon vs Nvidia: Can the Hyperscalers Break the GPU Monopoly?
Nvidia earns a 75% gross margin selling the chips that run AI, and every one of its biggest customers is now designing its own. Google sells TPUs to rivals, Amazon trains Anthropic's models on Trainium, OpenAI shipped its first chip this year and Broadcom's AI revenue is nearly tripling. This guide explains GPUs versus custom accelerators, profiles every challenger, follows the money through the supply chain, maps where Nvidia's software moat holds and where it is eroding, and lets you test when building your own chip pays off.
The 75% question
In the quarter to 26 July 2026, Nvidia reported revenue of $96.2 billion, up 106% on a year earlier. Its data center business alone brought in $89.0 billion, $48.7 billion of it from the big cloud companies. Its gross margin was 75%: for every dollar a hyperscaler spends on Nvidia hardware, roughly 75 cents is Nvidia's gross profit. It guided to $108 billion for the next quarter and says it expects about 70% revenue growth in its next fiscal year, limited mainly by how many chips it can get made.
That margin is the single biggest reason Google, Amazon, Microsoft, Meta and OpenAI all now design their own AI chips. If you are spending tens of billions of dollars a year on accelerators, capturing even part of that margin is worth a chip team of thousands. Jensen Huang's reply is that building AI infrastructure is "really hard" and most custom chip projects never reach production. Both things can be true.
GPU vs ASIC in plain words
A GPU is a general-purpose parallel processor: thousands of small cores that can run almost any maths, programmed through Nvidia's CUDA software. An ASIC (application-specific integrated circuit) is a chip designed for one kind of job. Google's TPU, Amazon's Trainium and Microsoft's Maia are ASICs built around the matrix multiplications that dominate neural networks. They give up some flexibility to spend more of their silicon, memory and power on exactly that work. Click a point on the spectrum.
The flexibility vs efficiency spectrum
Qualitative positions; real chips blur the lines (Nvidia GPUs contain specialised tensor cores, and TPUs are programmable).
Market share depends on how you count
Ask "how much of the market does Nvidia have?" and you get answers from about 62% to 85%. The gap is mostly method: Nvidia's chips are far more expensive than custom ones, so its share of revenue is much higher than its share of chips shipped.
Nvidia's share of AI accelerators, by estimate
Each bar is a different tracker or measure; none is authoritative. The unit figure comes from a low-authority tracker and is shown as a range midpoint.
The direction is clearer than the level. TrendForce forecast in October 2025 that AI servers built on cloud providers' custom chips would grow 44.6% in 2026, against 16.1% for GPU-based AI servers. Nvidia is still growing very fast in dollars; its share is what is slipping, slowly, at the edges.
The challengers
Filter by what each chip is mainly for. Specifications and performance claims are each company's own.
Custom AI chips in 2026
| Company ยท chip | What we know | Status |
|---|
Google's TPU story, 2015 to 2026
Google is the only hyperscaler with a decade of custom AI silicon behind it, and in 2026 it took a step none of the others has: selling TPUs to outside companies, including competitors. SemiAnalysis reported that of roughly a million TPUs for Anthropic, about 400,000 Ironwood chips would be sold as finished racks (worth about $10 billion) through Broadcom and about 600,000 rented through Google Cloud. The Information reported that Meta signed a multibillion-dollar lease of TPU capacity with an option to buy Ironwood outright from 2027. Bloomberg reported that Google DeepMind's own researchers are queuing for TPU time as a result.
TPU generations
Click a generation. Dates from Wikipedia's table and Google announcements as reported; v4's date differs between sources (2021 or 2022).
Follow the money: who designs, who builds, who makes
"Custom" rarely means a company does everything itself. Most hyperscalers design the architecture and hand the rest to a design partner, mainly Broadcom or Marvell, which turns it into a manufacturable chip, adds high-speed interfaces and manages production at TSMC. Broadcom and Marvell together are estimated to hold about 95% of this co-design market. Pick a chip to trace its chain.
The custom chip supply chain
Partner relationships as reported in the press; foundry and memory suppliers are typical for the generation, not confirmed for every chip.
AI revenue scale: Nvidia vs Broadcom
US$ billions per year. Nvidia's figure is our annualisation of one quarter's data center revenue (ร4); Broadcom's are management guidance and targets.
- Broadcom reported $8.4 billion of AI semiconductor revenue in its fiscal first quarter of 2026, up 106%. By its third quarter, custom accelerators ("XPUs") were 73% of AI revenue and shipments had more than tripled. It guides to about $58 billion of AI revenue for fiscal 2026, holds a $73 billion AI backlog and says it has "line of sight" to more than $100 billion of AI chip revenue in 2027.
- Its six custom-chip customers are Google, Meta, Anthropic, OpenAI and two it does not name. It is shipping Google's TPU 8i ahead of the MediaTek-built 8t.
- Marvell works with Amazon on Trainium and Microsoft on Maia; one source projects up to $11 billion of AI custom-chip revenue for it in 2026.
The CUDA moat, layer by layer
Nvidia's strongest defence is not the chip; it is two decades of software. But "CUDA lock-in" is not one thing. Click each layer of the stack to see where switching away is getting easier and where it is still hard.
Where lock-in lives
Green = eroding, amber = mixed, red = still strong. Summary of practitioner and analyst sources; mostly 2026 blogs, not primary research.
Build vs buy: when does a custom chip pay off?
The basic trade is a large fixed cost (designing the chip and its software) against a lower cost per chip (no Nvidia margin, though a design partner and foundry still take theirs). Whether it pays depends on volume and on how much useful work each custom chip does compared with a GPU.
Build vs buy calculator
All defaults are illustrative assumptions, not reported costs. "Manufacturing cost" is the cost to make one comparable accelerator; the GPU price is derived from it using Nvidia's margin. Ignores networking, software staff over time, power differences and timing.
Play with it and a pattern appears: at tens of thousands of chips, the fixed cost dominates and buying wins; at hundreds of thousands to millions, avoiding the margin dominates and building wins, if the custom chip does comparable work. That is exactly the scale at which Google, Amazon, Meta and OpenAI operate, and why smaller buyers still mostly buy Nvidia.
Why custom chips win inference first
| Training | Inference | |
|---|---|---|
| Workload | Changes with every research idea | A few known models served billions of times |
| Software | Hand-tuned kernels, deep CUDA dependence | Standard frameworks and serving engines; easier to port |
| Networking | Thousands of chips in lockstep (NCCL, NVLink) | Small groups of chips per request |
| What matters most | Time to result; flexibility | Cost per token and power, where specialisation pays |
| Share of compute | About one-third in 2026 (Deloitte estimate) | About two-thirds, and growing |
| Who is winning | Nvidia (Google TPUs the main exception) | Custom chips gaining: Maia 200, MTIA, TPU 8i, Trainium all lead with inference |
Google's 2026 split of its TPU line into a training chip (8t) and an inference chip (8i) makes the point in hardware. Meta says its MTIA chips have done better at inference than at generative-AI training, and Microsoft's Maia 200 is explicitly an inference accelerator.
Why custom chip programs fail
- Moving targets. Microsoft's next Maia chip (codenamed Braga) slipped at least six months to 2026 after, according to The Information, OpenAI requested new features late in development, which destabilised the design in simulation.
- People. The same reporting cited staffing shortages and turnover of up to 20% on some teams under deadline pressure. Chip teams are scarce and poached constantly.
- Architecture drift. A chip takes two to three years from design to deployment; model architectures change faster. A chip tuned for last year's models can miss this year's (mixture-of-experts, long context, new number formats).
- Training is harder than inference. Meta, by Reuters' account, has had more success with inference chips than with a generative-AI training chip.
- You still need partners. Custom does not escape Broadcom's or Marvell's margins, TSMC's capacity or the HBM shortage (see The Memory Wall).
- Software is the real cost. Compilers, kernels, debugging tools and framework support take as long as the silicon, and keep costing money every generation.
Nvidia strikes back
Nvidia's counter-moves
Three scenarios to 2030
๐ข Nvidia holds ~80% of revenue
Frontier training stays on Nvidia, new architectures keep outrunning ASICs, and Nvidia's yearly cadence and system integration (chips, networking, software) keep it ahead. Custom chips settle into each hyperscaler's own inference. Signal: Nvidia's gross margin stays above 70%.๐ก A 60/40 split
Inference, two-thirds of compute and rising, moves heavily to TPUs, Trainium, Maia, MTIA and OpenAI's chips; Nvidia keeps training and the long tail of customers. Analyst forecasts of 10โ25% custom-silicon share by 2028โ2030 sit at the low end of this. Signal: Broadcom hits its $100B+ 2027 target.๐ด Commoditisation
Framework-level portability makes accelerators interchangeable, Google sells TPUs widely, AMD's racks compete on price, and margins fall toward ordinary chip levels. Signal: frontier labs train flagship models off Nvidia; Nvidia margin below 60%.Reading the numbers: where sources disagree
| Claim | What the sources say |
|---|---|
| Nvidia's market share | About 73%, 83.5% or 85% of revenue depending on the tracker; about 62โ66% of units in one low-authority tracker. |
| Anthropic's TPU deal | "About 1 million TPUs" is consistent; dollar values and gigawatt figures vary (1 GW in 2026 and over 3 GW in 2027 per Broadcom's call; other reports cite 3.5 GW or 5 GW over five years). |
| Ironwood availability | Sources say general availability in late 2025 or April 2026. |
| Maia 200 transistor count | Over 100 billion in most reports; 140 billion in one. |
| Maia 200 vs Trainium3 and TPU | "3ร Trainium3 in FP4, beats TPU v7 in FP8" is Microsoft's own claim. |
| Trainium3 per-chip specs | 2.52 PFLOPS FP8 and 144 GB HBM3E appear in a secondary guide; AWS's own figures are system-level (4.4ร Trainium2 UltraServers). |
| ASICs "up to 70% better power efficiency" | An analyst claim that depends on volume and workload. |
| Unit tracker: ~1.9M custom chips in 2026 | ~900k TPU, 600k Trainium/Inferentia, 250k Maia, 180k MTIA from one low-authority tracker; not used in charts except the unit-share midpoint. |
| Marvell's $11B | A projection from a single secondary source. |
| Build-vs-buy calculator | All defaults are our illustrative assumptions. |
๐ References
- Company results and announcements
- Nvidia: results for Q2 fiscal 2027 ยท CNBC on the guidance
- Futurum: Broadcom Q1 FY2026 and XPUs ยท Broadcom Q3 FY2026 call summary
- DCD: Google's eighth-generation TPUs ยท ServeTheHome: TPU v8 at Hot Chips 2026
- TechCrunch: Trainium3 and an Nvidia-friendly roadmap
- Network World: Microsoft launches Maia 200
- VKTR: Meta expands its MTIA roadmap
- Maginative: OpenAI and Broadcom's 10 GW plan
- Analysis
- Tom's Hardware: the custom AI ASIC state of play (May 2026)
- SemiAnalysis: TPUv7, the 900lb gorilla
- SiliconANGLE: Google and Meta's TPU deal
- VentureBeat: the ASIC challenge to Nvidia
- TD Securities: reframing the GPU vs ASIC debate
- CUDA lock-in unpacked: the real switching costs
- TechSpot: Microsoft's custom AI chip hits delays
- Wikipedia: Tensor Processing Unit
- The Register: why Nvidia paid $20B for Groq's technology
- AI Compute Stack on this site
- Part 3: AI chips and compute infrastructure
- Part 4: The Memory Wall
- Part 2: Inside the AI Factory