The $5 Trillion Bottleneck
On 29 October 2025, NVIDIA became the first company in history to reach a $5 trillion market capitalisation — less than eighteen months after crossing $3 trillion. The number is a referendum on a single proposition: that demand for AI compute will keep compounding for a decade.
The fundamentals behind it are real. Data centre revenue in fiscal 2025 was $115.2 billion of $130.5 billion total, up from $47.5 billion the year before. Gross margins held at roughly 75% — a level unknown in the chip industry, where 40–60% is considered excellent. NVIDIA FY2025 10-K
But scarcity cuts both ways. NVIDIA is scarce because its customers have no comparably full-stack alternative at scale. That is a moat — and it is also the entire market's single point of failure.
Customer Concentration: The Four Buyers
NVIDIA's own filings disclose that two direct customers each accounted for more than 10% of revenue in fiscal 2025. Because hyperscalers also buy through contract manufacturers, the disclosed figures understate concentration: press analyses of purchasing patterns conclude that Microsoft, Meta, Google, and Amazon together account for a majority of data-centre GPU sales. NVIDIA 10-K · Bloomberg
Four companies — all building their own custom silicon — hold a majority of NVIDIA's largest revenue line. If even one shifts a meaningful share of purchases in-house, the growth curve that justifies a $5 trillion valuation bends. NVIDIA's FY2025 10-K lists customer concentration as a risk factor in its own words.
The concentration is also the mechanism of the circular economy: NVIDIA invests in OpenAI, OpenAI buys cloud from Oracle, Oracle buys 400,000 GB200s from NVIDIA. The chips flow in a circle through four buyers.
Blackwell: The Machine Itself
The current generation is Blackwell — the B200 GPU and the GB200 Grace-Blackwell superchip, sold as the GB200 NVL72: a liquid-cooled rack of 72 GPUs and 36 Grace CPUs wired as one machine, at an estimated ~$3 million per rack. NVIDIA GTC disclosures · analyst estimates
- Scale: Oracle's disclosed order is 400,000 GB200s — roughly $40 billion of racks for a single customer. Financial Times
- Physics: NVL72 racks draw ~120kW and require liquid cooling, which is reshaping data-centre design and energy demand simultaneously.
- Lock-in: The moat is not the silicon alone — it is CUDA, twenty years of accumulated libraries, kernels, and developer muscle memory. Porting a frontier training stack off CUDA is measured in engineer-years.
The Alternatives
AMD Instinct
AMD is the only merchant competitor at scale. The MI300X/MI350 line was guided to roughly $5 billion in 2025 data-centre GPU revenue, and in October 2025 OpenAI signed a 6-gigawatt deployment deal with a 10% equity warrant — a contract large enough to be a strategic hedge against NVIDIA, and a signal that even NVIDIA's biggest customer wants a second supplier. AMD guidance · The Register, Oct 2025
Custom silicon (the real erosion)
- Google TPU — sixth-generation Trillium; Anthropic runs a meaningful share of its workload on TPUs inside a $200B Google cloud commitment. The Information, May 2026
- Amazon Trainium2 — Project Rainier, a Trainium2 cluster built for Anthropic, is among the largest AI supercomputers in operation. Amazon disclosures
- Meta MTIA — third-generation in-house accelerator for recommendation and inference workloads. Meta engineering blog
- Microsoft Maia 100 — first-generation Azure accelerator, deployed for Copilot-class inference. Microsoft Ignite
Every one of NVIDIA's four largest customers is building its own chip. None can yet train a frontier model end-to-end at Blackwell economics — but every generation narrows the gap for inference, where the long-run volume is.
One Foundry: TSMC
Every NVIDIA AI GPU, every AMD Instinct chip, and virtually every custom accelerator is fabricated by Taiwan Semiconductor Manufacturing Company, which holds over 90% of the leading-edge logic market. The constraint is not just the wafer fab but CoWoS — the advanced packaging that stacks GPU dies next to high-bandwidth memory. CoWoS capacity has been sold out forward for over two years. TSMC earnings · industry estimates
Geography is the risk. Over 90% of leading-edge capacity sits on an island that Beijing claims and that the US has committed to arm. TSMC's announced US buildout — raised to $165 billion in March 2025, covering Arizona fabs and packaging — is the largest industrial relocation in chip history, and it will still leave the leading edge predominantly in Taiwan for years. TSMC, Mar 2025
Export Controls: The Chip Curtain
Since October 2022, the United States has run an escalating export-control regime aimed at freezing China's frontier compute: the October 2022 rules, the tightened October 2023 rules, and the January 2025 "AI diffusion" framework tiering the world by chip-access. In April 2025 the H20 — the China-specific chip NVIDIA had engineered to comply with earlier rules — was itself banned, forcing a ~$5.5 billion charge. BIS rules · NVIDIA filings, Apr 2025
The controls have two effects that pull in opposite directions:
- They cap China's frontier training capacity — Chinese labs train on clusters of slower, less efficient domestic or pre-ban chips, at higher cost per token.
- They force a domestic Chinese supply chain — Huawei's Ascend 910C and successors, built on SMIC's ~7nm process, are imperfect but improving under guaranteed state demand. DeepSeek's R1 was trained substantially on constrained hardware, which is precisely why it shocked the market. See The Global Race
Export controls trade short-term American revenue (China was ~13% of NVIDIA data-centre revenue before the bans) for long-term containment. Whether containment holds depends on lithography — where China remains two or more generations behind — and on smuggling, which US enforcement actions show is ongoing.
The Chart: Who Owns the Accelerator Market
AI accelerator market share, 2025
Share of data-centre AI accelerator revenue. Analyst estimates — TPUs and other custom silicon are largely consumed internally, so figures are approximate. Partly substantiated
Read the chart two ways. The bar lengths show today's dominance. The story is the direction: every grey bar except Intel's is a customer-funded hedge against the gold one.
Can 75% Gross Margins Survive?
No semiconductor franchise has sustained 75%+ gross margins through a full competitive cycle. Intel in its monopoly era peaked near 65%. The bear case says NVIDIA's margins must compress as custom silicon takes inference, AMD takes share under warranty-backed deals, and four monopsony buyers negotiate as a bloc in all but name.
The bull case says this cycle is different: the full stack (CUDA, NVLink, networking, software) compounds faster than alternatives can close, and demand is supply-constrained rather than price-sensitive — buyers are queueing for allocation, not negotiating discounts.
Data-centre gross margin at the four hyperscalers' own silicon programmes, AMD data-centre revenue trajectory, and any NVIDIA quarter where revenue guidance misses while inventory rises. Margin erosion at NVIDIA would not be a chip story — it would be the repricing of the entire AI equity complex.
Confidence Assessment
| Claim | Status | Confidence |
|---|---|---|
| NVIDIA holds ~85%+ of AI training accelerators | Analyst estimates converge; no audited market census exists | Partly substantiated |
| Top four customers account for a majority of data-centre revenue | 10-K discloses two >10% direct customers; press purchasing analyses support majority | Substantiated |
| 75%+ gross margins are sustainable for a decade | No semiconductor precedent; competition and custom silicon arriving | Unresolved |
| Custom silicon will displace merchant GPUs at scale | TPU/Trainium share rising; CUDA moat intact for frontier training | Unresolved |
| TSMC concentration is a systemic risk | Documented single-source dependency; geopolitical risk priced by analysts, not resolved | Substantiated |
| Export controls permanently contain China's frontier compute | Capability gap narrowed (DeepSeek, Huawei Ascend); lithography gap persists | Unresolved |