The Pickaxe Layer

Every gold rush enriches the people selling shovels, and NVIDIA's graphics processing units have become the pickaxes of this one. By 2025 the company controlled somewhere between 80 and 92 percent of the market for AI accelerators — the chips that train and run large models — depending on whether you count by revenue or units (Silicon Analysts, 2025; CarbonCredits, 2025). Its data-centre revenue climbed from about $15 billion in 2022 to well over $100 billion in 2024, a curve so steep it looks like a measurement error.

Why is this layer so lopsided? Part of the answer is manufacturing: the most advanced chips depend on a single supplier, Taiwan's TSMC, and on a scarce packaging technology for which NVIDIA has secured priority allocation. But the deeper moat is software. Over nearly two decades NVIDIA built CUDA, the programming framework that researchers use to make its chips do useful work. An entire generation of AI engineers learned their craft on CUDA. A competitor can match NVIDIA's silicon and still lose, because the tools, tutorials, and habits of the field all assume NVIDIA hardware underneath. That is a network effect masquerading as a hardware business, and it is why even AMD — the strongest challenger — held only single-digit market share in 2025.

The chip layer is the one place where "winner-takes-almost-all" is simply the observed reality. It is also, tellingly, the layer where the largest customers are trying hardest to escape. Amazon, Google, Microsoft, and Meta are all designing their own custom chips, precisely because dependence on a single supplier at 80-percent-plus market share is a strategic vulnerability. Whether they succeed is one of the open questions of the decade.

Why Compute Builds Walls

One layer up, the cloud providers rent out access to those chips — and here the concentration, while lower, is still striking. Amazon Web Services, Microsoft Azure, and Google Cloud together controlled about 63 percent of the global cloud infrastructure market through 2025, with AWS near 30 percent, Azure in the low-to-mid twenties, and Google around 12 to 13 percent (Synergy Research Group, 2025). If you want to train a large model and you do not own a fleet of data centres, you are renting from one of these three — quite possibly a company that also competes with you.

The reason compute creates such high barriers comes down to a number that anchors this entire chapter. In 2017, a research team could train a transformer-class model — the architecture that started the current era — for a few thousand dollars, on hardware a well-funded university lab could buy. By 2024, a single frontier training run cost between roughly $78 million (OpenAI's GPT-4) and $191 million (Google's Gemini Ultra), with Meta's Llama 3.1 405B around $170 million (Stanford AI Index, 2025; Epoch AI, 2024). That is a leap of roughly four orders of magnitude — on the order of twenty-thousand-fold — in about seven years.

Where did that increase come from? Epoch AI, which tracks these figures more carefully than anyone, found that the amortised cost of the final training run for frontier models has grown about 2.4 times per year since 2016. Compounded, that rate alone multiplies costs many hundredfold over the period; the rest comes from the ballooning of everything around the training run — the research staff, the failed experiments, the specialised data pipelines. Crucially, the cost is not rising because chips got more expensive. It is rising because the models kept getting bigger, trained on more data with more compute, and each generation of capability demanded roughly an order of magnitude more of everything than the last. At Epoch's growth rate, the most expensive publicly announced model will cross the billion-dollar mark around 2027.

This is what economists mean by a barrier to entry that emerges from scale rather than sabotage. Nobody conspired to price out the universities. The frontier simply walked away from anyone without nine or ten figures to spend, and it kept walking.

The Flywheel

What makes AI especially prone to concentration is that economies of scale operate at every layer of the stack at once, and they reinforce one another. Training a larger model costs less per unit of capability. Running that model — inference, in the jargon — costs less per query when you serve millions of users rather than thousands, because the fixed cost of the infrastructure spreads across more requests. Gathering and cleaning data is cheaper when you already run a platform that generates behavioural signals by the billion. And distributing the result costs almost nothing when you already own the channels people use every day.

Stack those advantages and you get a flywheel. A large platform deploys AI to improve a service, which attracts more users, who generate more data and revenue, which funds larger infrastructure, which lowers marginal costs, which funds better models, which attract more users still. Each turn widens the gap between the biggest players and everyone else.

Last updated: 2026-07-10