Microsoft, Google, Amazon, and Meta are projected to spend up to $660 billion on AI capex in 2026 alone. That is not a barrier to entry. That is a closed border.
The phrase "barriers to entry" was coined for an economy in which a determined founder with a few million dollars and a clever idea could disrupt an incumbent. Frontier AI has detached from that economy. The numbers no longer parse as competition; they parse as state-craft.
The cost curve runs straight through the sovereign tier
Training GPT-4 cost around $78 million in compute. Google's Gemini Ultra came in near $191 million. Meta's Llama 3.1 405B was approximately $170 million. Each headline number is a snapshot of a runaway curve: frontier training costs have grown by 2 to 3x per year for eight years running. Anthropic's Dario Amodei has openly forecast $10 billion training runs by 2028.
A $10 billion training run is not a venture-backable check. It is the annual research budget of a mid-sized nation. The only entities that write that check today are the four hyperscalers, two or three sovereign wealth funds, and the AI labs they have already absorbed by equity stake or compute partnership.
Notice what this does to the conventional startup story. A founder cannot raise around this. A university cannot reach it. A national lab in a non-G7 country cannot match it. The "competitive frontier" is a country club whose dues are denominated in fab capacity, power-grid access, and decade-long Nvidia allocations.
Concentration is not one bottleneck. It is four, stacked.
The naive picture of AI concentration treats it as a single chokepoint. The reality, as documented in Yale Law and Policy Review's antimonopoly analysis of AI, is that the stack is severely concentrated at multiple layers simultaneously, and each layer reinforces the others.
At the silicon layer, Nvidia is functionally a monopoly on training-grade GPUs. At the cloud layer, three hyperscalers — AWS, Azure, Google Cloud — capture the lion's share of large-scale model training. At the model layer, a handful of labs hold the top of every leaderboard. At the application layer, those same labs are now bundling consumer agents directly into operating systems and search.
A startup that wants to compete with OpenAI must rent GPUs from one of OpenAI's largest investors, on cloud infrastructure built by another of OpenAI's largest investors, and ship into application surfaces those investors increasingly control. The "moat" is not a single trench. It is four trenches dug at right angles to each other.
The DeepSeek leak is real but it does not refill the moat
The optimist's rebuttal is that the moat is leaking. DeepSeek shipped a frontier-tier model in late 2024 for a reported $5.6 million in training compute, benchmarked against U.S. closed-frontier rivals that cost half a billion to a billion dollars to produce. Open-weight models like Llama, Mistral, Qwen, and DeepSeek now sit roughly a year behind closed frontier and the gap is shrinking.
That trajectory is real. It is also not the same thing as breaking concentration. Even if frontier capability becomes a commodity input, the substrate underneath it — chips, data centers, power, capital — does not. Whoever owns the substrate captures the rents. Cheaper models produced by anyone running on rented hyperscaler compute is a transfer of value to the hyperscaler, not from it.
This is the lesson of every previous infrastructure cycle. Cheaper trains did not break the railroads' grip on freight. Cheaper websites did not break Amazon's grip on retail. Cheaper models will not break the grip of whoever owns the GPU floor and the substation feeding it.
What "barrier to entry" means at sovereign scale
The political consequence is that the next decade's most important productive technology is being capitalized inside a structure that classical antitrust was not built to address. The DOJ's case against Google's search monopoly took roughly five years from filing to a remedies order. The EU's parallel cases have run longer than ten. By the time a remedies order lands on hyperscaler-AI integration, the next cycle's monopoly will be priced in.
So the operative question is not "who wins the race to AGI." It is more boring and more decisive: who owns the substations, the GPU allocations, and the multi-year cloud commitments. The answer, today, is four companies. Whether that list grows, shrinks, or gets nationalized somewhere is the political economy story of the rest of the decade.