Most people think of AI as software — something you write, test, and ship. The reality of training a frontier AI model looks nothing like that. It looks more like building a power plant.
GPT-4, according to various estimates, cost over $100 million to train. Google’s Gemini Ultra likely cost more. The compute required runs on thousands of specialized chips running continuously for months. The electricity bill alone is staggering.
This is not a detail. It is the most important structural fact about the AI industry right now.
Only a Few Players Can Compete at the Frontier
When training a single model costs nine figures, the number of organizations that can participate drops to almost zero. A handful of American companies, one or two Chinese ones, and a few well-funded European labs. Everyone else is building on top of what those organizations produce.
This has profound implications for who controls the direction of AI development, what values get embedded into the most powerful models, and who benefits from the economic returns.
The open source movement — led by Meta’s Llama releases and others — is a partial counterweight to this concentration. But even open source frontier models require enormous compute to train. You can run them cheaply once they exist. Building them from scratch is a different matter.
The Compute Bottleneck Is Real
NVIDIA’s H100 chips — the primary tool for training large models — have been in short supply since 2023. The waiting list for cloud compute has pushed smaller AI labs to delay training runs by months. Some have changed their research direction entirely based on what compute they could access.
This is why the geopolitics of semiconductors matters so much. Export controls on advanced chips to China are not just trade policy — they are decisions about who gets to build the most capable AI systems.
What This Means for Everyone Else
For businesses, it means the AI tools you use are built on infrastructure controlled by a small number of companies. That concentration creates dependency risks. If your entire workflow runs through one AI provider and that provider changes its pricing, policies, or access terms, you have very little leverage.
Diversification across AI providers — and investment in understanding what you are actually dependent on — is not paranoia. It is basic risk management.
For individuals, it means the free or cheap AI tools available today are subsidized by venture capital and strategic positioning. The economics of AI access will change as companies move toward profitability. Plan accordingly.
The cost of AI is not just a business story. It is a story about power, access, and who gets to shape the most consequential technology of our era.