It’s become fashionable to talk about “open” and “democratized” AI. But the reality is that training frontier LLMs like GPT-4 is structurally out of reach for all but a few hyperscalers.
The reason isn’t just talent or data. It’s mainly infrastructure.
Modern AI training requires tens of thousands of GPUs operating in perfect sync across high-speed interconnects. You can’t crowdsource this with a few thousand laptops and consumer hardware. Unlike protein folding simulations like Folding@home, LLM training isn’t loosely parallel. It’s tightly coupled, sensitive to lag, and intolerant of inconsistency.
That makes true democratization impossible under current training prerequisites.
And it creates real strategic risks:
🚩 A small group of companies controls the foundation models everyone else must build on.
🚩 Cloud and GPU providers define the playing field, setting access, pricing, and pace of innovation.
🚩 Open-source alternatives depend on pre-trained weights, keeping the core advantage centralized.
🚩 Hardware supply chain issues and bottlenecks, like NVIDIA, are already a problem and risk, and will get worse
This isn’t just a technical limitation, it’s a concentration of power. And it means AI is not a level playing field. It’s already a platform lock-in problem, even though it’s not a monopoly.
Executives exploring generative AI must look beyond models and applications. The real leverage is in compute, infrastructure, and long-term control over value creation.
Until the economics or architecture of training shift, the companies that control the hardware will continue to control the future.
💡It’s turtles all the way down, till you reach NVIDIA.