The DeepSeek moment collapsed AI cost assumptions and redistributed competitive advantage overnight

In the early days of commercial aviation, aluminium was the material that defined what was possible. Every airframe, every structural calculation, every supplier relationship was built on the assumption that aluminium would always be the primary material and that its costs, processing requirements, and supply chains would set the boundaries of what aircraft could be built and at what price. At Boeing, I watched what happened when composite materials went from exotic and expensive to manufacturable and cost-competitive. It did not just change the material. It restructured the entire supplier network. Companies that had built decades of competitive position on aluminium expertise found that their advantage was not wrong, exactly. But it was suddenly less relevant than it had been the month before.

That is what a cost floor collapse looks like. But in aerospace, the shift took years. In AI, it is happening in weeks.

The infrastructure assumption

Until last month, the AI market operated on a shared belief that was so widely held it had stopped being examined. Building a frontier-level model required billions of dollars in compute infrastructure, massive datasets, and engineering teams that only a handful of organisations could afford to assemble. The cost of being competitive at the frontier was the moat. If you could not spend at that level, you were not a serious player. But nobody asked what would happen if the cost dropped.

NVIDIA's valuation reflected this assumption. The entire AI infrastructure trade reflected it. The venture capital flowing into AI startups reflected it. The competitive strategies of every company building on top of these models reflected it. The infrastructure assumption was the load-bearing wall of the AI market's competitive structure.

DeepSeek just removed the wall.

Their January release demonstrated frontier-level capabilities at a fraction of the cost that everyone assumed was the minimum. Not incrementally cheaper. Dramatically cheaper. The kind of cost reduction that does not adjust existing competitive positions. It scrambles them.

Every moat built on the assumption that this technology will always be expensive has a trapdoor. DeepSeek just opened it.

What actually just happened

The market reaction was immediate and severe. NVIDIA repriced. The broader AI infrastructure trade repriced. But the repricing in public markets is the least interesting consequence. The more significant repricing is happening inside every company that built its strategy on the infrastructure assumption.

I advise a startup founder who spent eighteen months building an AI product whose competitive moat was, fundamentally, a fine-tuned model. He had invested heavily in proprietary training data, custom fine-tuning pipelines, and the engineering expertise to run them. His pitch to customers was straightforward: our model is better at this specific task than anything you can access through a general-purpose API. That was true. But it was also, as of three weeks ago, a rapidly depreciating asset.

He called me on a Tuesday afternoon. Not panicking, but thinking out loud in that particular way people think when they realise a foundational assumption has shifted. His moat was not the model. But he had spent eighteen months acting as if it were. His actual moat had been there all along. His moat was the domain data he had collected from eighteen months of customer interactions, the workflow integrations he had built, and the switching costs his customers would face if they tried to replicate what he had built with a cheaper model. The model was the part he had treated as irreplaceable. It was the most replaceable part of the stack.

That realisation happened in a single afternoon.

The cost floor collapse

The cost floor collapse is not a new pattern. It has happened in solar panels, flat-screen televisions, genome sequencing, and, as I saw at Boeing, composite materials for aerospace. The pattern is consistent. An industry prices its competitive structure against a cost floor that everyone assumes is stable. A technological or methodological breakthrough drops the floor dramatically. The companies that built their position on the assumption that the floor was permanent find themselves exposed. The companies that built their position on something above the floor (domain expertise, distribution, customer relationships, workflow integration) discover that their advantages have actually increased in relative value.

But what makes the DeepSeek moment different from a slow cost decline is the speed. Solar panel costs dropped over a decade. Composite manufacturing costs declined over years. The AI cost floor collapsed in what felt like a single news cycle. Companies that were making strategic bets last month based on cost assumptions that no longer hold are now in the position of having to re-examine those bets in real time.

This is where it gets uncomfortable.

The infrastructure assumption was never a moat. It was a temporary height advantage on ground that was always going to flatten.

What this means right now

The immediate question for every company building with AI is simple: what part of your competitive advantage depends on the cost of the underlying technology staying high?

If the answer is "a significant part," you have a problem that did not exist four weeks ago. Not a theoretical future problem. A present one.

But here is the part that matters more. The companies that are in the strongest position right now are not the ones that spent the most on AI infrastructure. They are the ones that built the most value in the layers above the infrastructure. The data they collected. The workflows they integrated into. The customer relationships they built while the infrastructure was expensive and access was limited. The domain expertise they encoded into products, not into models.

The founder I advise is rewriting his investor narrative this week. But he is not rewriting his product. The product is fine. The customers are still using it. The workflows are still embedded. What changed is the story he tells about why his product is defensible. The old story was about the model. The new story is about everything around the model. It is a better story, honestly, because it is a truer one.

The redistribution

The most interesting consequence of the cost floor collapse is not who loses. It is how competitive advantage redistributes.

When a technology becomes cheap, the value migrates to the layers closest to the customer. That is the consistent pattern across every industry where this has happened. Cheap steel made distribution and brand more valuable, not less. Cheap cloud computing made product design and user experience more valuable, not less. Cheap AI infrastructure will do the same. But only for the companies that built those layers before the floor dropped.

But the redistribution is not automatic. It rewards the companies that already invested in those layers. It punishes the companies that assumed the infrastructure layer was where the value would permanently reside. The uncomfortable truth is that many of the most heavily funded AI companies of the past two years were built on the infrastructure assumption. They are the ones facing the hardest questions today.

Nobody knows yet how far this repricing goes. We are in the first weeks, and the full implications will take months to become clear. But one thing is already visible. The assumption that frontier AI would always require billions of dollars in capital expenditure, the assumption that made this market legible and this competitive structure stable, is no longer operative.

What remains when the cost advantage disappears is what was always underneath it: the quality of the problem you solve, the depth of your integration into the customer's work, and the trust you have built while everyone else was focused on the model. Those things were always the real moat. It just took a cost floor collapse to make that obvious.