You can feel it in the air, even if nobody in San Francisco wants to say it out loud. The ground is shifting under the feet of every frontier AI lab, and the tremor isn’t coming from a rival American startup. It’s coming from China, it’s coming in the form of open weights, and it’s coming at prices so low they look like a typo.
DeepSeek just dropped official V4 Pro GA benchmarks, and the numbers are, frankly, uncomfortable for anyone who just raised a billion dollars to train the next great model. We’re talking frontier-level performance — the kind that used to require entire data centers and the GDP of a small nation — at inference costs that make you double-check the decimal place.
The era where owning the smartest model was a moat is over. The moat has moved, and most people are still staring at the empty hole where it used to be.
Here’s what everyone is getting wrong about this moment. The tech press is obsessed with benchmark scores. Can DeepSeek match GPT-class reasoning? Does it hallucinate less? Can it code? These are the wrong questions. The right question is: what happens to your business model when the underlying intelligence becomes effectively free?
Because that’s what’s happening. Not “getting cheaper.” Not “trending downward.” Free, in the way that bandwidth became free, in the way that storage became free, in the way that every foundational commodity eventually becomes free once enough players crack the engineering.
Multiple open-weight Chinese models have landed in recent months, each one nipping at the heels of frontier performance, each one priced like a fire sale. DeepSeek’s pricing is so aggressively low that you almost feel bad for them — almost. Yes, it will go up. That’s not the point. The point is that the floor has been set, and it’s set at a number that makes the trillion-dollar capex narratives look like fantasy.
When a Chinese lab can match your flagship model for 1/20th the inference cost, you don’t have a technology problem. You have a business model problem.
Think about what the frontier labs have been selling. Not just intelligence — the promise of exclusive intelligence. The best brains, the biggest clusters, the most compute, locked behind an API that you pay for by the token. That entire value proposition collapses the moment someone gives away a model that’s 90% as good for nothing. And that’s exactly what’s happening, again and again, faster than anyone predicted.
Now here’s where it gets interesting, and where the real money will be made. If the model layer is commoditizing — and it is, violently — then value doesn’t disappear. It migrates. It moves up the stack to whoever owns the relationship with the user, the workflow, the data, the distribution.
The next trillion-dollar AI company won’t be the one that trains the best model. It’ll be the one that figures out what to do with a model that costs nothing.
If you’re building in AI right now, this should simultaneously terrify and electrify you. Terrify, because if your entire strategy is “we wrap GPT-5 and resell it,” you’re building on sand. The ground underneath you is liquefying. Electrify, because the cost of building genuinely intelligent products just plummeted, and it’s going to keep plummeting. The constraints that kept you out of the game — compute costs, API bills, model access — are dissolving in real time.
The frontier labs know this. You can see it in the frantic pivots, the enterprise deal pushes, the platform plays. They’re trying to become the distribution layer because they can feel the model layer slipping away. Whether they can make that transition fast enough is the only question that matters for their survival.
But for the rest of us — the builders, the developers, the startups — the message is clear. Stop obsessing over which model is on top this week. That leaderboard is going to look completely different next month, and the month after that, and the month after that. The models are converging. The prices are diverging — downward, fast.
Build for a world where intelligence is free and distribution is everything, because that world is arriving faster than anyone planned for.
The frontier labs had a good run. They did the hard, expensive, unglamorous work of pushing the frontier. But the frontier is catching up to them, and it’s bringing a receipt that’s 95% smaller. The question isn’t whether the incumbents survive this transition. The question is whether you’re positioned to profit from the wreckage.
FAQ
Q: But won't DeepSeek raise prices once they gain market share?
A: Yes, prices will rise from current fire-sale levels. But the floor has been set, and it's set dramatically lower than what frontier labs need to justify their capex. The pricing genie doesn't go back in the bottle once multiple players prove the cost can be this low.
Q: What does this mean for startups building AI products?
A: Stop building thin wrappers around a single model API. Your moat was never the model — it was always distribution, data, and workflow integration. Now that model costs are collapsing, double down on owning the user relationship and the application layer. That's where value is migrating.
Q: Are you saying frontier labs like OpenAI and Anthropic are actually doomed?
A: Not doomed — disrupted. They can survive if they successfully transition to owning distribution and enterprise relationships. But their core value proposition of 'we have the smartest exclusive model' is eroding faster than their capex cycles can justify. The model layer is commoditizing; whether they can pivot fast enough is the real question.