You’ve felt it, haven’t you? That quiet unease when you read about another Chinese open-source model matching GPT-4’s performance at a fraction of the cost. That little voice saying: wait, weren’t we supposed to be years ahead?
Ben Thompson says don’t worry. US frontier labs still have structural advantages — compute, capital, talent pipelines. The gap is real, and it favors America. Relax.
He’s wrong. And the labs panicking? They’re the sanest people in the room.
The moat isn’t shrinking. The moat is becoming irrelevant.
Here’s what Thompson and many analysts miss: they’re playing chess while the game is being rewritten as Go. The panic isn’t about electricity costs or whether DeepSeek-V3 trained for $5.6 million versus OpenAI’s hundreds of millions. That’s a spreadsheet argument. The real argument is about second-order effects — the feedback loops that compound quietly until one morning the center of gravity has moved and nobody can explain how.
Think about what happens when a capable model goes open-source. Not the model itself — the ecosystem that crystallizes around it. Developers build on it. Researchers improve it. Competitors fork it. Every download is a node in a decentralized compute network that no single entity controls, regulates, or can choke off.
OpenAI’s advantage was supposed to be scale. More GPUs, more data, more parameters. But what happens when thousands of distributed actors, each with modest resources, collectively out-iterate one centralized lab? It’s the same story as every disruption: the incumbent optimizes the old game while the challenger invents a new one.
You don’t beat a fortress by building a bigger fortress. You make walls irrelevant.
And the talent loop? This is where it gets genuinely scary for US labs. When DeepSeek publishes papers and releases weights, they signal something powerful to every ambitious researcher in the world: come here, your work will be seen, used, built upon. OpenAI and Anthropic increasingly operate in secrecy — redacted model cards, gated access, NDAs stacked like fortifications. Secrecy feels like strength. It’s actually a talent repellent.
I’ve talked to researchers who’ve left frontier labs. The pattern is consistent: they didn’t leave for money. They left because they couldn’t publish, couldn’t share, couldn’t feel the impact of their work. The open-source ecosystem offered something the closed labs couldn’t — the dopamine of seeing your contribution matter in real time across thousands of projects.
Now layer in geopolitics. US export controls on chips were supposed to kneecap Chinese AI development. Instead, they did what sanctions always do: they forced innovation under constraint. DeepSeek’s engineering — the multi-token prediction, the auxiliary-loss-free load balancing, the training efficiency — isn’t despite the chip shortage. It’s because of it. Constraint bred efficiency. Efficiency bred capability. Capability bred confidence.
Export controls didn’t slow them down. Export controls taught them to need less.
Thompson’s core argument is that structural advantages compound. He’s right about the principle and wrong about the application. Structural advantages compound — until the paradigm shifts. IBM had structural advantages in mainframes. Nokia had structural advantages in hardware. The advantage doesn’t disappear. It just stops mattering.
The labs that are panicking understand this. They’re not panicking because they’re weak. They’re panicking because they can feel the ground shifting — the realization that winning the scaling race doesn’t mean much when someone else has changed the destination.
And here’s the twist nobody wants to say out loud: the open-source approach isn’t just catching up. In several measurable dimensions — inference cost, deployment flexibility, community-driven improvement speed — it’s already ahead. The closed labs still lead on raw capability. But the gap between ‘best model’ and ‘best enough model that’s free’ is narrowing faster than anyone projected.
When ‘good enough’ is free, ‘best’ has to justify its price tag every single day.
The question isn’t whether US labs will lose their lead. The question is whether ‘lead’ will mean anything when the race has no finish line and everyone’s running different distances.
So yes, panic. Panic is the correct response when you realize your strengths are optimized for a world that’s being replaced. Panic is what happens right before adaptation. The dangerous ones aren’t the labs that are afraid. The dangerous ones are the commentators who say everything’s fine.
Complacency has never won a war. And this — whether we admit it or not — is one.
FAQ
Q: But aren't US electricity costs actually lower, making Thompson partially right?
A: On a spreadsheet, yes. But electricity costs are a first-order argument. The panic is about second-order effects — ecosystem feedback loops, talent flows, and paradigm shifts. Winning on electricity costs while losing the open-source ecosystem is like winning on fuel efficiency while everyone else switches to electric.
Q: What should US labs actually do differently?
A: Stop treating openness as a vulnerability. The labs that publish, share, and build communities will attract the talent that creates the next breakthrough. Secrecy protects today's model at the cost of tomorrow's researcher. Also: invest in efficiency engineering, not just scale — because the constraint-driven innovation we're seeing from competitors is a preview of where the field is heading.
Q: Isn't this just hyping Chinese AI to drive engagement?
A: No — this is about recognizing that the competitive dynamics have fundamentally changed. DeepSeek is a specific example of a broader pattern: open-source, decentralized, efficiency-first AI development. Whether it's China, Europe, or a distributed global community, the threat to closed-lab dominance is structural, not geographic. Ignoring it because it's uncomfortable is how incumbents die.