---

PURE SIGNAL February 20, 2026

Here's a question worth sitting with: if the most consequential AI race isn't about algorithms or talent—but about raw silicon—what does that mean for who wins?

The Hardware Gap: Why Export Controls Are Working Better Than Anyone Admits

Let's start with a finding that cuts against the conventional narrative. There's been growing pressure to loosen U.S. export controls on AI chips to China, partly on the argument that Huawei is becoming a credible Nvidia competitor anyway—so why bother restricting exports?

A detailed analysis of publicly available chip roadmaps says the opposite is true. The performance gap between U.S. and Chinese AI chips isn't narrowing. It's accelerating.

Right now, Nvidia's best chips are roughly five times more powerful than Huawei's best offerings. By twenty-twenty-seven, that gap widens to seventeen times. Here's the striking part—Huawei's own public roadmap shows their next-generation chip, due in twenty-twenty-six, will actually be less powerful than their current best. That's regression, not progress.

Why is this happening? SMIC—China's leading chip manufacturer—is stuck at seven nanometer process technology. That ceiling exists precisely because export controls block access to the advanced lithography equipment needed to go further. Huawei can't buy its way past physics without the right tools to fabricate at smaller scales.

Huawei's fallback strategy has been quantity over quality—flood the market with more chips even if each one underperforms. That's also failing. Even under the most aggressive assumptions—eight hundred thousand chips in twenty-twenty-five, scaling to four million by twenty-twenty-seven—Huawei would still represent only about two percent of Nvidia's aggregate computing output by twenty-twenty-seven. A hundredfold production increase still wouldn't bring them to half of Nvidia's capacity.

The implication is stark. Huawei isn't evidence that controls are failing. It's evidence they're working.

Now flip the scenario. If the U.S. exports three million H200 chips—Nvidia's most powerful chip currently approved for export—to China in twenty-twenty-six, that single decision would hand China more AI computing power than it could build domestically until twenty-twenty-eight or twenty-twenty-nine at the earliest. That's not a marginal shift. That's potentially enabling some of the world's largest AI data centers to be built under Chinese control.

The underlying logic matters here. In almost every other dimension of the AI stack—data, research talent, algorithmic innovation, software—China either matches or exceeds the United States. The hardware bottleneck is the one structural advantage the U.S. holds. And compute requirements for frontier AI training have been doubling roughly every six months since twenty-eighteen. The country that controls the most advanced compute, compounds its lead exponentially.

AGI Timelines and the Persistent Disagreement

Zooming out to the bigger picture, Yann LeCun—now at AMI Labs after his long tenure as Meta's chief AI scientist—continues to push back hard on near-term AGI timelines. Speaking at a recent AI summit, he reiterated his view that artificial general intelligence remains years away, not months.

LeCun's position here is worth taking seriously, not dismissing. His core argument has always been that current large language models—systems trained to predict text—are missing fundamental architectural ingredients for genuine world understanding. Things like persistent memory, causal reasoning grounded in physical reality, and the ability to plan hierarchically over long horizons.

This isn't a fringe view. It's a principled disagreement about what intelligence actually requires—and it shapes how you think about research priorities, safety timelines, and where to place your bets.

The Thread Connecting Both Stories

Here's what ties these themes together. Whether AGI is two years away or twenty, the path runs through compute. LeCun's skepticism about current architectures and the hardware analysis both point to the same underlying truth—we are in a regime where raw computational scale is the primary lever being pulled.

That makes the chip question not just a geopolitical issue, but a deeply technical one. Whoever controls the most advanced fabrication—and the most advanced chips running on it—is running the experiment at the frontier. Everyone else is catching up to yesterday's results.

The controls aren't just working. They may be the most consequential AI policy decision of the decade.

---