Nathan Lambert exposes the folly of protecting closed AI labs
By Bram de Vries · Reporting from Amsterdam ·
On the a16z podcast Nathan Lambert showed open AI models trailing by months, proving that subsidizing closed labs through state regulation will only throttle innovation across the broader economy.
The average researcher at OpenAI can call upon 5,000 Nvidia H100 chip equivalents at any moment. When Chinese pre-training scientists hear that figure, their jaws hit the floor.
Yet despite that grotesque disparity in silicon, open-weight models trail the frontier by a mere two to three months. On the a16z podcast, Trillium founder Nathan Lambert laid bare the mechanics of modern artificial intelligence. Lambert argued that OpenAI and Anthropic rely on massive capital moats and proprietary compute to pioneer training paradigms. Yet open models from Chinese labs and commercial startups fast-follow by harvesting knowledge and refining open architectures. What the closed labs call intellectual theft through distillation, the market recognizes as ordinary trade.
The fallacy of the silicon cartel
Silicon Valley has convinced politicians that AI development requires national protectionism. The incumbents want the state to police terms-of-service violations as if they were border breaches. Lambert pointed out the absurdity of calling in regulators because a foreign firm probed an API. A terms-of-service clause is not an international treaty. If a commercial laboratory dislikes how clients use its endpoints, it can shut off their access. Demanding that lawmakers cripple open research to protect closed balance sheets is rent-seeking in its purest form.
The strongest case for closing off frontier models rests on national security and capital recovery. Advocates argue that frontier models demand scarce computing clusters and power bottlenecked until 2028. If fast followers can distill the outputs of those investments for pennies, the commercial flywheel breaks. Venture backers stop funding massive clusters, private research stalls, and hostile states capture the frontier without paying the toll.
History tells a very different story. Every cartel makes this exact plea when its initial monopoly begins to leak. In maritime trade, British shipwrights demanded state navigation acts to ban Dutch vessels from carrying colonial cargo. The protectionists claimed free seas would ruin domestic shipping capital. Instead, open competition lowered freight tariffs, forced technical innovation in hull design, and expanded global commerce. To claim that closed models will outpace open ecosystems forever, labs must assume knowledge never diffuses. That has never happened in industrial history.
Fast followers expand the pie
Open trade does not destroy capital; it disciplines it. When closed labs spent fortunes generating reasoning traces, they seeded an entire supply chain. Commercial data firms now sell RL environments across the open market. Startups like Cursor, Decagon, and Harvey prove that fine-tuned open models can match frontier behemoths at lower cost. In software coding, specialized open implementations already deliver equal execution without the crushing overhead of frontier billing.
This dynamic expands the economic pie. Serving costs are projected to fall to one percent of current levels within four years. Incumbents see falling margins as a crisis because they must justify private valuations before impending initial public offerings. But lower inference costs are a blessing for every enterprise that buys and integrates software. The economy does not exist to guarantee margins for closed frontier labs.
Technological leadership cannot be preserved by decree. When a laboratory holds an edge, it must monetize it through superior distribution, infrastructure efficiency, and operational speed. Attempting to freeze software weights behind regulatory moats will merely handicap domestic industry while foreign rivals press ahead. Keep the borders of software trade open. Let the fast followers run. Force the giants to earn their lead in the market every day.