a16z Podcast: Why Open Models Could Still Win the AI Race | Nathan Lambert

SPEAKER_1 (0:01)

When AI model progress speeds up, everyone focuses on smarter software. But Nathan Lambert has a different take on the a16z Podcast. He says the real story is the race for three forms of capital: compute, data, and talent. Those are what decide who moves fastest. Right, and that explains why OpenAI, Anthropic, and Google have led the frontier.

Lambert argues these frontier labs can take the biggest bets. OpenAI helped ignite the reasoning model wave, while Anthropic focused on specific capabilities and found strong results in coding. He puts it pretty bluntly: "The frontier labs take the most ambitious bets to achieve breakthroughs." That willingness to spend heavily is part of what keeps them ahead. What's interesting is his take on Chinese labs. He describes them as technically excellent but compute-starved. That constraint actually pushes them toward open models as a practical business strategy. Not a charity play, in other words. As Lambert says, "They are not necessarily motivated by altruism, but by a pragmatic need to catch up." Open release helps them build brand, enter the conversation, and compete where global API sales are harder. That brings us to distillation, which is a major theme. Lambert says the industry is already built on competition, and attempts to stop distillation through policy are unlikely to work. The frontier model flywheel depends on training, revenue, and more compute, and fast followers can weaken that cycle.

He warns, "Anything that threatens this is perceived as a threat to the entire industry."

Building on that point, he compares open and closed systems. Closed labs are highly integrated and efficient, but the open model economy is fragmented. It takes longer to build, yet it gives companies real ownership of weights, the ability to run on their own hardware, and more control over cost and privacy. That's a genuine trade-off. Lambert also draws a useful line between domains. Math and code benefit from scaling and general reasoning, but biology, healthcare, and legal work depend on high-quality datasets and carefully designed environments. As he puts it, "The reality is that it comes down to data."

So in some cases, the best results come from buying or building the right data, not from magic generalization. He even points to Cursor research as evidence that smaller models can be better for long-horizon work on cost. Though he does warn the current RL boom is messy, with reward hacking and low-quality environments everywhere.

Still, he sees real momentum in synthetic data, open research infrastructure, and transparent work like Trillium and Open RSI. His closing message is pragmatic. The frontier will keep advancing, but not by secret formulas alone. Techniques diffuse, competition raises the floor, and specialized applications still have room to grow.

Exactly. As Lambert says, "There is room for many winners in the AI ecosystem." AI is not headed toward one winner.