Sarah Guo on Conviction's AI Investment Strategy

Sarah Guo keeps a headcount for the entire AI revolution: about two hundred and fifty people (07:41). Not companies, not labs, not billions of dollars of compute, but a specific number of researchers and entrepreneurs that she and her partner Mike consider the ones "doing the most interesting things on the frontier" (07:41). It is an old-fashioned way to think about a technology usually described in the language of scale: parameters, chips, gigawatts. Guo, who runs the four-year-old venture firm Conviction, is betting that individual people still bend outcomes, even inside a system that increasingly runs on compute scale alone.

That bet sits uneasily next to how frenzied the moment feels. Guo describes the mood as something closer to a car that will not slow down. "I keep saying I wanna press the brakes as hard as I can, but I'm not doing it, going ninety miles an hour," she says, relaying a friend's line about the market (03:11). The tension she is naming is specific to venture investing without a back test: nobody has lived through this exact cycle before, so nobody can point to precedent to know whether restraint or speed is the mistake. She calls her worldview a version of the great man and great woman theory of history: put highly capable people in the right place with the right support, and they change what happens next. It is why she resists the framing of an AI arms race between a handful of labs, even while she agrees the extreme version of that outcome, one or two frontier model owners capturing the whole economy, is a future she actively does not want.

Inside that community of 250, something has shifted in the last year. Researchers increasingly believe that recursive self-improvement, meaning AI systems that can improve the models building them, could deliver a burst of "exponential intelligence" within one to two years (11:49). Guo is careful to flag the caveat inside her own claim: the AI researcher Andrej Karpathy has reportedly said "two years away" for roughly a decade running, and may be saying it again now. What has actually changed is the psychology underneath the belief. When frontier labs employ thousands of people and require an estimated $750 billion in compute spend to keep scaling (12:26), an individual researcher can no longer feel like their personal effort moves the outcome. "The only thing that matters is compute scale, and both of those are somewhat disempowering," Guo says of the two beliefs now circulating among top researchers: that the model will simply do the work regardless of them, or that only capital and hardware decide the pace (12:50).

The Bottleneck Is Not the Algorithm

If compute is the new scarce resource, the constraint on getting more of it is not technical. Guo says the honest answer, delivered to her by an infrastructure leader at a major hyperscaler, is that nothing will meaningfully move capacity at scale before 2030 (14:21). Planning horizons inside the industry have already shifted out to 2032 (13:53). The limiting factor, in her telling, is regulatory and political: convincing communities to accept new data centers, and convincing the public that nuclear power, including small modular reactors, is safe enough to build at the pace the cost curve requires. She does not think America lacks the entrepreneurial or technical capacity to build cheap, abundant energy. She thinks it lacks permission.

That same logic extends to what Guo calls compute independence, a phrase she expects to become as central to policy debate as energy independence once was (41:01). The physical supply chain behind a data center, chips, cooling, power, is thin in places that are neither stable nor fully accessible to the United States and its allies. She points to a niche example: a specific kind of glass used in chip manufacturing that is effectively controlled by one company, with Taiwan Semiconductor Manufacturing Company (TSMC) holding a near monopoly on its supply (41:18). Efforts like Jacob Helberg's initiative, Pact Silica, are trying to map every link in that chain and find independent alternatives. Guo's own portfolio reflects the same theory: investments in the labor gap for data centers and robotics, in nuclear energy, and in alternative chip architectures, rather than in the flashier, more visible application layer alone.

The clearest evidence that individual insight still outruns brute infrastructure, for Guo, is a robotics company called Sunday Robotics, founded by two Stanford PhD students, Tony Zhou and Chang Xi, who had also worked at Toyota Research, DeepMind, and Tesla before starting the company in their mid-twenties. The prevailing assumption in robotics was that generalization required internet-scale data, data nobody actually has. Zhou and Xi treated that as a solvable engineering problem instead: find the cheapest possible way to collect data that still captures the real distribution of home environments and tasks, and figure out how that data transfers into model learning (20:39). Guo says her firm decided to invest in the company's first meeting. The team's own conviction is blunt: "Nothing is true until it is shipped," and they believe general semi-humanoid robots could be handling tasks in people's homes in beta by the end of this year (21:38).

Betting Before the Story Is Clean

Guo's own decision process mirrors that same comfort with incomplete information. In a first meeting, she rates a founder on a one-to-ten scale almost instinctively, often landing at an eight or nine before she fully understands the technology (22:42). What follows is not confirmation bias so much as an audit: days or weeks spent finding the holes in her own logic, writing a full investment memo (a practice she kept even when Conviction was a firm of one), and sending it to an outside investor for a second, more skeptical read (23:40). She worries that much of the current market has stopped doing this work. "There's a lot of proxying of judgment to pedigree or to other legible signals," she says, describing a debate with a fellow investor who could not explain a large bet except by citing the founder's track record (16:45). Not having a point of view on the business itself, she argues, beyond who the person is, is dangerous when capital is chasing unproven technical bets at scale.

Her language for the discipline underneath all this is conviction, the firm's namesake: finding a truth the market has mispriced and holding it until it is proven right or wrong.

"If you find the truth and it is wrongly priced and you hold onto that, you're in a good position." — Sarah Guo (50:24)

She applies the same test to biology. The conventional model in biotech was that the only way to make money was to build an actual drug, then license pieces of it away in staged "biobucks" deals, with venture investors often taking around 40 percent of a company built around a handful of academic founders (46:14). Guo spent five years watching computational biology companies under that model before concluding that AI changes the calculus enough to sell software directly to pharmaceutical companies, rather than only selling drugs. Her firm's evidence is Chai Discovery, now working with several top-ten pharma companies on parts of the drug discovery process, including at least one contract worth $10 million (47:19). She is candid that Conviction invested before knowing this would work: "We invested when we didn't know yet. That's part of the fun of venture," she says, adding that the real proof will arrive when a specific new drug's development timeline is clearly and publicly attributed to AI.

Guo's most personal argument concerns what happens to human effort once AI absorbs the mundane parts of work. She invokes Jevons paradox, the historical pattern in which making a resource more efficient to use increases total consumption of it rather than reducing it, and predicts the same will happen with labor (53:58). She already sees it inside her portfolio: one company's entire marketing function is now run by a single person, augmented by an autonomous system built to create leverage rather than reduce headcount. Asked whether she personally works less now that AI makes her more productive, her answer is immediate: she works more. It is a small, honest contradiction to end on. Guo's entire framework rests on the idea that individual people, not infrastructure, decide outcomes. Jevons paradox suggests that even as the machines get better, the people behind them will not get any less busy. Betting on greatness, in her telling, was never a way out of the work. It was just a way of choosing which work is worth doing.