This covers AI investor Sarah Guo's view that no single company will dominate AI. She sees computing power as the biggest bottleneck, with physical supply chain issues unlikely to resolve before 2030. She's bullish on Sunday Robotics (home robots by year-end), Harvey (legal AI handling 85% of M&A work), and Chai Discovery (biotech AI with million-dollar contracts). She notes open-source AI is unstoppable but requires safety testing.
Sarah Guo, Founder and Managing Partner of Conviction, shared her core perspective on AI frontier investing in this conversation: no single company can monopolize the future of AI. She believes that the best researchers in the AI field are facing a crisis of mission, while the entry of robots into h
Sarah Guo, founder and managing partner of Conviction, explores the core judgment of AI frontier investing in this conversation: No single company can monopolize the future of AI. She argues that the best AI researchers are facing a crisis of mission, while the entry of robots into homes and the true acceleration of scientific discovery are imminent. Key conclusions include: computing power remains the biggest bottleneck; open-source AI does not necessarily lead to the democratization of intelligence; and AI holds enormous opportunities in biology. Conviction focuses on investing in AI-native startups, typically as their earliest backer. Sarah is betting on small and medium-sized enterprises at the AI frontier rather than industry giants, and emphasizes that in a rapidly changing world, investors and founders need a "no backtest" mindset—making decisions based on conviction rather than historical data.
Sarah Guo argues that researchers at frontier AI labs are facing an unprecedented mission crisis, rooted in the fact that they no longer feel essential to the final outcome.
Historical Context and Mechanism Breakdown: Over the past 12 months, a new belief has emerged in the AI research community — that "through recursive self-improvement, models can enhance themselves, and we are only one or two years away from some form of exponential intelligence." Sarah points out that this belief has had a "disempowering" effect on researchers: "If the question is, I need $750 billion in compute spending, and we have tens of thousands of people working on this problem, I think people's sense of ownership over the outcome has diminished." (In other words: when researchers believe the sole determining factor is the scale of compute, the perceived value of individual effort is diluted.)
Data Chain and Magnitude: Sarah cites Andre Karpathy's self-reflection — "He thinks we are two years away from AGI, a judgment he has been making for about 10 years." The key change lies in the compute intensity and headcount at major labs (e.g., "OpenAI with 200 people"), which have significantly reduced the sense of individual contribution. Researchers face two choices: "Nothing I do matters because the model will do it itself" or "The only thing that matters is the scale of compute" — both are demoralizing.
Extrapolation and Falsification Conditions: If this trend persists, the best researchers may leave large labs to pursue entrepreneurship or academia. Falsification signal: when researchers once again feel that individual effort can change outcomes (e.g., achieving breakthrough progress in a small team).
Sarah Guo believes that compute power is the biggest bottleneck in current AI development, and that physical supply chain constraints are far more severe than those at the software level.
Data Chain & Timeline: Sarah cites a conversation with an infrastructure executive at a hyperscale cloud provider: "Nothing before 2030 can change the situation at sufficient scale." She points out that the U.S. needs adequate natural gas supply, but this is not a technology or capitalism issue—it is a "regulatory problem and a coordination problem"—"If you want to build a data center in New York, you need to convince New Yorkers that they should want a data center."
Mechanism Breakdown: The compute power supply chain involves multiple links—advanced process capacity (monopolized by TSMC), specialty glass ("basically controlled by one company"), cooling, and electricity. Sarah emphasizes: "The physical supply chain is a harsh reality. Learning how to build things, possessing tacit knowledge, labor, and raw materials—these cannot move as fast as software or even decision-making."
Extrapolation & Falsification Conditions: If the U.S. fails to resolve regulatory bottlenecks in energy and infrastructure, it may face the risk of "compute power non-independence." Sarah predicts that "we will start talking more about compute power independence." Falsification signals: substantial regulatory breakthroughs in U.S. nuclear energy (SMR) or natural gas power generation, or the commercialization of alternative chip architectures.
Sarah Guo argues that the proliferation of open-source AI is irreversible and crucial to the economic ecosystem, but security risks must be addressed head-on.
Data Chain & Current Status: Over the past three years, open-source models from China, the U.S., and Europe have become increasingly powerful ("Thinkie, Poolside, NVIDIA models, Mistral"). Sarah notes that even setting aside the economic interests of labs, "there are a large number of scenarios where using models from frontier providers is too expensive, too sensitive, or too slow"—and this trend will intensify as AI applications expand.
Mechanism Breakdown: Sarah distinguishes between two dimensions: "commercial health" and "security responsibility." On the commercial side, open source fosters capability diffusion, and enterprises need to control their own destiny (cost, capacity). On the security side, "if models can be used for defensive cybersecurity and biological work, they can obviously be used in similar ways for offensive use cases or biological weapons." She advocates for "testing and understanding frontier models" rather than trying to block technological progress.
Extrapolation & Falsification Conditions: Sarah believes the future where "intelligence is too cheap to meter" is approaching, but it is not inevitable—"I can easily imagine a world where we don't have that." Key variable: whether the U.S. can establish an effective security testing framework while remaining open. Falsification signals: verifiable backdoor behavior emerging in Chinese models, or the U.S. significantly restricting open source due to security concerns.
Sarah Guo believes the timeline for general-purpose semi-humanoid robots entering homes has shifted from "whether it is possible" to "when it will happen," and the Sunday Robotics team is advancing at an astonishing pace.
Historical Context and Mechanism Breakdown: Sarah describes her investment process in Sunday Robotics — founders Tony Zhao and Chang Chi, both around 25 years old and pursuing PhDs at Stanford, "have, over the past four years, almost single-handedly contributed the most interesting ideas in the field of robot AI." The core innovation lies in the data collection strategy: "How can we cleverly collect data in the cheapest way to support the distribution of real-world environments and tasks?" The team started with "a pile of cardboard in a Stanford basement" and completed hundreds of hardware-model-data collection iterations in less than two years.
Data Chain and Timeline: Sarah relays the team's assessment: "We will have general-purpose semi-humanoid robots doing things in people's homes before the end of this year, starting with a beta version." She emphasizes that "the entire team believes this" and notes that "the speed is incredible."
Extrapolation and Falsification Conditions: If successful, this will fundamentally reshape the labor landscape in areas such as home services, elderly care, and logistics. Falsification signals: unresolved generalization issues after actual deployment, or costs that cannot be reduced to a consumer-acceptable level.
Sarah Guo believes that AI models can create and capture significant value in the field of biology, a judgment that has shifted from "possible" to "strongly affirmative."
Data Chain and Evidence: Conviction was the first investor in Chai Discovery, a company that has already partnered with several "top ten pharmaceutical companies" to accelerate certain stages of the R&D process. Sarah notes that the conventional wisdom—"dumb software investors can't make money selling software to pharmaceutical companies"—has been upended by the model. Key evidence includes a $10 million contract and the fact that "the customer knows whether it has value."
Mechanism Breakdown: Sarah distinguishes between two constraints: "regulatory issues" and "the speed of the physical world." She argues that "safety is not something that can be easily overcome," but "we should see a significant acceleration in cures." The tipping point will be "when we have a new indication or new drug created by AI, whose development trajectory is clearly altered."
Extrapolation and Falsification Conditions: Sarah predicts that "we will see a massive wave of investment." Falsification signals: AI-assisted drug discovery shows no significant difference in clinical trial failure rates compared to traditional methods, or regulators adopt a conservative stance toward AI-generated data.
| Position | Guest Stance | Key Data |
|---|---|---|
| Sunday Robotics | Bullish (Invested) | Founder is 25 years old; completed hundreds of iterations in under 2 years; home beta version expected by end of this year |
| Harvey | Bullish (Invested) | Legal AI; leap from "reviewing California lease agreements" to "completing 85% of the Activision Blizzard M&A work" |
| Chai Discovery | Bullish (Invested) | Collaborating with multiple top-10 pharmaceutical companies; $10 million contract |
| Suno | Not Invested (Previously Declined) | Music generation; Sarah admits "underestimated the demand for expression and entertainment" |
| Sigma | Holding for Observation | No specific action disclosed |
| Notion | Holding for Observation | No specific action disclosed |
| Rippling | Holding for Observation | No specific action disclosed |
1. "Investing without backtesting" is the core challenge of the AI era (Sarah Guo)
Investors cannot rely on historical data to make decisions and must instead base judgments on first principles of technology and people. Sarah quotes an investor friend's metaphor: "I keep saying I want to slam on the brakes as hard as I can, but I'm not doing it while driving at 90 miles per hour."
2. The best researchers are experiencing a "crisis of mission" (Sarah Guo)
When researchers believe the sole determining factor is compute scale, the perceived value of individual effort is diluted. This could drive the best talent away from large labs.
3. "Nothing that can change the game before 2030" (Sarah Guo, citing an infrastructure head at a hyperscale cloud provider)
The compute bottleneck is a physical supply chain issue, not a technology or capitalism problem. The U.S. needs to solve regulatory and coordination issues, not a lack of technical capability.
4. The "cat is out of the bag" for open-source AI (Sarah Guo)
Attempting to restrict U.S. open-source models will only constrain law-abiding companies, not affect real attackers. The right approach is to "test and understand," not to block.
5. A future where "intelligence is too cheap to meter" is not inevitable (Sarah Guo)
If anti-capitalist sentiment hinders energy and infrastructure construction, the U.S. could lose its competitive edge. She predicts "compute independence" will become a core issue.
6. Sunday Robotics' two founders "single-handedly contributed the most interesting ideas in robot AI over the past four years" (Sarah Guo)
The core innovation lies in a methodology for collecting data at the lowest cost, rather than simply pursuing model scale. The team started from "a pile of cardboard in a Stanford basement."
7. Value creation from AI in biology has shifted from "possible" to "strongly affirmative" (Sarah Guo)
The conventional view held that selling software to pharmaceutical companies could not generate profits, but models have changed the equation. The key evidence is customers' actual adoption and willingness to pay.
8. "If you find the truth, and it is mispriced, you are in a favorable position" (Sarah Guo)
Sarah's investment philosophy is "asymmetric information + conviction in holding a view." She believes most investors spend too much time thinking about strategic frameworks like "which layer will win" and too little time thinking about "what the next 99% of diffusion will look like."