Sarah Guo - The Power of Conviction - [Invest Like the Best, EP.383]
What this covers
Patrick interviews Sarah Guo, who left nine years at Greylock in 2022 to found Conviction, an early-stage firm built for AI companies. The conversation starts with recruiting and the traits of good early-stage investors, then lessons from early deals and the decision to leave an established firm. From there it moves up and down the AI stack. Guo describes incubation as useful where the context for automating a domain sits far from researchers who understand model capability, such as enterprise software configuration and maintenance, which consultancies in the U.S. and India currently handle at high cost. She also rejects the GPT wrapper dismissal, arguing that workflows, customer relationships, and change management make up most of the work.
The infrastructure discussion is concrete. She explains why challenger chips struggle against CUDA, noting that failures in large training runs surface only at multi-thousand-node scale. She cites NVIDIA's roughly 90 percent margins and argues that next-generation clusters are close to single-use capital bets. She also compares today's AI infrastructure with the Infrastructure Maturity Ladder of web hosting, and says pretraining something interesting now costs over $0.5 billion, though later entrants get a steep discount.
The more contested material comes last. Guo asks whether reasoning generalizes or improves only in verifiable domains like math and code, and how to evaluate models that already beat expert benchmarks. She calls speculative existential risk a distraction next to fraud and misinformation. She warns that enterprise adoption in banking is marginal, so an air pocket in spending is possible. She is bullish on health care administration and government services if AI makes their data work about 100 times cheaper.
Guo argues that AI value will accrue to application companies that design around model limits to reach 'minimum viable quality', that narrative-driven thinking about incumbents and commoditization is dangerous, and that efficiency and distribution, not raw model ownership, will determine winners.
- Application companies win by crossing a use-case-specific quality threshold through system design, verification, and workflow wrapping rather than waiting for better base models
- Incumbents like foundation-model labs will not win every market because secondary products get deprioritized and distance from end users is large
- Efficiency and cost reduction will matter more as compute and data constraints bind, and most new foundation-model entrants face prohibitive entry costs
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Efficiency has never stopped mattering in computing because someone always pays for it; while a model company can absorb inefficiency for a long time with a strong other business, in a vacuum more efficiency improves latency, cost, and scalability.
“Nowhere in the history of computing has efficiency not mattered because somebody is paying for it in the end”
NVIDIA's dominance is protected by CUDA, which has been optimized for more than 15 years, and new chip challengers face very high barriers because chip failures in large-scale training only surface at multi-thousand node scale, so simulation and small tests cannot confirm a chip works.
“people understand the power of CUDA and the optimization and effort that's been going into that for more than 1.5 decades.”
Models that are better calibrated about whether their answers are correct, and that can tell when they have a good answer, would become much more useful, because hallucination management is the blocker for many enterprise use cases.
“Hallucination management is the blocker for many, many use cases in the enterprise”
NVIDIA's margins are about 90 percent, and pushing it with more competition would make compute cheaper, though NVIDIA is innovating quickly; she expects the industry to improve hardware utilization and efficiency at every layer over the next five to ten years.
“pushing NVIDIA and then eventually challenging a 90% margin product”
The popular view that all AI application companies are thin wrappers ignores that the workflows, customer relationships, proprietary data, change management, and distribution built on top of foundation models are most of the work, which Guo calls the '99-mile' problem.
“that last mile, that looks like 99 of the 100 miles to me”
Although frontier model training is getting more expensive, the number of people who know how to train large models is rising and the second-comer discount is high, since later entrants can use known expertise and modern hardware and skip training intermediate models.
“the second-comer discount is very high as that expertise is known”
An open question in AI is whether multistep reasoning improves in a general way or only by training on math and code, using verifiable domains such as Lean proofs; critics argue that approach is a dead end because qualitative reasoning like investment judgment has no formal proof, so richer human expert data and data quality are also needed.
“that, to some degree, a dead end because we cannot do this with the rest of -- think about the qualitative reasoning that you do on an investment, there's no proof.”
Clusters required for next-generation training are essentially single-use, gigantic capital investments that are not easily extensible to the next generation, so the strategy of building ever-larger data centers for better performance has limits.
“for training, essentially like single-use gigantic constructions of CapEx. They're not easily extensible to the next gen.”
Speculative existential risks such as bio risk or runaway optimizers are a distraction compared with near-term abuses such as misinformation and fraud, which already exist with older technologies and are amplified and made much cheaper by current models.
“I actually think they're like a huge distraction from the near-term abuses.”
Most people should not be training general foundation models because entry now costs more than $0.5 billion to do pretraining of something interesting, making a general language model startup an easy way to lose money unless it is highly differentiated.
“when the entry price is now more than $0.5 billion to do pretraining of something interesting or at least get into the game of a general language model, that's a really easy way to lose money.”
Despite heavy capital expenditure on AI model development, enterprise adoption in risk-sensitive sectors such as large financial institutions is marginal because the risk, compliance, and reliability bar has not been met, so an air pocket in spending is a real possibility.
“The idea that we're going to hit an air pocket because adoption of these tools in the enterprise where there's real economic value is lagging, I think that is real.”
Guo is not worried that application companies' margins are structurally worse in the AI era, because infrastructure costs should fall as the ecosystem invests in efficiency, and it is not clear this is a structural problem compared with software.
“I am not worried that the end margin structure of any application company is bad or any worse in this era.”
Models are surpassing human benchmarks such as MMLU and legal or medical exams, so evaluation becomes difficult because human evaluation fails once models outperform experts, requiring new evaluation methods such as ARC or measuring human-plus-model performance.
“We are surpassing the benchmarks in a bunch of different ways where it begins to open the question of, how do we do evaluation of these models”
As next-generation models hit limits of data center size and power, efficient use of resources in both training and inference will become a much bigger consideration, and the dimensions people optimize shift over time, such as the recent focus on data quality.
“even limits to data center size and power, thinking about using the resources that we have well both in training and in inference”
In AI products, the question is whether model output is good enough to sell, which means minimum viable quality must be defined per use case and per customer experience, and it is a moving target as the scope of use cases expands; this replaces the deterministic minimum viable product scope of traditional B2B software.
“are they good enough to sell?”
Challenger AI chips face high barriers because their failure modes in large-scale training often appear only at multi-thousand-node scale, which cannot be fully simulated or tested on FPGAs, so buyers need a patient-zero workload at scale to know whether they work.
“If you don't have people running large-scale training workloads at multi-thousand node cluster size on your chips, you do not know if they are going to work.”
Incubation is valuable where the context for automating a domain is far from the engineers and researchers who understand current AI capabilities, such as configuration, monitoring, and maintenance of enterprise software, which consulting firms in the U.S. and India currently perform at high cost.
“one of the really interesting opportunities that we think about and actually do support incubation around is places where the context of what people are doing that could be automated or improved is so far away from the engineering and research community”
There is no such thing as cheap enough intelligence: the computing industry has always kept finding ways to use more compute, so even if intelligence were free, people would still do more with it when it is not free.
“there is no such thing as cheap enough because we've been working on compute for a long time as an industry”
Believing the dominant narrative in AI is dangerous because narratives serve someone who is pushing them, and investors should instead reason from first principles about capability progression, vendor concentration, and the value of workflows, data, and distribution built on top of models.
“first of all, that narrative serves somebody, somebody is pushing that narrative.”
Many AI problems are more easily and cheaply verified than solved, so model systems that generate many candidates and then use verification or ranking to surface the best ones can deliver a much better end-user experience than any single generation.
“something more easily, more cheaply verified than solved.”
AI creates value in two ways: making highly paid professionals hundreds to thousands of dollars an hour more productive, and enabling work that is too costly to do at scale, such as searching 20,000 contracts for a term phrased differently, which would cost tens of thousands of legal hours.
“if that costs you X tens of thousands of legal hours, in or out of house, you're not going to do it. If it costs you a Harvey run, you might.”
Startups should pursue markets that incumbents are unlikely to win, because very capable incumbents will not dominate every market; founders should be honest about where the distance between the core technology and end users is far, and predict what incumbents care about most.
“I am very willing to bet that very large, very capable incumbents are not going to go win every market.”
A product that is secondary to an incumbent's enterprise value, like the 19th most important product at Google, is often deprioritized and frequently killed as priorities shift, so entrepreneurs may be less threatened than they assume, though incumbents will attack core revenue lines like ads violently.
“if you're working on the 19th most important product at Google, maybe that's okay.”
Health care administration and government services are large, inefficient parts of the economy where a huge number of hours are wasted on data movement and analysis, and if AI makes that work roughly 100 times cheaper, these markets become sellable despite complicated ecosystems and regulatory capture.
“if we make it 100 times cheaper no matter how complicated that ecosystem is, I think we can sell it.”
AI infrastructure is in an immature stage where companies build and deploy their own clusters in data centers under multi-year reservations, managing depreciation, whereas traditional web infrastructure progressed from on-prem servers to colocation, hosting, virtualization, and serverless.
“In AI, we're like step one or two. You are building and deploying your own clusters in data centers. You're in three-year reservation mode.”
Incubating companies is usually a bad idea because founders must live the problem themselves, but incubation is worthwhile where the context of a workflow and the capability to automate it rarely coincide, such as lawyers who know legal workflows and also have researcher roommates who understand models.
“where essentially the context and capability to build a particular type of company rarely come together”
Diffusion of innovation applies to talent as well as technology: people who have worked on large model training or tinkered with models longer have a deeper understanding than others, so there is a temporary edge for those who understand how to build products on these models.
“I think diffusion of innovation also happens with talent.”
A services business can be valued like a technology business if it is built around scalability, using AI to replace human labor rather than just selling human capital, since the most interesting application companies are replacing services with technology even when they are consumed as services.
“The answer is yes. But what would get me jumping out of my seat excited would just be somebody with the relentless ambition to figure out if we can make the margins as good as a software business.”
A likely big change in the AI ecosystem is using a mix of distilled models of different scales inside compound AI systems to reach performance and efficiency goals, since efficiency was not a major consideration for research labs and application companies until recently.
“whether or not people are going to start using a mix of distilled models, models of different scales, putting them in compound AI systems”
Startup strategy should not rely on early traction from a few weeks or months of work, because if something is simple to build and distribution is not unique, competition is coming quickly, so investors should be surprised by how short-term founders are about strategy.
“if something is really easy for you to build unless your distribution is totally unique and defensible, you're probably not going to be able to capture rent”
Without longitudinal exposure to outcomes, investors cannot be confident their judgment about early, high-potential people and markets is right; long build cycles, such as Figma's five years to public release, delay feedback and make pattern matching difficult.
“you don't get the feedback cycle for a while.”
AI represents a new era of software in which models are manipulated and much of the work is already done by capable foundation models, shifting from Software 1.0 explicit source code to Software 2.0 behavior learned from datasets via neural networks, which is a major opportunity.
“I genuinely think of this era as a new era of software.”
Billions of dollars spent yearly on configuration, monitoring, and maintenance of enterprise systems like ERP, HR, CRM, and ServiceNow, largely done by consulting shops, could be replaced by services organizations using code generation, with better margins and SLAs, potentially changing enterprise software's industry structure because configuration creates stickiness.
“there are many billions of dollars spent a year on configuration, monitoring and maintenance of enterprise software systems”
As AI crosses minimum viable quality in different domains, a leapfrog effect will occur in which the types of businesses and functions with the most inefficiency become the most ripe for selling new solutions, so previously unattractive sectors may be the first to adopt.
“the leapfrog effect is the types of businesses and the types of functions that had the most inefficiency can be most ripe for selling new solutions”
Entrepreneurs should not build features that foundation models will replicate in capability in the near term, because that effort is wasted and the output becomes commoditized; founders must predict model capability progress to avoid this.
“you don't want to build things that the foundation models are going to replicate in terms of capability in the near term because it's wasted effort and then like it's commoditized”
Selling something fundamentally probabilistic to a conservative, risk-aware, non-technical audience such as lawyers is hard, so companies must build credibility by demonstrating the size of the economic benefit to early partners.
“your ability to take something that is fundamentally probabilistic and sell it to a relatively conservative or risk-aware non-technical audience is challenging”
For writing products, adoption depends on whether the cost of managing errors, or moving from roughly 80 percent to acceptable quality, is very low; if editing AI output takes as long as writing it, users will not return.
“going from 80% quality to acceptable quality, really, really low cost for the end user.”
Guo hopes for a market with many foundation model options, open source, and robust competition, so that many companies can build economic value on top and walk the long last mile to human uses, rather than having one or two companies collect all rent.
“I want a million flowers to bloom”
The U.S. health care system is expensive partly because administration consumes a large number of employed people and hours, and the incentive structure, slow ecosystem, and regulatory capture make it an unattractive but high-potential target for AI.
“the incentives are a mess. It's super slow. The ecosystem is extremely complicated. There's regulatory capture on every zone.”
Beyond model improvement, companies can creatively wrap models in traditional product design, deal with failure cases, and improve user experience to reach minimum viable quality, which Guo says is a central lever in finding product-market fit.
“being creative about how to improve the quality of the model, use user experience, a traditional product to wrap it”
Teams should recruit for baseline technology understanding and core traits like competitiveness and team orientation, because the alpha is in hiring people before they are obvious, not in hiring those with long track records.
“the alpha obviously is hiring people before they're obvious.”
If an application can be built in weeks or months using a handful of templates, competition will arrive quickly, so such projects can serve as a wedge for customer engagement but are unlikely to create billions in enterprise value without unique distribution.
“if something is really easy for you to build unless your distribution is totally unique and defensible, you're probably not going to be able to capture rent”
NVIDIA's advantage rests partly on CUDA, whose optimization and developer investment across more than 15 years most challengers cannot easily match, and AMD's software stack has not been competitive to date.
“people understand the power of CUDA and the optimization and effort that's been going into that for more than 1.5 decades.”
The market is not infinitely deep for another general-purpose LLM player that is not particularly differentiated, and companies keep being funded because people envy the value creation at OpenAI and Anthropic, which is a flawed basis for investment.
“I don't think that the market is not infinitely deep for another LLM player that isn't particularly differentiated”
Conviction's portfolio intentionally overrepresents teams with research backgrounds; this has disadvantages because researchers often have never met the market, but a deeper understanding of models is a significant advantage while adoption and understanding are early, and it may not persist in five years.
“I'm saying we have chosen to back a lot of people in research, and it has some disadvantages because those people have often never met the market.”
A year and a half ago, focusing on data quality was not considered high-status research, but now people care a lot about it because the field has run up against limits on scaling internet data, showing that the dimensions of optimization change over time.
“A year and a half ago, I'd say a focus on data quality wasn't considered the highest status part of research.”
Large AI labs were quite surprised by how successful open-source models became over the last year, which illustrates that anyone claiming to know what will happen more than six months out is likely overconfident.
“the large labs were quite surprised by the success of open source over the last year.”
On SWE-bench, models reached only about 13 to 20 percent success, which is not good enough; Guo says that if software engineering interns produced acceptable output only 20 percent of the time they would be fired.
“we got to 13%, 15%, even 20%, that's not good enough.”
Incubating a company early in a venture career, with the investor doing hands-on work such as recruiting cofounders, talking to design partners, and drawing the marketecture, is educational because it teaches the substance of building and gives patience that investors with many competing opportunities lack.
“This was educational because it's just a cold start, and I spent half my time working on building the company.”
Silicon Valley is not a meritocracy, but it is the closest environment Guo has found to one, because its marketplace for ideas lets an unknown associate's blog post lead to a cold outreach from a leading researcher.
“the marketplace for ideas in Silicon Valley is pretty good.”
Knowing what it looks like when a product achieves genuine user love and adoption within a leading influential group in a domain is useful pattern matching, because that kind of pull is different from generic product-led growth.
“knowing what it looks like when you do have user-love adoption within the leading influential group of particular domain is just useful for pattern matching.”
Guo wishes she had started three to six months earlier; founders waiting to optimize another month or quarter before starting a venture are making a mistake, because the market is moving and it will always be hard.
“I wish I'd done this even three to six months earlier.”
Material science is an underexplored domain for foundation models that could be very valuable to society, but the required proprietary data is not generated or owned by large labs and requires talent and data strategy to build.
“material science as a domain for foundation models is not generated and owned at the large labs today”
The world is full of uncanny valleys in AI, and so far only a few have been crossed in small ways; each valley is crossed by improving foundational capabilities, product wrapping, and verification systems, and many remain.
“I think the world is full of valleys. It's all valleys from here, like we've crossed three of them in little ways.”
HeyGen's rapid growth came from crossing a minimum viable quality barrier, where generated video of a person speaking was good enough for customers to use commercially, even with input from consumer-quality webcam footage rather than studio capture.
“until this past year, until HeyGen broke through that minimum viable quality barrier, it just wasn't for like any player”
Seeing academic results, powerful models in large internet companies, and the 2017 transformer breakthrough made it clear that AI is a huge change in software rather than an incremental one, and once the opportunity looked too big to ignore, Guo chose to start a pure early-stage firm.
“If you think it is actually a huge change in the landscape of software, not an incremental thing, then the opportunity becomes a little too big to ignore.”
Intelligent systems built on foundation models could do increasingly useful work in health care and science better than humans in areas that matter, and could take operational toil out of every industry.
“maybe do it better than humans in areas that really matter, health care and science, but also in daily work and taking the operational toil out of every industry?”
Although abundance is possible, society's ability to absorb rapid productivity change is poor, so there will be winners and losers and real job displacement, and Guo cannot assign a clear probability to the 2030 outcome.
“the ability of society to ingest that much change in productivity is not great. There will be winners and losers, and the job displacement dynamic, I think, is real.”
Guo is not particularly concerned about OpenAI's operating expenses being existential, because the underlying wins behind them resemble the rest of computing's investment in efficiency, and whether ChatGPT is a good business depends on it being a differentiated market product.
“I am not particularly concerned about OpenAI's OpEx.”
The first concrete step in starting a fund was talking to limited partners who had already expressed interest, then raising the fund, while in parallel figuring out strategy around the goal of being dominant in some scope of an asset class or domain.
“So I guess step one would be go talk to LPs.”
Harvey's success combines a researcher cofounder's technical fit with a cofounder from elite legal practice, plus core entrepreneurial traits such as velocity, work ethic, and the ability to engage customers and get them to take risk.
“The combination of DNA and technology fit for purpose is very obvious where Gabe is a researcher, Winston comes from the legal domain”
The market gives credit to investors who take a genuine point of view rather than following consensus, which is part of why timing and conviction matter when launching a new firm.
“there's some credit assigned in the market for taking a point of view that is actually a point of view versus like following”
Enterprises have some challenge believing Facebook will be a great long-term partner for AI models, though engaging the ecosystem of deployment, inference, and consulting partners makes that partnership more credible.
“I still think that enterprises have some challenge believing that Facebook is going to be a great long-term partner to them.”
Just as web applications moved from owned servers to serverless and click-button cloud capacity, AI inference will eventually be abstracted so developers do not manage clusters, scheduling, or GPU failures, and companies like Base10 are building toward that.
“You want to do inference on GPUs of different types that is serverless and not write all of your own scheduling and GPU failure manage”
Leaving an established venture firm is complicated because one must responsibly honor personal commitments, such as remaining involved on boards, and leave things well with founders, which can take longer than one expects.
“You want to be responsible to the personal commitments you've made, I carried some Boards for Greylock still”
Her motivation is working with the most interesting and high-leverage people and being right over a long period, not financial reward, since money is not the main driver in an era of ambitious software change.
“it's really fun to be right, go compete, work with great people”
Early-stage venture capital has a very small sample size, so nothing on paper reliably predicts who will become a great early-stage investor; judgment is the hardest trait to assess in junior candidates, so proxies such as interest in how businesses work and reasoning about why products succeed or fail must be used.
“the sample size is so small, and there is nothing on paper that is going to, at least in my experience, going to predict well if somebody is going to be a great early-stage venture capitalist”
AI could let individuals become experts far faster and create desktop-Pixar-level creative tools and universal knowledge access, though the pace of societal change and job displacement will be limited by adoption and market structures.
“Somebody can become an expert in something at an incredibly accelerated rate because of that.”
Self-driving has reached the point where minimum viable quality for a specific use case, a driverless taxi, is provable to a safety regulator, as in San Francisco this year, which shows the concept generalizes across machine learning products.
“Now you have minimum viable quality as provable to safety regulator in the city of San Francisco for a taxi with no driver.”
Confident leaders give early-career people as much opportunity as they can handle based on their potential, even when résumés look unremarkable, because trust is earned and sponsorship can be transformative.
“give your people as much and perhaps a little too much opportunity if they prove they can take it.”
The core framework for evaluating software companies still applies to AI application companies: distribution, quality of the people, and size of opportunity matter most.
“The core framework for evaluating software companies remains the same, which is all those things, distribution and quality of the people and size of opportunity matter”
Greylock's long history taught her that innovation happens unevenly in technology, and recognizing this was critical to the success of some legendary partners, which informed her decision to focus on early-stage AI.
“seeing how unevenly the innovation happens in technology and how important recognizing that had been to the success of some of my legendary partners”
The uncanny valley concept applies to AI, and Patrick observes that an audio narration model reached good-enough quality for road-trip narration, which he judged better than a human guide because it is more knowledgeable and customizable.
“the valley is crossed, so to speak.”
Mistral's founders, Arthur Mensch, Guillaume Lample, and Timothée Lacroix, combine state-of-the-art performance with an efficiency focus, reflecting a view of faster, more democratized progress and economic value across the stack, which is why she invested early as a first-round investor.
“Arthur, Guillaume and Timothée are amazing researchers with a bent toward both state-of-the-art performance but also a view on efficiency.”
Harvey's product can be described as Copilot for lawyers that increasingly automates tasks which used to take junior legal staff hours, delivering an automated answer for those tasks.
“this is Copilot for lawyers that does increasingly sophisticated tasks”
OpenAI's SearchGPT announcement was predictable, since search was a likely future move, and the right response is awareness and a strategy against it, not avoidance.
“I think it was pretty predictable that they were going to do Search at some point.”
Conviction's first dated investment was in the legal AI company Harvey, made as an angel right before the fund fully closed, and the position was warehoused into the fund.
“I invested as an angel right before the fund was fully closed, and we kind of warehoused it.”