The compute bottleneck isn't technology, it's persuading people to accept data centers
An infrastructure lead at a hyperscaler told her nothing can change the picture at sufficient scale before 2030. She thinks this isn't a technology, capability or capitalism problem — it's a regulation and persuasion problem.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Knowing 250 people is enough
She repeats her partner Mike's framework: the people actually pushing the frontier forward are roughly 250 founder-and-researcher combinations — the ones who are really moving things forward each morning. Conviction's goal is to know all of them, be as close to them as possible, and support them in every way. She distinguishes this from ‘booking meetings from cheap seats’. It also explains why she says early-stage competition isn't as fierce as people imagine: if you genuinely understand the technology and the community and approach it from first principles, you can get better access and better judgment than less focused people — the rest she calls an execution problem.
— Sarah GuoResearchers are starting to feel irrelevant
She describes the moment as violently competitive, and the competition is narrative: over the past 12 months, many researchers have come to believe recursive self-improvement means exponential intelligence within a year or two — Karpathy jokes that he's probably been saying ‘two more years’ for a decade. What really concerns her is the loss of a sense of ownership. When OpenAI had 200 people, everyone could feel they were pushing something forward; when the problem becomes ‘I need $750 billion in compute spend, with thousands of people working on it’, people feel the outcome has little to do with them. She reduces this to two equally disempowering beliefs: nothing I do matters because the model will do it; or the only thing that matters is compute scale.
— Sarah GuoThe compute bottleneck is regulation, not technology
This week she talked to an infrastructure lead at a hyperscaler, who said nothing can change the picture at sufficient scale before 2030 — she used the word depressing. She thinks this isn't a technology, capability or capitalism problem, but a regulation problem and an alignment problem, and she stresses it's not AI alignment: to build a data center in New York, you first have to get New Yorkers to want a data center there; to make nuclear power competitive as baseload, you have to convince people it's safe and allow enough reactors to be built to push the cost curve down. She also says the physical supply chain — tacit knowledge, labor and raw materials — can't move at software speed; the only way is to invest in it.
— Sarah GuoBetting on résumés is dangerous
She points to a phenomenon: many people should be, and are, making technical bets, but their depth of understanding of the technology varies, so judgment gets delegated to résumés and other legible signals — who invested, who recommended it — and decisions become less fundamental. She describes a disagreement with a friend: she asked why this company could become huge, and the answer was basically ‘do you know how strong this person is?’ She's known this person for eight years, and he himself admitted he couldn't explain the technical logic of the business. Her conclusion: people are making large research-driven bets with no intuition or opinion about the technology, and no view of the business beyond the person's résumé, and that's dangerous.
— Sarah GuoTwo people produced the key ideas in robot AI
She and her partner Pranav met Tony Zhao and Chang Chi when they were still Stanford PhD students — they'd been at Toyota Research, DeepMind and Tesla, were about 25 when they founded the company, and one didn't finish his degree. After reading their work, she concluded that most of the interesting ideas in robot AI over the past four years came from them. They're attacking generalization and robustness in robots — the industry believes that once you have an internet of robot data, general-purpose robots will arrive on their own, and the bottleneck is where to get that data. So they treat ‘how to collect data covering real environments and task distributions as cheaply as possible’ as a technical problem. The company is less than two years old, and the team believes that by the end of this year there will be a beta of a general-purpose semi-humanoid robot doing work in homes.
— Sarah GuoRestricting open source only hits law-abiding players
Her judgment is that the cat is out of the bag: over the past three years, open models from all sides have become increasingly competitive, mainly from China but also from the US and Europe. Even setting aside lab economics and looking only at how capabilities diffuse into the economy, there are many scenarios where using frontier vendors' models is too expensive, too sensitive or too slow, and that share will only rise. So if the US restricts the use of open models, the practical effect is to restrict law-abiding American companies, slow them down, and shift profits to other pockets; real attackers are unaffected. She advocates strict safety testing at the frontier, actually investigating rather than speculating that ‘Chinese models might have backdoors’. She also doesn't think abundant intelligence is inevitable: if people turn anti-capitalist out of fear of losing jobs or resentment of a few companies collecting rents, that sentiment will slow energy and infrastructure.
— Sarah GuoModels changed the verdict on making money in biology
She says biology is the field where she ‘strongly took a side after seeing empirical data’. The old conclusion was: in biotech or serving pharma companies, the only way to make money is to make a drug, get a Biobux deal, and then decide how much risk you're willing to take; the corresponding capital structure is a good traditional biotech firm finding a principal investigator, owning 40% of the company, and assembling things into a drug candidate, and most fail. So ‘you can't make money selling software to pharma companies, and you can't build a platform business on top of pharma companies’. Her debate is: will models change this, and her answer is a strong ‘yes’. The company she invested in, Chai Discovery, is working deeply with several top-10 pharma companies to accelerate a segment of R&D. She also acknowledges that regulation and the speed of the physical world remain unsolved.
— Sarah GuoA one-person marketing department will become the norm
What she most wants to see a year from now is the Jevons paradox playing out in reality: agents and products doing more of the trivial things in every area of life better, with an effect like the transformation that has already happened in software engineering. She gives an example from a portfolio company: it serves many customers, and its marketing department is ‘like one person in a room’; to create leverage for himself, the marketing person built an autonomous marketing department that handles traditional work like sales enablement content. Her conclusion is that in every function, if you learn faster and do fewer trivial things, the time saved always gets used elsewhere — so when asked whether she works less, the answer is more, and she attributes this to Jensen's core wisdom: we'll need more people, and the key is to make sure people can get and learn to use these tools.
— Sarah GuoIn their own words · checked verbatim
there's sort of 250-ish people that he thinks about or you guys think about that are some combination of entrepreneurs and researchers and doing the most interesting things on the frontier
Sarah Guo7:08
Now, if the question is, well, I need $750 billion of compute spend, and we have many thousands of people work on this problem. I think people feel less ownership of the outcome.
Sarah Guo12:25
I don't think it is a technology or capability or capitalism problem. I think it is a regulatory problem and an alignment problem. And I don't mean like AI alignment.
Sarah Guo14:29
But I think that they have contributed dual-handedly most of the interesting ideas in robotics AI over the last four years.
Sarah Guo19:49
Because the actual attackers or people who have adversarial uses of these things are not affected by your restrictions. You're restricting your own people.
Sarah Guo37:29
I don't think it is inevitable that we are competitive. And I think we need to make that decision actively.
Sarah Guo39:34
Dumb software investors, you can't make money selling software to pharma or build platform businesses at pharma. The debate is, does it change with models? I'm a strong yes.
Sarah Guo45:51
I'm hopeful that a year from now, we see Jevin's paradox in practice as we have agents and products that do more of the mundane more effectively in all the domains of our lives.
Sarah Guo53:07
Figures
| Conviction's core frontier list | About 250 people (founders and researchers) | 7:08 |
| Compute spend envisioned by researchers | $750 billion | 12:25 |
| Hyperscaler infrastructure lead's judgment | Nothing can change the picture at sufficient scale before 2030 | 14:29 |
| Age of Sunday Robotics' two founders at founding | About 25 | 19:49 |
| New companies Sarah sees per week now | 4 to 6 | 27:05 |
| Companies Sarah saw in her first months at her previous firm | 500 | 27:05 |
| Hourly wage mentioned in the automation narrative | $13/hour | 39:34 |
| Pharma companies Chai Discovery works with | Several top-10 pharma companies | 45:51 |
| Traditional biotech firm's stake in a new drug company | 40% | 45:51 |
Glossary
- compute independence
- She predicts this will be the next big topic: supply chain links like wafers and glass that don't depend on a single region.
- intelligence too cheap to meter
- Sam Altman's phrase for intelligence becoming so cheap, like electricity, that you don't meter it; she thinks open source will underpin it.
- Jevons paradox
- Efficiency gains increase total consumption; she uses it to argue AI will make people busier, not less busy.
How to listen
Early-stage investors looking at AI applications, robotics and compute, and engineers who care about where open models and US compute policy are heading.
30:49 to 33:21, the segment on her parents and upbringing — the lowest information density.