In-Depth · XiaoHu Explains

Silicon Valley investor Sarah Guo on AI startups: what to do when "the AI giants will crush us"

Worrying that Anthropic will turn your idea into a feature is usually the wrong question — the real one is which team inside that company you're actually competing with.
The one-minute version
  • The pushback founders hear most often now is that Anthropic or OpenAI will just build this themselves. Five years ago, the same line named Google and Amazon instead.
  • Sarah Guo's answer: don't ask whether the big company will build it — ask who they'd assign to build it. Every company only has one A-team, and it's parked on the core business; everything else doesn't get their attention.
  • Her read: Anthropic's core is the model itself, and its core use case is code — everything else is secondary. It deliberately skips video and image models, partly for safety reasons, partly just to stay focused.
  • AI doesn't exempt anyone from the fundamentals of business: Anthropic's $40-billion-plus revenue came from developers showing up on their own — selling there is mostly just order-taking.
  • Conviction's biggest shift in thinking lately is dropping Silicon Valley's old orthodoxy of "nail one thing first." Customers don't want a million vendors — being a single-point solution is just a temporary state.
Opening

The line founders dread most right now

Five years ago, pitching a new product meant fielding the same challenge every time: Google will just build this themselves, Amazon will turn you into a feature, you'll get steamrolled — so what's your edge?

Pitch investors today and the same challenge hits even harder. In the AI era, a single model update can bury a product in a matter of months, sometimes weeks. The names in the challenge have changed too: now it's Anthropic that's going to eat you, OpenAI that's going to turn you into a feature.

The threat, five years ago

A big company copies what you built. That takes a project charter, a roadmap slot, staffing — which bought you at least a year or two to build a moat and lock in customers.

The threat, now

A single model update can wipe you out. Nobody has to target you specifically — the moment the capability ships, your thin layer of value is gone.

That question landed on Sarah Guo.

She's the founder of Conviction, a Silicon Valley venture firm. Before that she spent nearly a decade at Greylock, then struck out on her own in 2022. The reason she's known is simple: everyone else piled into AI in 2023, but she bet the whole firm on it in 2022 — and the bet paid off. Her earliest investments included Harvey, the AI legal tool that went on to become one of the fastest-growing AI applications; Sierra and Open Evidence followed. In four years, Conviction went from nothing to a name Silicon Valley actually knows.

Host Matt MacInnis put the challenge to her directly: does that logic still hold? Would you still pass on a company because Anthropic might eventually build it? Where do you draw the line?

She didn't answer yes or no. She swapped out the question: first figure out who inside that company you're actually competing with.
Reframe the question and it stops being an unanswerable guess and becomes something you can actually check: where does what you're building rank on that company's priority list, and what caliber of person would they assign to it?
Full interview, 54 minutes, with bilingual Chinese-English subtitles added by this site. Host Matt MacInnis is on the left, Sarah Guo on the right. The rest of this piece follows the conversation in order — watch here if you want the original audio. Source: episode one of the Rippling podcast First Principles.
Core argument

Big companies only have one A-team

She starts by conceding one case where the fear is real: go up against search and advertising, and you're facing Google's actual A-team — and those people are genuinely good.

But most startups aren't in that situation. She says a big company's product portfolio is so wide that most of it isn't staffed with the A-team — because there's only one A-team, and it doesn't stretch that far.

If your product is Google's priority number 27, that's really not so scary. Sarah Guo

She has a name for this: organizational physics. The idea is that there's a hard physical limit to how many things a company can do well at once. Unless it was built from day one to fight on multiple fronts, it'll only excel at its core business and be mediocre everywhere else. That's not about good or bad management — there just aren't enough people to go around.

An analogy

It's like a team with only one starting lineup. Play three leagues at once and two of those games are necessarily fielding backups. If you happen to land in the game where they've sent the bench, "they're stronger than you" isn't really the whole story anymore.

A giant's product-line grid Dark cells: the core, guarded by the A-team. Light cells: everything else Core business A-team here Second pillar Gets some talent too B-teamB-teamC-team C-teamC-teamOutsourced MaintenanceNear-cut You are here The question: where does this rank for them, and who do they send? The lower it ranks, the fewer people they commit — and the less urgency they feel
Diagram drawn by this site based on Sarah Guo's remarks in the interview; the number and division of cells are illustrative, not an actual product lineup of any real company.

Her point, then, is that founders should feel more confident going after everything outside a big company's core business. Those areas get B-team and C-team staffing — which, for a startup, is open space.

So where exactly is Anthropic's core?

Which raises the question: how do you know where a company's core actually is, and whether you're about to run into its A-team? The host pressed further: what was Anthropic built to be good at?

Her read is that the model itself is the business — the core use case is code and models that can improve themselves, pushing capability forward toward AGI — and everything else is fairly secondary. What backs that up is what it has deliberately chosen not to do.

The evidence

Anthropic doesn't build video or image models. She thinks that's partly a safety call and partly just focus. What a company deliberately gives up tells you more about where its core is than anything it claims to prioritize.

So why don't the big labs just scoop up the application layer while they're at it? Her answer: they've actually tried both horizontal and vertical applications, and it hasn't worked out — not because they can't see the money in enterprise. Every step of running a software business that actually satisfies customers is hard: