Jason Cohen：融资算术、AI定价与Build-to-Prod建议

The deeper cut on a week where capital concentrated harder than I’ve ever seen, and what founders should actually do about it

Harry, Rory and I covered a lot this week: a $20M ARR profitable company that can’t get a meeting, Anthropic’s $16.8B quarter, CoreWeave’s 25% drop, Modal at $15.75B, and Berkshire doubling down on Bending Spoons. Here’s my cut for founders.

1. Top-Decile Numbers in an Unsexy Category Means 100 VC Meetings

For years, top-decile growth was an automatic term sheet. You’d walk into Rory’s office at Scale, he’d know every comp since the dawn of time, and if he believed in you, you walked out with a term sheet. That world is gone if you’re outside the 3-4 categories everyone wants.

The company Harry met is $20M ARR, growing 6-7x, profitable. Crickets from the top firms. Someone will still fund it. It just won’t be the seven Benchmark partners. It will be a growth investor whose name isn’t Index or a16z, who sees it as the best deal of their quarter.

My learning → If you’re vertical software or anything without “AI infra” in the pitch, plan the round like it’s 2009. Build a list of 100 investors, not 15. Budget 2-3 months. Treat the top 10 firms as a lottery ticket, not the plan.

2. Know the Fund Math of Whoever You’re Pitching

A $5B fund raising every 18 months needs billion-dollar positions. Not billion-dollar exits. Billion-dollar gains on their stake. In their IC memo, a $1B gain is the B case. The A case is $25B+.

So when you pitch a $2B outcome to a mega-fund, you’re not being rejected on quality. You’re being rejected on arithmetic. Rory’s right that only a handful of funds can play this game, but those are exactly the ones founders chase first.

My learning → Before every pitch, divide the fund size by 20 and ask whether your realistic outcome gets them a position that size. If it doesn’t, you’re pitching the wrong fund, no matter how good the meeting feels.

3. One-and-Done Used to Be Nirvana. In the AI Era, It’s a Trap.

I used to love one-and-done. Raise a couple million, never raise again, no pressure, keep growing. Zapier is the canonical example: one real venture round, about $300M in revenue, generational in its own way, probably won’t IPO.

But what if Zapier had raised nothing and the founders just split $200M a year? That’s 37signals. Once you have VCs on the cap table, dividending out profits gets hard. Buying them out from cash flow gets expensive. If an investor owns 30% and you pay 6x cash flow, you’ve tied up two full years of cash.

My learning → If you can get going without venture, do it. With AI cutting the team you need to get to $10M ARR, none-and-done is now realistic for a lot more founders than it was three years ago. One-and-done gets you the worst of both: a cap table that wants an exit and a business built not to have one.

4. Pick Seed Investors Who Can Carry You

The portfolio company I mentioned raised about a $3M seed and plans to never raise again. It’s growing 15-18% a month, fully agentic, checks every box. Nobody wants to meet. That means its early investors may have to carry it, like the old days.

That only works if they have reserves and the conviction to use them. A lot of seed funds don’t. And the ones that are actually a multi-stage firm’s seed program will crush everyone else’s allocation and may not care about a $500M outcome.

My learning → Ask every seed investor two questions: what are your reserves per company, and what did you do the last time a portfolio company was growing well but couldn’t raise? The answer tells you whether they’ll carry you or just send a nice note.

5. Take the Valuation That Wins the Day

Anthropic was expected at $2T for its IPO. The talk dropped to $1.5T. Chamath’s take was right: raise whatever you need at $1.5T and move on. It’s just a valuation. Win the day.

The same applies at every stage. A billion is the new $100M at Series B. But if you stretch to the top number and your growth slows from 300% to 70%, you wake up inside a great company that’s simply priced wrong.

My learning → Pick the valuation you can comfortably grow into in 18 months, not the one that tops TechCrunch. The extra 30% on the headline price isn’t worth a flat or down next round.

6. Your AI Revenue Has Less “R” in It Than You Think

Anthropic’s quarter was unprecedented. But developers aren’t loyal. The share swings between Anthropic, OpenAI, open weights and brand-new entrants just since this show started have been enormous. It wouldn’t shock me if Anthropic’s revenue share is down 10% a year from now while the business is still great in absolute terms.

If that’s true for the best AI company in the world, it’s true for your usage-based AI product too.

My learning → Don’t model AI revenue like classic B2B ARR. Run a plan where 20-30% of usage revenue is at risk to a cheaper or better alternative in any 12-month window, and know what you’d do if it happens.

7. Route Your Models Like Replit Does

Replit flipped its default from Opus and Fable to OpenAI. The press called it a breakup. It wasn’t. We use Replit 14-15 hours a day. They have their own router, and OpenAI was 30-40% cheaper for roughly the same output. Next week it may flip back.

The bigger tell: Replit, Lovable, Base44 and v0 are each around $1B in revenue, with extremely cost-sensitive customers at every tier. They have the strongest incentive on Earth to route well. None of them moved users to Kimi. That tells you more about open-weight quality in production than any benchmark.

My learning → If you’re building on LLMs, own a routing layer and re-test your defaults monthly. A 30-40% COGS swing on the same quality is the difference between a 50% and a 70% gross margin business.

8. Build-to-Prod Is the Most Important Category in Software

Everyone is building 10-14 hours a day in Claude Code, Cursor, Replit. The code is pretty darn good. It’s never production-ready. The time to review, test and harden it is still an order of magnitude more than the time to generate it. The engineer builds the feature on their phone at a barbecue and spends the next three days fixing it.

Generation is a solved-enough problem. Let Sam and Dario fight over it every day. We all benefit. The open question is how anything gets from “works in dev” to production. That’s where I’d make 10 bets today, knowing some may get killed by Anthropic. Claude Code Review launching proves the point rather than ending the category.

My learning → If you’re building in the coding stack, build for the 90% of time spent after the code is written. If you’re a team shipping with AI, budget review time explicitly: when we scope a feature now, the build is the small number.

9. Don’t Build a Company That Exists Because Someone Else Needs It For Now

CoreWeave dropped about 25% in a week. It exists because Nvidia needs it to exist right now. It isn’t producing silicon or software nobody else can. Oracle is renting excess capacity because it can’t sell enough enterprise software. Those are B-tier assets, and B-tier assets get hit hardest by every shock.

Founders build B-tier companies all the time without realizing it: the wrapper a platform tolerates, the integration a partner hasn’t gotten around to building.

My learning → Write down the one thing you do that your biggest partner or supplier couldn’t do themselves in 12 months. If the honest answer is “nothing, they just haven’t prioritized it,” your valuation multiple and your shock risk should reflect it.

10. AI Versions of Low-Multiple Categories Will Settle Toward Their Peers

Marketing tech is the worst public category. HubSpot and Klaviyo aren’t dying and trade under 3x revenue. Braze was the best exit in public GTM at about 6x. Now look at the agentic players in those same categories raising at 40x+.