20VC复盘Cursor收购：定价逻辑、多模态转型与并购博弈

Plus the math that has to be true for AI to reach $600B in revenue: $100K of tokens per engineer, and 30% fewer engineers.

A heavy week on 20VC x SaaStr with Harry Stebbings, Rory O’Driscoll and Jason Lemkin:

• SpaceX closed its $60B all-stock takeover of Cursor
• Stripe bought OpenRouter for around $7B, four months after OpenRouter was valued at $1.3B
• Anthropic turned its first profit
• Silver Lake circled Workday at $43B
• And Higgsfield and Lovable raised at $5.5B and $13.3B on roughly the same revenue.

Here’s what we got into:

#1. By the time the Cursor deal closed, $60B wasn’t an expensive price.

Rory’s math: At roughly 15 times current revenue, or under 10 times the $6B the company is tracking toward by year end, this is a normal multiple for the fastest-growing asset in the largest AI category. SpaceX itself trades around 40 times revenue. Buying at 15 with paper that trades at 40 is accretive the day it closes.

Jason’s read: The best AI deals can often look absurdly expensive at signing and by closing, look almost conservative at closing. Don’t look at ARR multiples, look at the multiples of forward revenue and forward growth.That’s the bet.

#2. Cursor was declared almost dead on this show not 12 months ago. Going multimodel is what saved it.

Jason’s read: We’ve been doing this show about 70-something weeks and there have probably been three different Cursors in that window. It rocketed to half a billion in revenue. Then Claude Code came out and it genuinely looked over. Nobody’s portfolio companies were using it. I’m not even sure Claude Code had launched when we started the show, which is insane to say out loud. Cursor went multimodel very early, and that one call turned into a $60B outcome.

This was not linear progress. Many teams would have quit on that journey. Being agile isn’t enough at this rate of change. You have to be able to absorb a competitor shipping something better than you and keep executing your own plan anyway.

#3. The gross margin criticism was correct. It just stopped being the deciding factor.

Early Cursor was selling a dollar of tokens for well under a dollar. That was the standard VC mockery when we started the show.

Rory’s read: The negatives you can cite along the way tend to be true. They were real. It’s easy to sound financially sophisticated and say everything eventually has to generate free cash flow, so a business without gross margins is worth nothing. That’s wrong when the market underneath is exploding, because the buyer looks past it. Coding is the biggest AI market there is, and a large share of Anthropic’s value is predicated on it. If you own the number two asset in that category and it’s compounding, margin structure is a second-order issue.

#4. Elon bought a gross margin problem he was already positioned to convert into revenue.

The cleanest line of the episode, and the actual logic of the deal: your gross margin problem is my revenue opportunity for my Colossus cluster.

Rory’s read: The set of buyers for whom this works is tiny. You need enormous compute and no obvious business sitting on top of it. That describes two companies, Meta and SpaceX. One of them bought.

Cursor’s worst line item was inference paid to someone else. Owned by a buyer who owns the compute, that cost becomes internal revenue. The same P&L looks completely different depending on who holds it.

#5. How you take a company off a round it already has.

Cursor was closing roughly $2B at a $50B valuation from Andreessen and others. Elon bid $60B, added a $10B breakup fee if the deal failed, and granted autonomy. That combination removed the objections.

Jason’s read: Zuck did Instagram and WhatsApp in about an hour on the back of a napkin and paid high. This is the same move executed better. When a company is mid-round at a known price, you don’t run a process against it. You put a number on the table that makes the round irrelevant and you underwrite the seller’s downside if you can’t close. Very few people on earth have the authority and the currency to move that fast.

#6. Why Meta (potentially) passed, and why the seller may not have wanted the Meta deal anyway.

Harry’s read: Meta is building model capability with Alex Wang and company and is missing the enterprise piece. Cursor would have solved that.

Three reasons it went the other way:

• Deal certainty. SpaceX filed and got antitrust clearance quickly. A seller would have far less confidence a Meta deal clears given the scrutiny Meta attracts. Certainty is a term, and sellers price it.
• Necessity. SpaceX built AI capability out of an existential need to use compute it already owned, and it now rents that compute to Anthropic and Google. Meta’s AI push is driven by interest rather than necessity, because a $100B+ advertising business already works.
• Who you want to work for. For a founder already running a $50B company, the calculus is who executes. Mass layoffs and pressure to produce a frontier model is a different offer than what Elon put in front of them. The broader founder advice on the show: in M&A, ignore brand reputation and the job description you’re promised, because the business will look nothing like it in 24 months.

#7. Microsoft is arguably the real loser in the Cursor deal.

Rory’s read: owning a developer platform is existential over the medium term for exactly one company, and it isn’t AWS and it isn’t Meta. Microsoft lost its connection to developers, the group that made it in the first place. GitHub is now a trailing-edge product. That’s a serious medium-term loss even though the company is operationally fine.

Also worth flagging, per Jason: the popular VC theory that one big acquisition triggers an industry-wide land grab is mostly wrong. Buyers don’t panic-buy the second or third choice because a rival moved. They buy when they realize they have less time than they thought.

#8. Stripe paid roughly $7B for OpenRouter, four months after a $1.3B round.

OpenRouter, led by Alex Atallah, is the leading LLM routing layer. Stripe’s business has always been absorbing complexity in exchange for a share of the flow. Model selection and API integration is the same shape of problem. Stripe already benefits from AI growth indirectly, taking a cut every time somebody pays OpenAI or Anthropic with a card.

Rory’s read: current revenue is close to irrelevant to the price in deals like this. The acquirer is buying future extractable value, and can usually generate cash from the asset faster than the standalone company could have. In a market moving this fast, the buy-versus-build analysis that assumes a five-year build window gets abandoned. A few billion in cash and stock buys you a live capability inside a week. This is what venture returns are made of, and it happened repeatedly in the early internet buildout.

#9. The bear case on OpenRouter is that enterprises only want two or three models for most workflows. And they may want their router as part of the harness.

Jason’s read: routing is genuinely useful for developer tools and chatbots, where availability matters more than perfectly consistent output. It’s much less useful for high-reasoning B2B workflows, where switching models introduces drift and forces constant re-testing and QA. Rippling picked a small set of models on price, performance and speed rather than managing a large menu, and that’s the more common enterprise posture. Databricks, Replit, Lovable and Vercel all do their own routing.

The counterpoint: enterprises that don’t want to be captive to one frontier lab use a router as plan B, and having plan B is what keeps the lab honest on price.