The latest 20VC roundtable with Harry Stebbings, Rory and Jason ran through nine of the biggest stories in B2B and AI right now. Anthropic filing to go public the same week it closed a $65B round. Cognition at $26B. The best public software earnings week in two years. And a budget fight that is going to redraw the org chart of every engineering team on the planet.
Here are the threads worth pulling on, with the most quotable moments from each at the end.
We Are About to Start Choosing Tokens Over Humans
The single most important number in B2B right now is the ratio of dollars spent on engineers to dollars spent on tokens. Nobody knows what it settles at yet. The whole AI market gets priced off the answer.
Run the thought experiment from Jason’s Adobe days. You manage a 400-person product and engineering org. Old world, your budget was just headcount. Everybody cost roughly $300K, office manager or staff engineer, and you didn’t think about anything else. Now you get a fixed dollar budget and a choice: keep 400 people, or go to 300 people and spend the other 100 salaries on tokens.
In 2024 you keep all 400, because you needed bodies to pick up the phone and hold the customer. In 2026, if your team will commit to even 1.5x output, the cuts are obvious and they take about ten minutes to make. The bottom of the QA list, the marginal CS roles, the inbound reps closing $3K deals. Whatever survived the first layoff wave gets cut for tokens in the second.
The data points stacking up:
- Uber is capping engineers at $1,500/month in token spend. That is ~$18K a year, roughly 10% on top of a $200K engineer. A first-pass guess at what “normal” looks like.
- The EDA software market, the most tool-heavy engineering market that exists, runs about 13% of engineering spend. That has been the historical ceiling for “tools as a fraction of engineer cost.”
- Jason’s read is that for the best shops it goes well past 10%, maybe a third or more. Rory thinks that math is too aggressive. The honest answer is nobody has run a real end-to-end org at that ratio yet.
- The startups are already there. Brandon at McCor says they now spend more on tokens than on engineering salaries with ~80 engineers. The conceit in those stories is the headcount. 80 is not 1,200.
The reason this matters for valuations: you can get to the Anthropic and OpenAI growth curves on even a 10% allocation across corporate America. At 33%, you buy the IPO at any price, and you also watch one in three or one in four roles across the product and engineering stack get replaced.
SaaStr already made the call. We got rid of the B players and would rather have tokens. The bet only works if your VP of Engineering can actually ship product, not just pull requests, with that level of cut. The faster a company is growing, the more they tell you yes. The slower it grows, the more they tell you it can’t be done.
Founder takeaway: Start measuring your token spend as a percentage of fully loaded engineering cost now. That number is your single best leading indicator for where your headcount lands in 2027.
Anthropic Files to IPO, and the “Billion-Dollar Position” Bar Resets
Anthropic raised $65B and filed to go public in the same week. ARR is up 28% since the prior episode. It is on track to be the fastest company ever to IPO anywhere near its scale, and SpaceX is formalizing at $1.75T for early June.
When the best company of the decade goes from zero to a trillion in five years, it warps the whole bar. Jason’s version of the reset: he won’t take a meeting now that he would have taken in 2024. Not because the founders aren’t good or the companies aren’t real. The bar for a position just moved. His new line is that he is not interested unless it can be a billion-dollar position, not a billion-dollar outcome.
The distinction matters. A billion-dollar outcome can happen with a pretty-good CTO, a mid-size TAM, and some luck. A billion-dollar position requires you to underwrite ownership times a much bigger number. So the screen becomes: are there tangible blockers? An A-minus CTO who won’t launch 17 products at once. A small TAM with no real path to a large one. Complainers. Those are blockers, and now they’re a fast no.
Rory’s pushback is the base rates. In a normal year there are maybe four or five $10B-plus software outcomes total, and one to three $100B-plus outcomes per decade. You cannot make 20 to 30 seed or Series A bets and credibly believe each one is a billion-dollar personal position. So he underwrites to a realistic base case and never caps the upside. Same place, different math.
Where they fully agree: the floor for “aspirational enough” has gone up. People have now seen what 10x growth for three years actually looks like, and that anchors everyone. Founders who pitch below that line read as boring, even when the business is healthy.
Founder takeaway: The grandiosity-versus-credibility band has shifted up. You need a credible upside story that gets to a real position, told without sounding delusional. The cost of pitching “solid but capped” is a faster no than it was a year ago.
The Trillion-Dollar Land Grab: Everyone Is Rushing the Public Markets at Once
The stay-private era is over. Google announced an $80B equity raise, its largest in a very long time. SpaceX priced at $1.75T. Anthropic and OpenAI both pointed at roughly October. Add it up across those names and you get $300B to $400B of equity issuance, nearly all of it AI-related.
The framing that fits: it’s like an airline gate in a country with no orderly queue. The door opens and it’s a mad rush to board. Smart boards all looked up at the same time, saw the scale of capital required, and decided to jostle to the front. Google grabbed the first $80B. Why issue equity when you’re the second most profitable company on earth and could borrow? Because the stock is high, equity is cheap at an all-time premium, and a strong balance sheet insulates you from the micro-panics in the debt markets. You can also do both, layering debt on top.
The deeper signal is the business-model shift underneath all of it. These companies have gone from capex-light cash-flow machines to capex-heavy cash-consumptive machines. Across history, things that spin out cash are great investments and things that eat money are usually bad ones. On a trailing basis the numbers still look incredible. The forward picture is the open question.
Founder takeaway: When the best-capitalized companies in the world are all raising at once, the message is that the capital required to compete just went up an order of magnitude. Plan your own raise timing accordingly.
Is the SaaSpocalypse Over?
Best public software earnings week in two years. Salesforce, Snowflake, MongoDB all delivered and the stocks surged. The WCLD cloud ETF was down ~30% a month ago and has round-tripped 25 to 30% back. The catch: that only gets it back to flat on the year. For most of the last month software actually outperformed semis, but for the full year semis have murdered software to the upside because software just clawed back to even.
The rule that held up: you either reaccelerate or you attach to AI spend, and the winners did both. The losers with messy stories went down.
The clearest pattern is the split between software agents use and software only humans use:
- Twilio is up ~57% this year, growth gone from 4 to 5% back to ~20%, because agents and AI need more voice and messaging.
- Okta is up ~56%, because everyone doing identity is blowing up.
- Datadog is up ~100%, because every AI leader runs on it.
The human-per-seat model is dying. Not dead, but nobody wants to buy another seat for a person who doesn’t do work in your project management tool, and they’re actively cutting those seats to fund tokens. Gartner has AI software spend up 60% this year, which has to come out of somewhere.
Even Salesforce, which reaccelerated through Agentforce and hard work, told the market plainly: the classic software business will sit in perpetual single-digit growth, while the agentic side grows 12 to 13%. They split the company into two lines for the first time. That bifurcation is the whole story.
What’s gone is the freefall. The panic part of the selloff is over because the market realized these companies aren’t going to zero, and if they’re not going to zero they have cash-flow value. What hasn’t changed are the fundamentals. Atlassian sits around 4x ARR, HubSpot ~3.8x, Salesforce ~4x. The multiples bounced off the hard deck and went from worthless back to merely crappy.
Founder takeaway: The Jeff Lawson test is the one to run. Ask honestly whether agents and agentic workloads consume more of your product. If yes, you get lift, and the only thing that proves it is faster growth. If no, you’re not dying fast, but you have a value-creation and value-realization problem to solve.
Cognition Raises $1B at $26B, and the Autonomous-Engineer Vision Gets Real
Cognition raised $1B at a $26B valuation, with Devin at $492M ARR and some of the largest enterprises in the world as customers. For reference, Cursor sits around $3B ARR and is being acquired for $60B if the deal closes.
The vision here is more compelling than “your engineers are 1.1x more productive.” The interesting version is the autonomous AI engineer you task in Slack, that goes off and makes the commit on its own. Early on it was mediocre because the underlying models were. The bet is that the model curve fixes that. The logic is the same one Jason applies to sales: don’t empower the mediocre rep, automate the role. Devin doesn’t argue, doesn’t only want the interesting problems, and doesn’t turn down the boring edges of the app.
The risk is that this is a market where the lead changes hands at a furious pace. Every company worth a trillion dollars wants to eat your lunch. High risk, enormous market, exactly what it should be.
Founder takeaway: “More productive humans” is a feature. “Autonomous agent that ships the work” is a company. Know which one you’re actually building.
Kirkland & Ellis Commits $500M to Build Its Own AI. Build vs. Buy, Again.
Kirkland is committing roughly $100M a year for five years to build in-house AI rather than lean on Harvey or Legora. It reads as a big slight to those vendors. It probably isn’t as interesting as it looks.
Kirkland does ~$11B in revenue growing 20% and pays partners $11M bonuses. $500M over five years is less than 1% of revenue, likely reallocated out of some legacy software budget. It doesn’t preclude them from also buying third-party tools, and it doesn’t mean they’ll keep building if it doesn’t work. By announcing first, they already won the PR. They look AI-forward for committing 1% of revenue they may or may not fully spend.
The real signal is the old build-vs-buy rule, applied to AI. You buy horizontal products with no unique differentiation, where you’re not handing over secret sauce and can’t monetize it differently by owning it. Law firms have always bought case management, document storage, and research that way and competed on the relentlessness of their senior counsel instead. If AI stays in that bucket, they keep buying it and the drama leaves the deal.
The pause comes if firms believe the AI encapsulates their actual secret sauce. No matter what a vendor promises about not training on your data, an $11B firm’s managing committee is going to ask whether paying Harvey $10M to learn the “K&E way” is smart. And the model providers sit one layer below that with the same IP question.
Founder takeaway: If you sell AI into a knowledge-work vertical, the existential mistake is any hint that you might go full-stack and compete with your own customers. The day they think you want their book of business is the day they build their own.
Robinhood Lets AI Agents Invest, and the Real Opportunity Is Planning, Not Trading
Two very different problems get bundled together here, and they should be separated.
Financial planning is knowable and should be automated. The correct portfolio allocation relative to net worth, goals, and risk tolerance is well understood. Most of the job is extracting the right information from the client and then keeping them from doing something dumb. Wealth management is arguably the lowest-quality professional service most people ever pay for, with everyone funneled into the same models and the same in-house products. An agent that looks at your full picture and tells you the truth, “you don’t need more private equity, you’re carrying too much risk here,” is real expertise that should be available to everyone. Range, which Scale backed, does a version of this for the lower end of high-net-worth.
Beating the market is not there yet. The record on LLMs outperforming as a pod manager isn’t proven. If there were a durable edge, Jane Street and Citadel would already be quietly running it and telling no one.
The bigger idea underneath Robinhood’s move is the one Andrew Bialecki described at SaaStr AI Annual about Klaviyo at $1.4B in revenue. The most valuable agents they built weren’t for marketing or support. They were the agents that make every single Klaviyo user a true marketing expert from day one. That’s the aspiration for every application: a user logs in and is instantly an expert in your domain. The YouTube analytics agent is a good example, it tells you more about your video’s performance than any human could because it sees data you can’t.
Founder takeaway: Your product should make all of your users experts in your domain on day one. That’s the durable agentic moat, whether the agent educates the user or just executes for them.
Apollo Warns PE Software Returns Will Be Disastrous
Apollo is talking its book, and is probably also right. Private credit is the senior lender across a huge swath of PE software deals at ~5x EBITDA leverage. If the debt is struggling, the equity sitting below it in the stack is worse.
The math is harsh. Buy Salesforce at 14x revenue in 2021, half equity and half debt, and today the public market values it at five or six times revenue. You have some growth, but your equity is challenged. In a company growing 100% you can overpay and get bailed out. In one growing 20% that slows to 8 or 9%, there’s no accelerant and no magic. You just own a mature B2B company you paid too much for. The likely path is owning it for ten years, grinding out bolt-on acquisitions, and limping to a 1.2 to 1.3x.
The system depends on those GPs not getting to “move on.” As part of the management fee, you spend the next five to seven years finding an exit, because the gap between giving up at 0.5x and grinding to 1.5x is the whole job. Capital commitment is what keeps people in the chair. Skin in the game, and the desire to raise again, determines whether a manager fights or folds.
The flip side is the distribution wave. Menlo and Spark make ~$10B in carry from Anthropic, Founders Fund more than that from SpaceX. When that much cash hits a team, humans are humans and structures change. OpenView and most of Emergence wound down after generational outcomes, which were rational choices, not failures. The honest framing on huge carry: if a future fund can’t credibly return more than your big win, and you’ve already made nine figures, the marginal utility math points to retiring or, like Peter Thiel, taking a third of the fund so the LP economics make it worth showing up.
Founder takeaway: Ask your investors where their capital is committed and whether they’re playing a multi-period game. In a hard fund, that answer predicts whether they grind for your exit or quietly move on.
The 996 Work Ethic: Performative Theater or Startup Reality?
It’s both, and it’s not new. Jason’s first startup job, he showed up at 9am on a Saturday because every services job he’d had before ran six and a half days a week, they just didn’t brand it. The founder was already there.
The honest version: different jobs pay differently, carry different risk and intensity, and you get to pick where on that curve you sit. Startups have always run hotter than most companies. The Apple history is full of it, heart attacks and all. Sometimes hard things require a small number of people to concentrate around the clock and will something into existence. It is not sustainable for 50 years, and it isn’t a way to live forever, but it’s real and it’s old.
The two rules that matter:
- Drop the performance. The Cognition CEO explaining the Windsurf layoffs by saying plainly “we work seven days a week” was thoughtful, not douchey. Just hire the people who want that, pay up, give them four times the equity, and let everyone else go work somewhere saner.
- You better deliver. If you’re implicitly promising eight figures to your first 50 or 100 people, fine. A $150M exit does not justify that intensity. The quid pro quo is non-negotiable.
And watch the trap: when you’re cranking and stressed, you put in the extra ten hours but you can lose your judgment doing it. If part of your job is judgment, step back, take the walk, cut the grass. Otherwise you’re not rage-baiting, you’re rage-working. Performatively busy, not actually shipping.
The contradiction is the funny part. The entire Valley has a plan to automate white-collar work into mass unemployment in three years, and every single person you talk to says they’ve never worked harder and can’t hire fast enough. Both things can’t stay true. The likely resolution is the boring one: things get better, the “A” players can’t be hired fast enough and get paid more, and the “B” players have to accept that a moment of being wildly overpaid has passed. They still get jobs. Just maybe not the $400K work-from-home-three-days version.
Founder takeaway: Intensity is a legitimate strategy for your first 50 to 100 people if the equity is real and you deliver the outcome. Past that scale it stops being uniform. Don’t sell it as new, and don’t let it quietly erode the judgment you’re actually being paid for.
Quotable Moments
Jason Lemkin (SaaStr AI)
“I’m not interested if it can’t be a billion-dollar position anymore. That’s how Anthropic has changed my mind.”
“We’ve already made that choice at SaaStr. We got rid of all our B’s. We’d much rather have tokens.”
“I would quit as a developer if you told me I could not use the model of my choice. I would quit. It’s not worth my time.”
Rory O’Driscoll (Scale)
“Losing money is like sex. You can talk about it all you like, but until you feel it, you don’t know what it’s like.”
“All these businesses have gone from capex-light cash-flow machines to capex-heavy cash-consumptive machines. Generally that’s never good for stock prices over the medium term.”
“The weakest link in the chain determines the speed of the convoy. It may not be the amazing productivity lift you think, and therefore maybe the budget won’t be as big either.”
Harry Stebbings (20VC)
“We’ve just given a company credit card to every employee and said there’s no limits, spend away. And that’s the token-spend budget today.”
“We’re all here in the Valley with a plan to automate white-collar work, and yet you talk to every single person and they say I’ve never worked this hard. The contradiction at the heart of it is hilarious.”
“I have this feeling about you that you always instantly know the people you want to get rid of. And it kind of chills me a little. Which of us is going?”
