20VC x SaaStr: The Token ROI Crisis Comes for Everyone, Anthropic Wants Chinese Open Source Banned, and Microsoft Has Its Worst Month Since 2000

With Harry Stebbings, Jason Lemkin, and Rory O’Driscoll

We recorded this on the last day of June, and one theme ran through almost every story: Companies quintupled their token spend in the first half of the year. Almost nobody can point to the revenue lift that justified it.

That’s the real state of the AI era right now, even with all the incredible things we’ve all shipped, built and analyzed. And it’s about to stress even the best B2B companies, not just the strugglers.

Here’s what the three of us got into this week.

1. The Token ROI Crisis: (Almost) Nobody Can Clearly Show the Lift

The trigger was Brian Armstrong’s Coinbase post: AI spend down 50% this quarter, usage actually up, driven by a shift to open-source models and away from frontier usage.

The 50% cut only brought spend back to roughly November levels. Everyone ramped aggressively in November and December when agentic coding exploded. It took about five months for cost discipline to catch up, and now it’s getting clipped back down.

At SaaStr Fund we have a portfolio company where every number is green. Hit the first-half plan. Overloaded with investors. You would love the dashboard. They came to the last board meeting wanting to double their token spend, and it was enough to move the burn from no-big-deal to big-deal. For the first time, the board tied approval to ROI, and this excellent team couldn’t draw the line from the spend to the curve. The velocity is there. Everyone wants to invest. It just didn’t tilt the revenue line.

If you’re a software company, AI spend on engineering should credibly produce revenue lift. You’re making more of the exact thing you sell. If you’re not a pure token reseller and you still can’t show the lift, or at a minimum the savings, you’re going to start looking at that spend with a very jaundiced eye. It’s time for the next mature phase of token spending. It’s time to grow up.

2. Software Is Either Accelerating or Irrelevant

This is the principle we’ve been consistent on for a year, and Rory paraphrased it back to me this week: software companies in the age of AI are either accelerating or irrelevant.

The Coinbase post gets a partial pass from Rory because he lumps them in financial, not software, and crypto doesn’t get an automatic AI lift. Fair. But apply the principle to a pure software company and it holds. Adobe announcing “$500M of agentic revenue” and then missing the quarter is performative too.

The counterexample is Aaron Levie. Some of Box’s AI messaging runs a little hot. But he ties every bit of it to revenue and the business model. Here’s how it works at Box, here’s how we process documents and content, here’s how it maps to the P&L. It hasn’t made Box a 100% grower. It got Box back to double-digit growth. That’s the whole difference. Tie it to your business, or it’s just noise.

3. Is Open Source Killing Frontier Model Revenue Growth?

Anthropic went from a $1B run rate at the start of last year to $9B by year end to $44B mid this year. The Coinbase move sharpens a question: does the rise of open source cannibalize the path to a trillion in revenue faster than anyone modeled?

Rory’s framing is the right one. The bulk of tokens generated may well shift to open-source models. The bulk of the revenue still comes from state-of-the-art frontier models. There’s clearly a very big business here regardless.

The trap is structural. If you’ve built a cost structure and a capex commitment that only works at a trillion in revenue, the last thing you need is cheap open-source alternatives at a fifth of the price. Land at half a trillion, still the largest digital company on the planet, and it’s a consolation prize that breaks your model. Nothing in the Coinbase memo says these won’t be great companies with real differentiation. It just says the pricing structure some of them are counting on may not hold.

4. Anthropic Wants Chinese Open Source Banned, and Why That Would Be So Dumb

Anthropic wrote to the Senate Banking Committee alleging that Chinese open-source companies are bootstrapping their models by sending millions of prompts to Anthropic, recording the answers, and training on them. Distillation. It’s a clear breach of terms of service, and arguably touches copyright and trade-secret law.

You have to power through the irony first. Every foundation model, Anthropic included, was trained on other people’s IP. Anthropic recently settled with a large group of book copyright holders. So being appalled when someone does a version of it to you is a particular kind of position to take.

Set the irony aside and look at the play. A terms-of-service breach is a contractual problem between two companies. Good luck with that lawsuit in Beijing. The reason the Senate Banking letter is interesting is that it’s an attempt to turn a contractual dispute into the US government putting its thumb on the scale. You can see where this goes: legislation that says no Chinese model proven in a US court to have distilled on US foundation-model technology can be used by a US company.

The fallback win is quieter and more achievable. They can’t stop Cursor and Harvey directly. What they can do is make every Fortune 500 company nervous enough about security to ban open-source use internally. Startups cut corners enterprises won’t. Make it look dangerous, and enterprise procurement does the rest.

The problem is two separate issues getting brilliantly conflated. One, did they do the naughty thing with distillation? If so, pay the naughty tax, exactly like Anthropic paid the book publishers. Nobody’s banning Anthropic or OpenAI for how they trained. Two, is a Chinese-origin model running on US inference an actual national-security risk? If the weights are open, the code is inspectable, and there’s no telemetry going back to China, there’s no real danger. Conflating the naughty tax with the security scare to protect a capex position is regulatory capture in the extreme.

Rory’s analogy is the one to remember. It would be like the US government in 1982 banning Compaq and Dell to keep IBM’s stock price up, and keeping the PC industry a small, profitable business for IBM alone. The loser, as always with trade restrictions, is the rest of the economy, who get dear intelligence instead of cheap intelligence. The AI stocks in your 401k keep making money. The stocks that should be getting the lift from cheap AI lose it.

My counter is that we might do the dumb thing anyway. Forty percent of the S&P is tied to this bubble. Nobody wants their 401k touched. AI has become the oil of the Persian Gulf. We’re all addicted, and I think we do everything we can as a society to prop it up, whether or not that’s the smart long-run move.

5. Microsoft’s Worst Month Since 2000

Microsoft is down about 16.5%, its worst month since 2000. Ironically it happened right after a big Satya announcement on their AI direction, which may be exactly what opened people’s eyes.

The core issue is the one we’ve been flagging for a year. Microsoft doesn’t have a standalone frontier model. The AI story is capex to support OpenAI plus a 30% ownership stake in OpenAI. Meanwhile the two historic pillars of the franchise are under direct attack. Cowork is eating the individual knowledge-worker seat. Claude Code is eating developers, developers, developers.

Then Azure guided deceleration from roughly 40% to 37%. In a market that only rewards beat-raise-accelerate, guiding down three points is a fail even at that scale. And a big slug of Azure growth is simply inference for OpenAI, so it isn’t a compelling end-customer growth story that Microsoft owns itself. If we’re all supposedly about to run 20 agents 24/7, Azure should be accelerating, not decelerating. That’s the canary.

Contrast with Google. Three years after the “make Google dance” comment, Google has massively outperformed, because for all its risks it at least has its own model and its own product to sell. Microsoft has 30% of OpenAI and no state-of-the-art model of its own. That’s a real gap.

6. Kalshi at $40B and the Casino Economy

Kalshi is raising at $40B, up from $22B in May, on about $2B in revenue, roughly 70%+ of it sports betting. People like to bet. The US banned it for decades, and now it’s a couple-hundred-billion-dollar industry doing the same thing.

Does Kalshi become a $100B company in 12 months? I doubt it. You get there one of two ways. Either sports betting keeps expanding and they take outsized share, or the non-sports side, crypto perpetuals and financial prediction, turns out much bigger than we realize.

The tell that it might is that Intercontinental Exchange, which owns the NYSE and runs real financial markets, took a roughly 20% stake in Polymarket. That’s a serious operator signaling this is a $50B to $100B category.

Predicting the next president is fun and it’s my favorite part of these products, but it’s a small market. The big markets are the primal ones. People love sports. People love money. As Rory put it, if you could bet on sex, you’d have the trifecta. I did a deal into FOMO, which just raised from Index and USV. Their Perks product lets consumers bet up or down on stock prices in the simplest possible way. That’s a brilliant way to expand a TAM into a mega market.

7. Did SpaceX’s IPO Freeze the AI IPO Market?

The question was whether SpaceX’s IPO volatility makes OpenAI and Anthropic nervous about going out. The evidence says the window isn’t shut. Bending Spoons went public July 1 at a $20B valuation, and they own AOL and Evernote. That’s the anti-AI IPO, and it got out fine.

As long as the window stays open, the frontier labs will be fine. But the volatility has been high enough that Dario and the board are almost certainly thinking about timing weekly. Anthropic looks perfectly linear and up-and-to-the-right while it’s private and raising rounds. Once it’s public, a hint of bad news moves it hard. If Microsoft can have this kind of month, the frontier labs will be some of the most volatile stocks out there.

8. Bending Spoons IPO at $20B+, Trades Up 40% in IPO

Bending Spoons went out at $20B on about $1.5B in trailing revenue, with $600M in Q1, so call it 8 to 9x forward. That’s a healthy multiple for the antithesis of an AI company.

Read the S1 and you see the model. They don’t get organic growth. They don’t add new users. They buy tired consumer products growing at 10%, raise prices, cut costs, and optimize. Slow-growth single-product B2B trades at 2 to 3x revenue. Constellation Software, a roll-up of B2B companies, trades near 9x. Bending Spoons is running a version of that arbitrage on consumer.

Is it frothy? A little. But it’s well-executed, and value is clearly being created. If they’ve genuinely identified a thousand material targets, they can maintain outlier growth for five-plus years, and outlier growth earns a premium. I’d bet it does well over the medium term.

9. Bending Spoons for B2B: The Roll-Up Nobody’s Built Yet (Even Though There are Tons of PE Plays)

Here’s the one Harry actually wants to see built. A Bending Spoons for B2B. Buy the horrific, sticky, decaying B2B products, put a real operator in, and boost net revenue retention.

Our worst and most expensive product today is Marketo. The API breaks. It hit us with rate limits. They threaten us, they told us they’re raising prices 20% next year, and the site went down for a full day recently. Now take whatever Marketo-style revenue is left, say $300M that’s slowly decaying, put a kid in who actually gives a crap, launch some features, pull the rate limits off the API, stop threatening the customers, and retain the base. It’s not that hard. Stack ten of those at $200M each, growing 30 to 40%, and you’ve built a $2B business.

The target list writes itself. Marketing automation and Marketo. PagerDuty, which has commanding share and 15,000 customers but never shipped AI-enabled incident resolution. Semrush, where every customer obviously needs GEO and AI optimization next, though Adobe already bought it for under 2x revenue, which was a great deal. Asana, where the billionaire founder up and quit 18 months ago.

The catch, and Rory is right about this, is that it’s harder than Bending Spoons. Bending Spoons just raised prices and optimized on consumer products that needed almost no innovation. Buying a pre-AI B2B company in 2026 and only pressing the Vista cost-cutting buttons won’t be enough. You have to re-engineer the company with AI to generate genuinely new revenue. Bigger managerial task, bigger upside.

The real problem is the cultures, not the products. These teams have given up. Customer success has become a force of evil that threatens you with lawsuits. It reminds me of a guy I once had on my sales team whose prior job was at the Yellow Pages, where he got a big bonus if his patch shrank less than 20% a year. At some point you have to leave that job. Turning around a team that’s mentally checked out isn’t hard if the base is sticky, but you have to find people who want it, and the PE playbook keeps installing recycled 2021 executives who go on a 90-day listening tour and come back with no ideas. When Bending Spoons buys something, stuff happens in the first 30 days. People move out, people move in, products get shut down. That’s the level of action this requires.

10. Chamath Goes CEO of 8090

Chamath raised $135M for 8090 and took the CEO seat. It’s a software-factory platform for the full development cycle, with governance. Credit for getting in the arena. It’s a genuinely exciting market with a ton of competition.

My skepticism is general, not personal. I have too much scar tissue writing small checks into successful founders doing multiple things “just for fun,” and watching them all go to zero. When wealthy VCs decide to be CEO but aren’t grinding every waking hour, they run out of energy when the going gets hard. It’s all fun in the early days: pull a team together, use your brand to line up partners, whiteboard the vision. It stops being fun when the real work starts. Show me the paunch, the hair loss, the right-hand person quitting on him. Then I’ll believe it’s the only thing he’s doing, and then I’d invest. It’s just too easy to start up today. Seeds are for suckers.

11. Why $1.5M to $5M ARR Isn’t a Series A Anymore

Harry turned down a founder this week finishing the year at $1.5M ARR and next year at $5M, and tweeted that it isn’t good enough to raise a good Series A because the opportunity cost of cash is real. He deleted it when it only got 30 likes, then reposted it, and it became a thing.

As a factual statement about today’s venture market, he’s correct. Those growth rates would have been top-quartile in the age of B2B SaaS. They’re the exception now in the age of AI, and you’d need some other extenuating factor.

There are actually two tweets buried in one. The state of the Series A market, and the opportunity cost of cash. The second is the more triggering claim, and it’s the one people reacted to.

Here’s the part I care about most, because it’s a subtly toxic thing a lot of nice VCs do. I’ve watched two capital-efficient portfolio companies get told “go have a go” at a round that everyone around the table knew wouldn’t get done, instead of getting honest feedback in the boardroom. I run their decks through the AI pitch-deck analyzer I built on the actual benchmarks, and it tells them straight that they’re a B for that round, not an A. Not one of their existing VCs would be that honest with them. Sending founders out to collect a harsh message from the market, just because you’re too conflict-averse to deliver it yourself, is a pretty pathetic act.

The nuance Harry’s tweet left out is that $1.5M to $5M can still become a generational company. The correlation between initial growth rate and final outcome is real but modest. Procore was a slow grower and became a huge outcome. What that founder should hear is: nothing’s wrong with your numbers, put it on a spreadsheet, if your burn is low and you don’t quit you can build something great. But cut your cloth accordingly. You may have to meet 150 investors instead of 15. Don’t run a fast process, don’t build a data room, give people a week and ask for checks. And you may not be a venture asset in the new world, which is completely fine.

Meanwhile Higgsfield, which I invested in at seed and Harry did too, just crossed $500M in revenue in under 18 months. That’s what everyone’s actually hunting. When I first mentioned Higgsfield on the show, nobody had heard of it. That’s the point. Go find the next one.

The one real mistake Harry made was deleting the tweet. That looked like blinking.

12. Claude Tag in Slack: Trojan Horse or Nothing?

Claude Tag puts Claude in a Slack channel as a fully present, autonomous member scoped to a function, your legal team, your data team, whatever, running across platforms including Salesforce and HubSpot.

The existential version is real. If the agent runs 24/7, pulls in all your data, builds the analytics and the dashboards, and operates autonomously, then your data flows between apps and you stop caring where it lives. Every fear about headless comes true, because Claude becomes the head, and Salesforce and HubSpot become dumb databases. If Anthropic puts its best people on this and doesn’t quit, it could be the biggest deal for traditional software there ever is.

The skeptical version is also real. Is it Zapier on steroids? Is Slackbot already better? We tried to deploy Tag for the show and weren’t on the right enterprise plan, so I haven’t run it. Same as Claude Design and Figma: the internet declared the world over, and there’s no evidence yet that it’s an actual Figma killer. Give it a week before declaring anything.

Why does Salesforce allow it at all, when they own Slack and have their own Slackbot? Because you can’t be the cross-platform comms layer for a company and then block the agent that helps people do better work. Rory’s point is the sharp one: Slack is where the context lives. All the weird stuff people do on top of the apps, the exceptions, how Jason and Rory actually handle a given problem, that’s the context graph, and capturing it is exactly what lets an AI automate the work. So it’s a genuinely interesting entry point. Salesforce is right to make sure Slackbot stays better.

The bigger frame is the one that reframes everything. By the end of this year, Anthropic may have more revenue than every public software company combined. Materiality has always been roughly 10%, that’s basically the old SEC disclosure line. Anthropic may not get out of bed for anything under $10B of revenue. It may simply not be worth Dario’s time to worry about disrupting Salesforce’s $42B, or even Salesforce’s ~$8B of net new bookings. That’s probably why the Figma-board drama read the way it did, with Anthropic essentially saying “oh, we didn’t realize you’d care.” When elephants dance, the little people get trampled.

Quotable Moments

Jason Lemkin

“I’m getting burned out on struggling CEOs sharing performative AI data when they’re not AI companies. Show me the money. Show me the revenue growth.”

“It’s time for the next mature phase of token spending. It’s time, boys, to grow up.”

“I don’t know if Anthropic can get out of bed for less than $10 billion of revenue by the end of the year. Ten Percent. That’s always been the definition of materiality.”

Harry Stebbings

“Turned down a founder this week. Finishing the year at $1.5M ARR, finishing next year at $5M. Brutal as it is, this isn’t good enough to raise a good Series A. The opportunity cost of cash is real.”

“I’m really sorry if that’s going to ruin your day. Don’t be a founder. Life’s harder than a VC tweet ruining your day.”

“You’ve got Sam and Dario both advocating for it and the people around the administration advocating for it. Isn’t a ban almost inevitable, not just plausible?”

Rory O’Driscoll

“If you can be the largest tech company on the planet and still not make money, you might have oversized your ambitions a little. It might pay to come back a bit.”

“The loser, as always with trade restrictions, is the rest of the economy, who won’t get cheap intelligence. They’ll get dear intelligence.”

“People love sports. People love money. If you could bet on sex, you’d have the trifecta.”

This post is part of the ongoing 20VC x SaaStr collaboration with Harry Stebbings and Rory O’Driscoll.