20VC x SaaStr: Anthropic at $900B+, Salesforce Spends $300M on Tokens, SpaceX Files for the Biggest IPO in History, and Why the Tech Lash Is Just Beginning

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

Anthropic is closing $30 billion at a $900 billion valuation. That’s nearly triple the $380 billion price from February. Andre Karpathy joined Anthropic the same week. Cerebras priced its IPO at $185, popped 68% on day one, and broke $300. SpaceX set June 12th as the date for the largest IPO in history at a suspected $1.75 trillion market cap.

Meanwhile, Marc Benioff casually mentioned on All-In that Salesforce will spend $300 million on Anthropic tokens this year, almost entirely on coding. That works out to roughly $15,000 to $20,000 per engineer per year.

And on the other end of the AI boom: Cisco cuts 4,000 jobs, LinkedIn cuts 875, Meta cuts 8,000, Intuit 16,000. Eric Schmidt got booed at a graduation speech. Three years ago, kids were applauding ChatGPT at the same kind of ceremonies. The political direction of travel is exactly one way.

This is the most consequential week in AI venture in months. Here’s what it means.

Top Takeaways

1. Anthropic at $900B Is the Best Trade in Venture Right Now

If you take Anthropic’s June run rate, $900 billion is roughly 18x revenue. Compare that to hot seed and Series A rounds in AI for $10 million ARR companies trading at 30-40-100x ARR five years away from an IPO. Anthropic is so far post-IPO scale that there’s no IPO risk, no “it will go away” risk. The only risk left is valuation risk.

“Every time the multiple on this has been significantly lower than the multiple on your median Series A or B or Series C, for a higher growth rate,” Rory said. “It’s been the best trade out there.”

The flip side: why does Dario Amodei do this if it’s such an obvious deal? Because giving away 3% of the business to derisk another year of monstrous burn is rational when you’re committing to 5-6 gigawatts of compute this year at a rule-of-thumb cost of $40-50 billion per gigawatt. You’re not spending that capital directly. You’re persuading hyperscalers to spend on your behalf. But the balance sheet war is real.

There’s also a philosophical difference between Anthropic and OpenAI on fundraising. Sam Altman pushes valuations to the absolute maximum because his thesis requires infinite capital. Dario does the opposite. The last two Anthropic rounds appear to have intentionally traded at a discount to get done low-drama in days. The Anthropic philosophy: “If I want to raise $30 billion, I should probably get a check for $30 billion and call it a day.”

Compare that to OpenAI’s last round: Amazon gives $50 billion, but $20 billion is up front and $30 billion is contingent on going public. SoftBank gives $40 billion, but $30 billion has to be borrowed. “It’s like, give me a break.”

2. The Salesforce $300M Token Bill Is Both Nothing and Everything

Mark Benioff dropped on All-In that Salesforce will spend $300 million on Anthropic tokens in 2026, almost entirely on coding. Sounds enormous. Do the math.

Salesforce has roughly 20,000 developers out of 83,000 total heads. That’s $15,000 per developer per year, or $1,200 per developer per month. Salesforce’s total engineering wage bill is around $5.8 billion. So $300 million is roughly 4% additional spend on top of fully-loaded engineering cost.

Rory and Scale ran a survey across 40 portfolio companies and external companies. The average came in at $1.2K-$1.3K per developer per month. The median was lower. Salesforce is in the strike zone of normal. It’s only big because they have so many developers.

But this is the only product Salesforce buys that comes anywhere close to this amount. If Salesforce bought a brand new ERP system from SAP, they’d maybe spend $5-10 million. Anthropic tokens are the largest single external software spend at most enterprise software companies, period. From nothing two years ago.

The real question is the trajectory. To justify Anthropic’s $900B valuation and OpenAI’s path to a trillion dollars of token revenue, the math requires something like 5-7% of every knowledge worker salary and 20% of every engineering salary getting converted to tokens. Salesforce at 4% is maybe a quarter of the way there.

One of two things has to happen: either they stay at $300 million (in which case the TAM for token businesses is overestimated and there’s a real correction coming), or they 4x their token spend from here. In two years Benioff is on All-In saying “we spend a billion dollars on tokens.” Pick a side.

3. The Klaviyo Counterpoint: Most Companies Massively Overestimate Token Costs

An underrated moment from SaaStr AI 2026 last week was Andrew Bialecki’s session. Klaviyo requires every employee anywhere close to product to commit code. Every single person has to be running AI or agents to do their job. 100% adoption. They built their own custom agentic framework to enforce it.

Backstage, Jason asked Andrew how much he thought it costs to run an AI VP of marketing. Andrew guessed $250 for both. The actual answer was $257 combined. He was the only person who got that number right, because he’s actually doing it.

His point: “A lot of this stuff is not as expensive as we think. We do not need to worry about token maxing at Klaviyo, and everything we’re doing is agentic. The most junior product person, every single person has to be doing this. But if you do this right, it’s not as expensive as people think.”

The implication: there’s a bear case where the models get better and more efficient, the orchestration gets smarter, and most companies need a half or a third or a quarter as many tokens as they think they do. The Mark Benioff number is roughly right as the bear case for companies not at the bleeding edge of token-maxing. But it’s not the floor that some people assume.

4. Public B2B Has Hit Bottom and Is Quietly Reaccelerating

The bounce off the SaaS apocalypse lows is real but uneven. Figma reaccelerating to almost 50% growth. Atlassian reaccelerating north of 30% on the back of Rovo. Datadog hitting its first billion-dollar revenue quarter at 32% growth. Twilio coming back from the dead to 20% growth.

The meta question: is there more time than we thought?

“The whole world is not in San Francisco of all the buyers, of all the users,” Rory said. “Maybe Andrew at Klaviyo and Mark Benioff and these founder-led companies have enough time. Maybe another year, if they’re just getting going on their agentic journey.”

But the bigger insight: every tech in cycle has a set of industries that, once in their life, get valued on prospects and futures. Five years ago, that was B2B SaaS. Today, it’s AI. Once you lose that veneer, you’re valued for the rest of your life on revenue, growth, and cash flow. You almost never get back to 50x ARR.

Figma reaccelerating, Datadog crushing, Atlassian inflecting. These are wins. They’re just bounded wins. Datadog at 17-18x sales is the outer edge of good. Figma at 6-10x sales is the standard for good. Wix at 1x is the floor.

5. The Figma Vibe Coding Miss Cost Them $500M+, But the Captain Obvious Point Was Bigger

The honest founder take: Figma Make is the worst vibe coding product anyone on this pod has used. It was clearly not important to senior management. They left $500 million or more on the table by not building a Replit-level competitor.

But the captain obvious point that got missed: Figma is in the business of making it easier to build software. The software explosion is happening whether or not Figma owns vibe coding. Even though they lost the vibe coding race, they benefit from the software explosion the same way Work OS and RevenueCat have exploded because there’s a software explosion going on.

What Figma still hasn’t built: the workflow where you create a design, loop in the product team to approve it, and then click one button to push it into a full production prototype. Replit and Lovable both have “upload a Figma design” as a native insertion point in their prompt UI, because they know that’s exactly what their ICP wants to