20VC x SaaStr This Week: Anthropic Is Eating OpenAI’s Enterprise Lunch, Figma’s Real Problem Isn’t Stitch, and the Unicorn Math Nobody Wants to Talk About

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

We’re back!

Ramp data put a number on what OpenAI has been avoiding: 73% of new enterprise AI spending is going to Anthropic. Ten weeks ago it was 50-50. That shift — plus a Figma stock rout, a $100B Bezos bet on AI-transformed manufacturing, a $20B Grok deal that cost founders roughly 60% in effective taxes, and the uncomfortable math on unicorn exits — made for a dense week.

The bigger story underneath all of it: we are past the point where enterprise buyers are browsing. They are locking in. And for the companies on the wrong side of that lock-in, the window is closing faster than anyone is willing to say out loud.

Top Takeaways

Anthropic Is Winning the Marginal Enterprise Buyer — and That’s the Only Number That Matters

OpenAI still has more total enterprise spend than Anthropic. That’s not the point. The point is the marginal buyer — the company making a new AI decision right now — is going 70% Anthropic. That’s the leading indicator. And OpenAI’s response to the Ramp data, snarking about “extrapolating from a lemonade stand,” was exactly the wrong move.

Ramp touches somewhere between 0.5% and 1% of US GDP in transactions. Their data scientists are good. And their customer base skews toward the digital companies that are making these decisions first. The data is real.

What’s driving it isn’t one thing. Claude Sonnet and Opus 4 through 4.7 have been a genuine step function since December — if you’re deep in vibe coding or building AI agents, you felt it immediately. More importantly, once a company has spent weeks dialing in an AI workflow on Claude — QA’d the outputs, built the scaffolding, trained it on their context — they are not switching. The soft costs of switching are enormous even when the hard costs (token pricing) look attractive. Anthropic has been consistent about what it is and what it’s building. OpenAI has been lurching: keep headcount flat, no double headcount; go deep on agentic commerce, Walmart says it doesn’t work; launch hardware, hardware gets deprioritized. That inconsistency has a smell to it now.

OpenAI still owns the consumer market. ChatGPT has the muscle memory. That’s real and shouldn’t be dismissed. But the enterprise coding market — which is the motherload app for enterprise AI spend — is locking in right now. If they let Claude be the default for another six to twelve months, they’ve sacrificed value they won’t get back.

The Real Figma Problem Isn’t Google Stitch

Stitch is a proof of concept from a company that abandons most of what it launches. The odds Google decides to compete with Figma for a decade approach zero. The market overreacted. That’s the first-order read.

But the market isn’t wrong. It’s just reacting to the right thing for the wrong stated reason. The actual worry isn’t Stitch. It’s that Figma Make — Figma’s own AI product — is among the worst vibe coding tools available. It can’t pull context from a live website. Every other tool in the space, including Replit, Lovable, and yes, Stitch, can do this now. A company doing $35% growth, with an enormous installed base and a smart CEO in Dylan Field, somehow has an AI product it isn’t even charging for because it isn’t good enough to charge for.

That is the diagnosis. And the installed base is part of the cause. At scale, the existing customer base becomes a resource trap. It demands constant attention — 50 years of features, offline integrations, non-agentic gaps. If you’re not deliberate about ring-fencing resources for the new thing, the old thing consumes 98% of what you have. Intercom had to consciously let its core business enter partial decline to build Fin. That’s a hard decision for a private company. For a public company headed toward IPO, it’s almost impossible without enormous conviction at the board level.

The market’s test is simple and harsh: are you charging for your AI product? If no, you’re not an AI company yet. And right now, almost no public SaaS company is passing that test.

What It Takes to Actually Pass the AI Monetization Test

There are two signals worth tracking. For SMB products: is ARPU 50% higher than pre-AI? Notion appears to have done this — their AI tier is $20/month versus a $10 base, and they’ve reportedly doubled ARPU. That passes. For enterprise: is revenue reaccelerating? Something has to be moving — new product attach rate, ACV expansion, net revenue retention. If nothing is moving, AI is a feature, not a business.

Token costs are a red herring for most applications. Yes, there’s an open router world where cost-sensitive builders are switching between Kimi, Haiku, and Mini to optimize spend. That market exists and is real. But there are many applications — AI agents doing sponsor management, marketing ops, customer success — where token spend is $2,000 a month and the delivered value is 100x that. For those applications, the idea of saving $500 a month by switching models isn’t a business decision, it’s noise. The companies building in this zone are not going to churn off their tuned, dialed-in Claude setup for a marginal cost reduction.

Jeff Bezos Wants $100B to Do What Amazon Did to Retail, But for All of Manufacturing

The clearest way to understand this: when Amazon was starting, there were three ways to bet on the internet transforming retail. Build software and sell it to retailers (Shopify, $200B outcome). Buy an existing retailer and AI-transform it (Walmart, roughly a 2x on a lot of capital). Or build the full-stack retailer from scratch and kill everyone (Amazon, $2T from zero).

Bezos did the third. Now he’s 60, he has more money than time, and he doesn’t want to build from scratch again. The Bezos $100B fund is the Walmart play for AI and manufacturing: buy existing companies across semiconductors, space, and defense, inject AI, and capture the transformation upside without the 25-year grind. It’s less IRR but faster and more comfortable from Indian Creek Island. It’s what you do when you've already proven you can do the hardest version.

SpaceX at $2 Trillion: How You Build the Math

The TerraFab announcement — effectively building a chip fab at 70% of TSMC’s capacity for roughly $25B CapEx, primarily to serve SpaceX’s data center and Tesla needs — is the latest step function in a pattern. SpaceX’s operating model isn’t linear growth. It’s big, chunky technical milestones every five to seven years, each of which unlocks a new layer of TAM. Government launch contracts. Starlink internet. Remote cellular. Data centers in space. Now: chip manufacturing to support all of it.

Starlink’s profit margins are reportedly exceptional. If you believe that’s the baseline and TerraFab extends that flywheel by two orders of magnitude, a DCF case for $2T can be built on a spreadsheet. Polymarket put the probability of that valuation at IPO at 50-60%. Tesla stock didn’t move on the announcement, which is meaningful — it suggests the market sees this as SpaceX value, not shared value.

The Grok/NVIDIA Deal: $20B for Sub-$100M Revenue

When does someone pay $20B for a company doing under $100M in revenue? When the strategic value to the acquirer justifies it. NVIDIA has a $5T market cap and just announced Grok is going into production. The value isn’t in the current revenue, but in the strategic integration into the AI infrastructure stack.