Meta's AI Agent 'Muse' Adds $100B in Market Cap in a Single Week

Source: 20VC with Harry Stebbings | Published: 2026-09-24T13:57:48Z

Meta surged ~$100B in market cap within a week of launching Muse — a free AI agent with more generous token limits than ChatGPT, genuinely usable agentic capabilities, and a built-in apology letter feature when it makes mistakes.

He uses it for exactly one thing: deciding which two people in a startup community should meet.

Jev: A Model That Returns No Text, and It's Eating LLMs' Lunch

Jev is, at its core, a classifier. It doesn't generate text — it returns true/false, rankings, or scores. System 1 from Daniel Kahneman's Thinking, Fast and Slow: fast, cheap, precise. Because output tokens are minimal, it doesn't charge for output at all.

Rory ran the numbers: global LLM spend is roughly $100 billion, and about 20% of those calls don't require complex reasoning — just a simple judgment. That's a $20 billion market Jev is attacking at one-hundredth the price. Then he immediately walked it back: even if they capture it, prices will collapse. That $20 billion in a price war might become $4 billion, maybe less. It's Jev's market to take, but the pie shrinks as they take it.

Anthropic and OpenAI's position is telling. Jason was blunt: Mini and Haiku are the worst models he's ever used. Both major AI labs have strategically abandoned the low end — they're more worried about open-source catching up than a small classification company.

Model Selection Is Getting Harder, Not Easier

Jason's weekend revealed a practical dilemma.

He switched Replit to auto-routing mode — dynamic model selection across providers. Performance degraded. He switched everything back to Claude. Now Jev comes along, adding another layer. When do you use Jev? When do you use Claude? Where's the line?

"Can you QA ten thousand Jev calls?" he asked. The answer is no. Using Jev for real-time routing — having it decide in milliseconds which model to call — sounds clever, but Jev's error rate is around 20%. If you're routing calls for critical feature code, 20% wrong routing means you're fighting a system riddled with bugs.

His conclusion: model complexity is growing exponentially, but the engineering infrastructure to manage it hasn't kept up. Every team serious about AI applications will need a dedicated team of engineers running evals and managing model selection — and that team is only going to get larger, not smaller.

Meta Muse Hit #1 on the App Store, and It's a Trojan Horse

The most significant product event of the week was Meta's Muse.

Jason did one thing with it: built a complete CRM. He used Muse to track 150 sponsors, pull in every relevant email in real time, and organize all interaction history. Within two weeks, the system was running. Cost: zero. It needed a database, intelligence, and email access — Muse assembled all the pieces and delivered.

Why does this matter? Anthropic never really built Claude as a consumer product. ChatGPT's autonomous agent capabilities are limited. Muse does both: it's a capable general-purpose LLM and a genuinely autonomous agent. For ordinary users, it's free, with more generous token limits than ChatGPT, and it writes you an apology when it makes a mistake.

Meta's market cap rose roughly $100 billion this week on the back of this product. Stock up 34% in a single month. Rory noted the figure was a reminder: when you're addressing a truly large end market, a single product can generate value at this scale.

Amazon Blocks, Shopify Opens Up — Both Decisions Are Rational

Muse's agent capabilities let it purchase things on behalf of users. Amazon announced it would block this. Shopify chose to partner. The logic behind both decisions isn't contradictory at all.

Amazon's core profit doesn't come from merchandise — it comes from advertising. Ad revenue now exceeds e-commerce profit. Agent-driven purchases bypass the search page: no "related product" recommendations, no ad impressions. On top of that, Walmart data found that agent shopping compresses basket size — users buy exactly what they need and don't add three extra items to the cart. Amazon's calculation: block it, because consumers will come back anyway. Because it's Amazon.

Shopify is different. It represents a long tail of merchants — another traffic source is another dollar of revenue, and Shopify still takes its cut at checkout.

Jason's take was sharper: these platforms will all be "maimed" — not killed, but weakened. Agents will find every exposed API and route around every intermediary layer they can. He gave an example from that morning: a key supplier suddenly raised prices significantly. Their agent immediately mapped out a response and generated a 12-month migration plan. Rory added: even if it only hits 10% of Amazon's ad revenue, that's a serious problem.

Enterprises Don't Trust Anthropic and OpenAI With Their Code

Factory closed a new round this week at a $5 billion valuation. It's an enterprise AI coding agent company. Their flagship product is called Droid.

Jason spent time at Dreamforce talking with C-level executives and got one consistent signal: they don't trust Anthropic or OpenAI to hold their data. This isn't manufactured hype on X — it's real enterprise decision-making. Upload proprietary code to a frontier model, and nobody can give a satisfying answer to whether that data gets used for training or ends up benefiting a competitor.

Jason's memory is specific: when he was an executive at Adobe, source code was a core strategic asset — treated with code-red-level protection. The team wanted to adopt GitHub, and it took countless security audits and high-level debates before it was approved. That same concern is now spreading through enterprises even faster. His conclusion was blunt: they trained on everything — every YouTube video, every open-source repo, every piece of closed-source code. Of course they'll train on yours too.

Rory framed it in a larger context: coding is the most value-dense area where AI is delivering returns right now. Investing anywhere in the coding stack is a reasonable bet. Factory's value proposition is simple: you get the capabilities of frontier models while keeping your code in an environment you control.

Anthropic Pushed the IPO to November, and the Bankers' Logic Is Clear

Anthropic moved its IPO from October to November. Some read this as cracks in the market.

Rory explained it from a banker's perspective: they had a strong Q2, Q3 is probably better, but IPO in October and Q3 numbers can't be included in the filing. IPO in November and Q3 goes into the prospectus — a clean story told by the numbers. October requires a lot of explaining. November lets the numbers speak for themselves.

The timing decision itself sends a signal: Anthropic believes it has time and the position to wait — it doesn't need to rush out before the window closes. Rory puts the odds at 90% correct. The remaining 10% is the risk that if something unexpected hits markets in November, they'll wish they'd moved in October.

On the insurance question: no insurer is willing to underwrite Anthropic. Rory's answer was straightforward: a company worth $2 trillion can self-insure — it doesn't need Munich Re to backstop it. And anyone buying the stock who hasn't noticed that Dario himself has acknowledged he may be building one of the most dangerous things in history simply hasn't read the public statements.

OpenAI Will Burn $278 Billion, and It Might Not Be Enough

OpenAI's projections call for net cash consumption of $278 billion through 2030. Current balance: $122 billion. Already an enormous number. But Rory pointed to a more staggering figure: the total capex required to get there is roughly $700 billion — most of it off OpenAI's balance sheet, carried by Oracle, Nvidia, and others through lease agreements.

"AI isn't software. It's an extraordinarily capital-intensive business." Growing from $35 billion ARR to $350 billion requires directly or indirectly consuming $700 billion in capex. Intelligence is not cheap.

Jason's addition was more grounded: this company has never come in on budget. Historically, it runs 30% to 50% over projections every time. What they actually need may not be $278 billion — it may be $400 billion.

Seed Rounds Are Getting More Expensive — Some of It Is Inflation, Some of It Is FOMO

Jev raised a $40 million seed round. Some said it's nothing — any decent company coming out now starts at $20 million.

Rory offered a purely mathematical explanation: from 2010 to today, nominal GDP is up roughly 2.5x. Software engineer salaries have kept pace, at minimum. So a $4 million check written in 2010 represents roughly $10 million in real purchasing power today — and that's before any market premium, just raw inflation. Add in the growth expectations of today's AI market, and $20 million isn't surprising at all.

But Jason and Harry both noted where this logic breaks down: the big wins of past cycles came from getting in when prices were relatively cheap. The same companies today might be valued three to four times higher. More expensive entry means less room for error. Rory quoted a line attributed to Baruch — "if every day you look in the mirror and remind yourself that two plus two equals four, you can avoid a lot of mistakes" — to make a simple point: the obvious things are the easiest to overlook, right up until they bite you.

a16z announced a $40 million "university" program this week focused on pre-inception investing — getting in before founders have an idea. Peter Thiel did the same thing 18 years ago. a16z is just doing it at scale.

Data Centers Are a Good Business, But They're Highly Leveraged

Crusoe raised $3.9 billion at a $30.9 billion valuation, with roughly $140 billion in contracted order backlog.

Jason's analysis was purely financial: full-stack from power to inference, with the capability to build modular data centers, placing it near the top of the second tier. But at current valuation, the numbers fall just short. He passed.

Rory's framework was broader: data centers are high-leverage assets. They perform well when AI usage is growing fast. If growth slows, that leverage amplifies the downside. He compared these assets to application-layer companies like Factory — more resilient to cycles, not dependent on AI overall growing at hyperspeed, just requiring that software development remain relevant.

The Wall Street Journal provided a fitting footnote: the morning headline read "Data center deals stall on contraction fears." By close, it had changed to "Nasdaq surges as AI capex fears dissipate." The market completed a full fear-to-greed cycle within eight hours. Rory said watching it happen in real time was something.