SpaceX Opens at $2.7T While Anthropic Gets Hit with an AI Export Ban — One Wild Week

Source: 20VC with Harry Stebbings | Published: 2026-06-18T14:00:11Z

SpaceX surged exactly 19% on its first day of trading, commanding a $2.7T valuation on just 4% free float. That same week, Anthropic became the first AI company ever hit with a U.S. government export control over model capabilities.

SpaceX IPO: No Price Discovery, No Roadshow, Perfect Landing

SpaceX's IPO didn't follow the conventional playbook. Elon Musk skipped price discovery entirely—no roadshow, no bookbuilding. He named his price and got it. The stock closed up 19% on day one, right at the upper bound of a "perfect first-day pop." It continued climbing over the following days, pushing the market cap to $2.7 trillion.

Former Benchmark partner Bill Gurley's benchmark: too high a first-day pop means the CEO left money on the table. It touched 30% intraday, which was already pushing it, but closed back at 19%—a number that looks almost engineered. A CEO who did zero price discovery still landed the plane squarely on the runway.

4% Float and Gamma Squeeze: The Stock Price Is Meaningless for Now

SpaceX currently has only about 4% of shares in public float, which makes the stock price extremely fragile. Options trading just opened, and all the conditions for a gamma squeeze are in place—retail investors pile into call options, market makers are forced to buy shares to hedge, rising share prices trigger more call buying, creating a self-reinforcing loop.

With a float this thin, that loop can push the price in any direction. Every insider—VCs, management, employees—is locked up for six months and cannot sell. This is the same dynamic as private-market valuations: the number looks great, but you can't cash out. The price when lockup expires in six months is the real report card.

When asked whether the stock would be higher or lower in six months, the investors on the show leaned toward "lower." Too many structural forces are inflating the price in the short term—passive buying from index inclusion, options-driven gamma squeeze, retail enthusiasm. But once those forces dissipate, there's $1.8 trillion of value between the current $2.6 trillion and the last private round at $400 billion that needs to be digested.

Elon's Core Asset: Long-Dated Call Options That Always Pay Off

One of the guests shorted Tesla in college and lost his entire savings. That painful lesson became the most important cognitive correction of his investing career. Years later, when he led SpaceX's investment at Kleiner Perkins, his pitch to LPs was: the numbers alone get you a decent return, but ignoring what Elon can do is stupid.

Elon's pattern is selling the market "long-dated call options"—massive technological promises that take eight to ten years to verify. In the Tesla era, it was full self-driving. Now it's the Optimus robot. SpaceX started with reusable rockets, then Starship, and now orbital data centers, Mars missions, and lunar launchers. The key: historically, he's delivered on nearly every one.

This gives him a cost-of-capital advantage no one else can touch. Nobody else can walk into a bank and say "I think AI is going to take off, give me $24 billion to build data centers, I don't have contracts yet" and walk out with the money. But Elon can, because every single person who blindly trusted him with capital became extraordinarily wealthy. There's a small-village-sized group of investors whose cumulative portfolios are 95% SpaceX and other Elon companies, all massively in the green.

Anyone who blindly gave Elon money when he asked for it—no matter what he was asking you to invest in—made a fortune. Every single one.

This trust capital creates a snowball effect: low cost of capital lets him make bets nobody else can, winning reinforces trust, which lowers his cost of capital further. Rationalists will argue capitalism shouldn't work this way, but it has for 20 years.

xAI's AI Business: Quietly Surpassing SpaceX's Core Revenue

Before the IPO, Elon systematically solved the "how to monetize Colossus compute" problem. First, the Cursor acquisition—exercise price of $60 billion, when Cursor was running at $4 billion in annualized revenue with a potential year-end run rate of $6 billion. The deal was attractive at signing and even more so after SpaceX listed, since the acquisition price amounted to roughly one-third of a single day's stock price movement.

Then he signed compute contracts with Anthropic at $1.25 billion per month and Google at roughly $920 million per month, totaling about $2 billion monthly. From the moment those contracts kick in this September, xAI's AI business—Google and Anthropic compute revenue plus Cursor—will have annualized revenue exceeding SpaceX's entire Starlink business. While others spent two years debating whether to build data centers, he built two, acquired a company to fill compute capacity when his models weren't good enough, and when that still wasn't enough, signed two massive contracts. Execution speed itself is the moat.

Anthropic: Said the Most Dangerous Thing, Then Got Burned by Its Own Words

Anthropic launched Claude Fable on Monday, a consumer-facing product built on the Mythos model. Mythos had previously been described by Dario Amodei as "extremely dangerous" in the cybersecurity domain, which is why it wasn't made broadly available. Fable was theoretically not supposed to have those cybersecurity capabilities.

Then Amazon discovered a case where Fable could be prompted in a specific way to produce cybersecurity-related outputs. Amazon called the government. The problem is, both sides already distrusted each other—the relationship between Anthropic and the current administration had been tense for a while.

Dario's explanation was technically coherent: Mythos's real threat isn't that it can find any single vulnerability, but that it can discover vulnerabilities at a scale humans can't defend against—spinning up ten thousand instances simultaneously. This particular "jailbreak" didn't constitute that kind of scalable threat. As a logical argument from a top scientist, it holds up perfectly.

But the people on the government side don't need to understand "cybersecurity risk as a function of parallelizable instances." The deputy chief of staff needs to brief the president, and "you said this thing was dangerous, and now it's doing dangerous things" is a yes-or-no decision that can be made in 90 minutes. They invoked export control law—a statute that doesn't require judicial review and falls within executive authority. Anthropic had virtually no legal room to maneuver.

A private company can't go around telling everyone it invented something more dangerous than the atomic bomb while also saying "but we get to decide who uses it." The moment you claim you've built the most dangerous thing in existence, you're part of the political process—and you'd better learn to play politics fast.

This Is AI's "Crossing the Rubicon" Moment

The implications extend far beyond Anthropic. This is the first time in U.S. history that AI has been subjected to export controls based on model capabilities. Previous friction with Anthropic was about contract terms and government access rights. Now the government has explicitly stated: access is restricted for non-U.S. citizens based on capability alone.

The critical test comes in the next three to six months—when OpenAI or Google reaches model capabilities equivalent to Fable, will the government apply the same restrictions? If not, Anthropic may have due process grounds for appeal. If so, it means the U.S. government is genuinely starting to control the distribution rights of intelligence.

Extend that logic: if we actually achieve superintelligence—recursive self-improvement, genius-level cognition running in data centers—who gets to use it? NATO allies? Citizens of specific countries only? Imagine the U.S. and China with superintelligence while Greece or Brazil goes without. The economic implications are deeply unsettling. This transforms "sovereign AI" from a slogan into a serious strategic question.

The Awkward Reality of Open-Source Models and the Sovereign AI Dilemma

A fact being overlooked: with enough inference compute and the right orchestration framework, open-source models can replicate most of the vulnerabilities Fable discovered. This makes the government's regulatory logic inherently unsustainable—if you ban Mythos based on capability, you logically should also restrict open-source models given sufficient compute.

Mistral is raising $3 billion at a $20 billion valuation. But the brutal reality is: outside of China, there are essentially no good open-source models. The U.S. doesn't even have one. Achieving top-tier performance across pretraining, mid-training, and post-training requires extremely scarce talent, capital, and sustained focus. Mistral itself has fallen far behind the frontier on model capabilities, pivoting to an inference platform and fine-tuned models. Most countries' realistic path is probably fine-tuning Chinese open-weight models with post-training—which hardly qualifies as true sovereignty.

Europe's typical dilemma: enough willingness to fund, but nowhere near $100 billion to compete. Five years from now, it'll most likely still be a two-or-three-player oligopoly of American companies, with some European affiliates tagging along.

The Fin Acquisition: The Golden Playbook for Pre-AI SaaS Companies

Salesforce's $3.6 billion acquisition of Fin (formerly Intercom) makes this company the defining case study of a legacy SaaS company successfully pivoting to AI.

Fin's transformation path is worth dissecting. They shifted from per-seat pricing to outcome-based pricing—99 cents per resolved customer issue. This wasn't a pricing tweak; it forced the company to the front line of value creation. To make money, your software has to be in the resolution path, and your answers have to be right. This pushed the entire organization closer to the customer.

The numbers tell the story: a company doing $300 million in revenue growing at 7% was pushed to $400 million in revenue growing at 25% within two to three years. Pushing an already-heavy rock uphill is harder than building a company from scratch.

A Benchmark partner said that for pre-AI SaaS companies, any form of liquidity is a top-10% outcome. Any liquidity at all.

The Intercom team turned a "$300 million revenue, 7% growth, essentially zombie company" asset into $3.6 billion in hard currency. Any pre-AI SaaS company board that doesn't put this case study as the first agenda item at their next meeting is failing in its fiduciary duty.

Wix at 1x Revenue Multiple: Time to Go Private

Wix cut its 2026 guidance, laid off 20% of staff (a thousand people), and trimmed revenue expectations by $25 million. The stock has fallen to a 1x price-to-sales ratio.

They've checked nearly every box on the "bad attributes" list: the product is easily replicated by AI-native tools like Lovable and Replit, the per-seat pricing model is legacy, as market leader they can only lose share, and the product experience feels dated compared to AI-native competitors. The Base44 acquisition was directionally correct, but not enough. The earlier stock buyback now looks like a mistake—capital deployed at the highs, and the stock has since been cut in half.

At a 1x revenue multiple, an interesting option emerges: going private. Michael Dell found Silver Lake, took Dell private when the market hated him, and is now a top-ten richest person in the world. If Wix's management believes they can execute an AI transformation over five years, it might be better to find a PE firm, take the company private, and do the work away from public-market scrutiny rather than getting ground down quarter after quarter.

Adobe at 8x Free Cash Flow: So Cheap Nobody Cares

Adobe beat earnings and raised guidance, but the stock dropped 6%—because the CFO announced he was leaving for Marvell Technology. When the first sentence of a turnaround story is "the CFO left," nobody sticks around for the second sentence.

Adobe now trades at 8x LTM free cash flow. As a value play, that number is compelling. But here's the problem: you can buy Adobe at 8x cash flow with a pile of headwinds, or you can buy Nvidia at 16x earnings with all of AI's tailwinds at your back. For a hedge fund manager benchmarked against an index, this isn't a difficult choice.

Over the past five years, the Bessemer Emerging Cloud Index is down 44% while the semiconductor ETF is up 325%. Short SaaS, long semis, and your returns beat virtually every hedge fund manager on the planet. Trend-following is lesson one for momentum traders, and this trend has been running for five years.

Adobe has checked almost every bad-attribute box: 80% market share means there's only share to lose, the product is increasingly replicable by AI, the business model is the type the market currently hates most, and AI talent is scarce internally. Even more fatal: two years ago their stock price was high enough to use equity to acquire AI companies; now the stock is down 60% while top AI companies' valuations have multiplied several times over—they can't afford them anymore. They missed the acquisition window and now have neither internal talent nor external acquisition currency.

The SaaS Survival Ledger

Public markets are conducting a brutal triage of SaaS companies, and the criteria are getting sharper.

Good attributes: usage-based billing tied to AI token consumption (e.g., Datadog, Snowflake); operating in a sector with clear AI tailwinds (e.g., cybersecurity); being a challenger that can use AI to accelerate share gains.

Bad attributes: per-seat pricing model exposure; product value easily replicated by coding agents; already the market leader with only share to lose; product sitting on the chopping block as enterprise IT budgets get reallocated.

Palo Alto, CrowdStrike, Cloudflare, Datadog, and Palantir have climbed back above 15x forward revenue multiples. The market isn't handing a death sentence to all of SaaS—it's doing precision sorting. The problem is, even if you sort correctly and pick the right SaaS stocks, your returns over the past five years still massively underperform semiconductors. A fund manager's job isn't "being smart within SaaS"—it's making money.

Robotics and the Poly Bag Problem

Standard Bots raised $200 million, betting on a sweet spot between humanoid robots and traditional industrial robot arms. Their thesis: humanoid robots spend enormous sums on legs, and most industrial settings don't need legs. Traditional robot arms can do the work, but their software is terrible. The middle ground—a new generation of arms with integrated hardware and software—is the sweet spot.

There are roughly 3 million robots operating worldwide, while 1 billion people perform manual labor. Thirty years in, less than 1% of human labor has been replaced. When you actually walk a factory floor, human dexterity is staggering—for anything that isn't highly repetitive, humans still crush machines. And warehouse operators calculate ROI down to the penny: if a robot can't come in at less than half the cost of a minimum-wage worker, they won't switch.

The biggest obstacle for a robotics company landing a large order: when a plastic poly bag isn't lying flat, the label scanner can't read the barcode, and the entire production line grinds to a halt. You need a human to smooth out the bag—and if you need a human for that, you don't need a robot. This kind of detail never appears in any investment memo.

But LLMs are changing robot flexibility. Under traditional programming, any deviation in the workflow requires reprogramming. The new generation of LLM-powered robot software lets you say "put this thing into that thing" and it figures out the rest. In one demo, someone moved an object to a different position, and the robot paused, thought for a few seconds like a velociraptor in Jurassic Park, then reached over to the new position and picked it up. That pause—that's the brain working. That's the future.

By 2060, the ratio of digital agents to humans could reach ten thousand to one or higher, while the ratio of physical robots to humans may have just reached one to one. Scaling robot AI is far slower than scaling ChatGPT—you can't deploy a robot through a website or API; you have to buy it, assemble it, and put it to work. But among the sectors where trillion-dollar companies will be born, this remains one of the most likely candidates.