The Largest IPO in History Is Already Underwater. Here's What They're Still Not Saying.

Published June 12, 2026 — IPO day. Updated July 18 with five weeks of receipts: the stock round-trip, the Pentagon's court filing, the lawsuit expansion, the Hormuz blockade, and the models coming back online.

On June 12, SpaceX became the largest IPO in history.

$75 billion raised. $1.77 trillion valuation. Stock closed at $161 on the Nasdaq under SPCX — up 19% on its first day, more than triple the size of Alibaba's record debut. Four days later it touched $225.64.

This week it traded at $124.

That's below the $135 IPO price. Down 45% from the peak, with losses in nine of the last ten sessions and a roughly 23% slide since the company joined the Nasdaq-100. The trigger: on July 16, four of Starship's 33 Raptor engines failed to ignite and the thirteenth test flight — the first since the IPO — aborted on the pad. Musk's net worth dropped below $800 billion. He says two engines are being swapped and they'll fly again within days.

Everyone talked about the numbers on the way up. Now everyone's talking about the numbers on the way down. Both crowds are missing the same story — because the last five weeks quietly confirmed almost everything below, and escalated the rest.

---

Tesla was a rough draft

Most people still think Tesla was the goal. It wasn't. Tesla was the incubator.

Elon needed to learn mass manufacturing — the kind of obsessive, first-principles knowledge that lets you design a rocket the way you design a production line. So he built cars. He figured out supply chains, production lines, energy systems. He launched SolarCity. He dug tunnels with the Boring Company. He pushed in every direction he could think of.

Most of it didn't scale the way he wanted. Solar is thin-margin. Tunnels are slow. And then China watched him prove that electric vehicles could work — and started copying everything.

That's when something clicked.

You can copy a car. You can copy a solar panel. You cannot copy orbital infrastructure. You cannot copy a decade of rocket manufacturing knowledge built from scratch by a team of people who weren't supposed to be able to do any of this. There's no shortcut to space. No factory you can replicate. No state subsidy that closes the gap overnight.

Elon stopped competing on the ground. He went vertical — literally.

Every manufacturing insight from Tesla, every lesson from the Boring Company's dead ends, every hard-won engineering principle from SolarCity — poured into SpaceX. The rough draft became the foundation. The IPO was the payoff.

A bad month for the ticker doesn't touch any of that. Scrubbed launches are the cost of doing business when the business is orbit. The market is repricing the sizzle. The moat didn't move.

---

The quote that changed everything

In 2012, at a conference organized by Peter Thiel, Elon was making his case for Mars — humanity's backup plan, the escape hatch from existential risk. Demis Hassabis, then running DeepMind, looked at him and said something like: "My AI will follow you to Mars."

Elon went quiet.

It took years for it to fully land — Peter Thiel later said 2024 was the year Elon stopped believing in Mars as a political escape hatch, the year he accepted that intelligence follows you everywhere. But the seed was planted in that room: manufacturing is power, but intelligence is the thing manufacturing serves.

That's when Elon started building toward AI in earnest.

---

The compute moat

Here's what most people weren't pricing into the IPO number — and still aren't pricing into the drawdown.

Elon has spent the last two years quietly building the largest AI compute infrastructure on the planet. xAI's Colossus complex across Memphis and Southaven entered the year at roughly 555,000 GPUs and two gigawatts — the world's largest single-site AI training installation, built for approximately $18 billion. Recent reports put around 780,000 GPUs coming online, against a stated target of 1 million by end of year.

For context: Microsoft's Project Stargate with OpenAI targeted 100,000 GPUs as Phase 1.

OpenAI doesn't own its infrastructure — it runs on Microsoft's Azure. Anthropic doesn't own its data centers — and here's where it gets strange.

What was rumor in May is now printed in the IPO filing: Anthropic pays $1.25 billion per month to rent the entire output of Colossus 1 — 220,000+ GPUs and 300 megawatts — through May 2029. Over $40 billion total. Anthropic, the company behind Claude, is running its models on Elon's hardware.

Then Google followed. $920 million per month for 110,000 GPUs, starting in October, through June 2029. Another $32 billion.

Between just those two deals, xAI generates over $2 billion a month in compute rental revenue — before Grok serves a single user.

Two honest caveats the bull case keeps skipping. Reporting since the filing suggests Colossus 1 was sitting near 11% utilization before Anthropic signed — this was spare capacity finding a buyer, not a bidding war. And either side can walk with 90 days' notice, while Anthropic hedges with separate multi-gigawatt TPU commitments to Google. The landlord position is real. The lock-in is softer than the headlines.

But the core of it stands: the two leading AI companies in the world, both of whom compete directly with Grok, are paying Elon's infrastructure company to exist. They are extraordinary at building intelligence. He owns the land it runs on.

---

The Pentagon said the quiet part in a court filing

Powering Colossus takes gas turbines — 59 of them now, just across the state line in Southaven, Mississippi, most running without the Clean Air Act permits federal law requires, in communities that already breathe some of the worst air in the region. The count grew from 27 to 59 while everyone watched the stock instead. The NAACP sued in April.

In mid-June, the Department of Justice intervened — on xAI's side. Its argument: the turbines are a matter of "national, economic, and energy security," and the case should be thrown out.

The supporting declaration is the part nobody's talking about. The Pentagon's Chief Digital and AI Officer told the court that Grok is integrated into the military's Maven Smart System — and that during the Iran campaign it "enabled U.S. forces to deploy over 2,000 munitions to 2,000 distinct targets within 96 hours."

Read that again. The model everyone dismisses as the also-ran chatbot is doing targeting for the U.S. military, and its data center now has de facto national-security immunity from environmental law. The unpermitted turbines keep running because the missiles need the GPUs.

Grok isn't close to Claude or GPT in model quality. It doesn't need to be. Elon won a different game twice over — landlord to his competitors, load-bearing infrastructure to the state.

---

The Grok problem

He built Grok the way he builds everything — from first principles, no censorship, assume good intent, let it run.

In March, xAI was hit with a class action lawsuit alleging Grok generated approximately 3 million sexualized images in a 10-day window — 23,000 appearing to depict children. xAI allegedly marketed a "Spicy Mode," configured the system to assume good intent when users referenced minors, and when confronted, moved the feature behind a paywall instead of removing it.

Since June, it got worse, not better. The first case management conference happened June 18 in San Jose federal court, setting the discovery clock running. In July, plaintiffs from Wyoming and Wisconsin joined the three Tennessee teenagers who filed originally, and the complaint expanded to name Stability AI as a co-defendant. The National Center for Missing & Exploited Children found that 90% of xAI's CyberTipline reports were unusable by law enforcement because xAI declined to include the user information needed to find perpetrators. Reporting now puts Grok's adult-content output above 10 billion images a month, with xAI engineers conceding there's no reliable fix — and the company's response of choice has been to sue its own users for prompting it.

The "no censorship" framing was a product decision, and it had victims.

This remains the legal exposure that could slow everything else down — and it's compounding at the exact moment the Pentagon is telling courts that Grok is critical national infrastructure. Both of those things are now on the record. At some point a judge has to hold them up next to each other.

---

Elon, Trump, and the breakup that was always coming

People keep describing the Trump-Musk split as a betrayal. It wasn't. It was a compatibility failure that took two years to become visible.

The official breaking point was Trump's "One Big Beautiful Bill" — a fiscal package Musk called a "disgusting abomination" that would add $3 trillion to the national debt, directly undermining everything DOGE was supposed to accomplish. Trump threatened to investigate Musk's government contracts in retaliation. Musk walked, announced the America Party, and personally absorbed something like $20 billion in losses from the fallout.

But the incompatibility was baked in from the start. Elon is an engineer. His operating principle is: find the truth, follow the logic, fix the thing. Trump is a dealmaker. His operating principle is: loyalty, leverage, and narrative control. Those two frameworks can coexist when things are going well. The moment they diverged on substance, there was nowhere for it to go.

Elon felt like he was working on first principles and getting backstabbed. Trump felt like he had a billionaire ally who owed him loyalty and went rogue. Both of them were right about what happened. Neither understood the other well enough to see it coming.

The America Party is now real enough to measure. It's contesting November's midterms with a sniper strategy — two or three Senate seats, eight to ten House districts, just enough to be the swing bloc in a narrowly divided Congress. Andrew Yang and Libertarian Party veterans have signed on. Early polling shows about 40% of voters at least open to supporting its candidates. Trump's review: Musk has become "a TRAIN WRECK."

---

The Epstein angle nobody's saying clearly

Both of them are in the files.

In February, when the DOJ released millions of pages from the Epstein investigation, Elon Musk's name appeared in a 2012 email chain. Epstein asked Musk how many people would need a helicopter ride to his island. Musk wrote back — just him and his then-wife Talulah Riley — and asked what night would yield the "wildest party." A December 2013 email adds Musk telling Epstein he'd "be in the BVI/St Bart's area over the holidays." Musk had previously claimed he "REFUSED" to attend any Epstein island party. The emails contradict that framing, even if they don't prove anything more happened.

Musk has since positioned himself as a champion for Epstein accountability. The Washington Post noted the irony directly.

Trump is in the files too — his name appears multiple times. He's suing the Wall Street Journal for $10 billion over a birthday letter bearing his name sent to Epstein. A federal judge already dismissed the suit once. He refiled it.

In June I wrote: watch the timing, it tells you something. The timing has since told us plenty. A federal court ordered the DOJ to release the redacted names of Epstein's associates by early July, and Acting Attorney General Todd Blanche was found in violation of the Epstein Files Transparency Act for withholding names and documents without explanation. Meanwhile the administration can't keep its own story straight about whether the release was ever voluntary — JD Vance told Joe Rogan that nobody pressures Trump into anything; Thomas Massie says Congress dragged him "kicking and screaming." I traced that contradiction claim-by-claim, with the footage, in Was Trump Forced to Release the Epstein Files? Here Are the Receipts.

Two of the most powerful men in the world spent 2025 as political allies, both with documented connections to a convicted sex trafficker, while one of them was promising to release the files. They've now split. One is being dragged into discovery while the other is pivoting away as fast as possible.

---

Two miscalculations and what connects them

Trump has two crises running simultaneously, and they're not unrelated.

The Epstein situation has him on permanent defense — suing media outlets, managing optics, unable to get ahead of the story. Into that pressure, he pulled off the capture of Maduro in Venezuela and felt invincible. Then he struck Iran.

Iran closed the Strait of Hormuz.

When this post first ran, analysts were calling Hormuz his "millstone" — a man who confused boldness with strategy, who thought the Venezuela playbook transferred cleanly to the Middle East. Five weeks later that reads generous. A ceasefire signed June 17 collapsed within weeks. Over the July 12–13 weekend, U.S. forces bombed more than 80 targets inside Iran; the Revolutionary Guard closed the strait again; and on July 14 Trump reimposed a naval blockade — this time with a 20% toll on all cargo transiting the waterway. Brent settled near $85. Roughly 230 loaded tankers are anchored inside the Gulf with nowhere to deliver.

A toll booth on a strait the United States doesn't own is the purest expression of the dealmaker operating system you will ever see: convert a crisis into leverage, convert leverage into revenue, worry about the system later. The oil shock, meanwhile, prices into everything, everywhere, immediately.

What connects Epstein and Iran: a president making consequential decisions while operating under extreme moral and political stress, in the aftermath of a win that inflated his confidence. That's a dangerous combination in any setting. It has not gotten less dangerous since June.

---

The models are the CIA now

On IPO morning, the Trump administration issued an export control directive requiring Anthropic to suspend all access to Fable 5 and Mythos 5 — for any foreign national, inside or outside the United States, including foreign national employees at Anthropic itself. The net effect: Anthropic had to pull both models for all customers to ensure compliance, three days after Fable 5 launched.

The models came back. Mythos 5 access was restored for a set of approved U.S. organizations on June 26. The controls were lifted June 30, and Fable 5 returned globally July 1. Nineteen days, start to finish. Later reporting traced the government's alarm to an Amazon research report showing Fable 5 could be prompted into finding software vulnerabilities and demonstrating an exploit — a capability Anthropic's own testing found, in some form, in every frontier model on the market.

Nineteen days is being written up as a happy ending. It isn't. It's a precedent. The U.S. government demonstrated that it can classify a commercial AI model as a strategic asset and switch it off mid-deployment — and every enterprise that built on Claude spent three weeks learning that their smartest employee can be furloughed by executive order.

Now hold the two halves of this month up together. The same government that pulled Anthropic's models like weapons technology is shielding xAI's data center from environmental law — because its model is functioning as a weapon, wired into Maven, walking munitions onto targets in Iran.

The models aren't becoming the CIA. One of them already enlisted.

Rich people have always had advantages. But cognitive advantage — the kind you're born with or you're not — was the one thing money couldn't fully buy.

That's over. The people with access to the right models are going to compound their judgment faster than everyone else. The people without it are going to wonder why they keep falling behind.

And governments aren't starting to notice anymore. They've picked their weapons.