The Option Musk Never Had to Ask For

These pieces were drafted before June 12. The IPO happened. Cursor closed at $60 billion. The analysis below did not need updating. That's the problem.

Previous in this series: Cursor Has No Moat. SpaceX Has No Cash. Perfect Match. / The Landlord, The Exit, and The Ghost of Houston / SpaceX's S1 Is Out / SpaceX IPO Valuation: Am I Missing Something? / SpaceX IPO (SPCX): You're Already In

Disclaimer: All analysis here is speculation and speculation only. Written for fun and intellectual curiosity. Not financial advice. Not an accusation of anyone, any organization, or anything. Just someone connecting dots and asking questions when there is nothing better to do.

Now, let's get started.

Don't get me wrong. I genuinely adore Dario Amodei. He is the kind of person I would happily catch up with over a coffee break. His thinking on humanity, on the long arc of what AI means for society, even his instincts on UBI, I have written about why the math doesn't add up, but I have never doubted the sincerity behind the question. Dario asks the right things. That matters.

This piece is not about Dario the person. It is about Dario the CEO, a person wearing more hats than one human should, moving faster than any organization probably should, in an industry that does not slow down to let you think. Sam Altman wears even more hats. Michael Truell is luckier in the hat-selling business. He is also unlucky to be in AI, where one second is other industries' one day.

Something got through without the scrutiny it deserved. Let's talk about what that was.

Everyone called it a partnership

When Anthropic signed its compute deal with SpaceX, $1.25 billion a month, through 2029, with a 90-day termination window on both sides, the narrative wrote itself. Two frontier AI organizations finding common ground. Infrastructure meets intelligence. xAI generates meaningful revenue. A signal that the AI economy is real, that money is flowing, that the sector deserves the valuations being assigned to it.

Institutional investors noticed. The tide lifted. If you were wondering whether SpaceX's IPO had an AI credibility problem, this helped answer it. The AI economy is monetizable. Here is the invoice to prove it.

What nobody said loudly enough: Anthropic may have just helped validate the IPO it is racing against for the same investor dollars. Same window. Same capital. Same institutional appetite being divided between two companies with overlapping pitches.

Set that aside for a moment. The deeper issue is what the contract actually is, structurally, when you read it correctly.

It is not a partnership. It is an option. And Musk holds it.

Google understood the assignment

Let's look at how Google approached its relationship with SpaceX, because the contrast is instructive.

Google holds a meaningful equity stake in SpaceX. That is not an accident and it is not sentiment. It is a calculated position built on three things that all had to be true simultaneously.

First, Google is fine if SpaceX succeeds. Their equity participates in the upside. A SpaceX IPO is not a threat to Google. It is a return.

Second, SpaceX cannot hurt Google deeply. If the relationship sours, Google has the cash, the infrastructure alternatives, and the institutional memory to absorb it. It stings. It does not bleed.

Third, the precedent was already made. xAI got its revenue story. Google got what it needed. The geometry worked.

All three were true. Google's math was clean. Then Google signed its own compute deal, $920 million a month, June 2026, same 90-day clause. Google did the math and decided it was still fine. The three conditions still held.

Anthropic's math was not

Run the same three questions for Anthropic.

Is Anthropic fine if SpaceX succeeds? Complicated. A successful SpaceX IPO means Musk has the capital and the narrative. xAI is no longer a story about potential. Anthropic's differentiation gets harder, not easier.

Can SpaceX hurt Anthropic deeply? Yes. Materially. Anthropic is burning capital at a rate that makes the eyes water. The IPO is not optional. It is existential. The S-1 will be read line by line by institutional investors who are already doing the math on burn rate, on dependency, on single points of failure in the cost structure. A compute contract held by a counterparty who has publicly said Anthropic is "genuinely trying to make AI that kills everyone" is not a footnote. It is a risk factor.

Was the precedent already made before Anthropic signed? Yes, but not in Anthropic's favor. Musk already had his revenue story from the Google relationship. Anthropic gave him a second one, with more leverage, against a weaker counterparty.

None of the three conditions that made the Google deal sensible applied to Anthropic. Not one.

The option has a clause nobody put in the press release

There is one more thing about this contract that did not make most headlines.

In a post on X, Musk wrote that SpaceX reserves the right to reclaim the compute if Anthropic's AI engages in actions that harm humanity. The clause was not in the formal press release. Whether it is enforceable is a separate question.

What it does, regardless of enforceability, is reframe the relationship. A landlord sets the rent and collects the check. A landlord who also decides whether the tenant's work is morally acceptable is something else. Anthropic, one of the three leading AI labs in the world, is now partly dependent on infrastructure controlled by a direct competitor, with a moral termination clause written by that competitor, announced on social media.

Who decides what harms humanity is not a contract question. It is a power question.

I know how to train. I don't know what comes after.

Here is the quieter problem underneath the contract.

The leading AI labs, all of them, Anthropic and OpenAI included, were built by people who are extraordinarily good at one thing: making models smarter. The research is real. The capability gains are real. What comes after the training run, how you sell it, to whom, against whom, and with what strategic position, that is a different set of skills. Largely absent from the founding DNA of these organizations.

So what fills the gap? Motion. Let's sell. Let's partner. Let's sign. The enterprise sales hires go in, the deals get done, the revenue numbers start climbing. None of that is wrong. The problem is when the motion happens before the direction is set. When you are selling before you know what you are selling, or to whom, or why a customer should choose you over the model that runs inside the tool they already use every day.

The Cursor deal is the same pattern from a different angle. Two parties, both under pressure, both without a clear picture of what their maximum benefit actually looks like, finding each other at exactly the right moment of mutual need. The MIT kids are extremely smart. Smart people under capital pressure make the deal that is available, not necessarily the deal that is optimal. Cursor took SpaceX's option. Anthropic took SpaceX's compute. OpenAI is running the same playbook with different counterparties.

The real Silicon Valley move would have been different. Test everyone publicly. Use Nvidia's hardware, ask Oracle to run it, release the results right before the IPO. Watch Grok speak for itself. $0. Instead, he wrote a check for $1.25 billion a month.

Everyone in the Valley with scars on their back knows this playbook. Dario, apparently, did not.

Anthropic's specific cost is that its entire public identity is built on the claim that it has thought further ahead than anyone else. The gap between that claim and the reality of this contract is not a scandal. It is just expensive.

The floor beneath the floor

Now add the part that nobody is pricing in.

The 220,000 GPUs Anthropic is renting sit in Memphis. The power that runs them comes partly from a fleet of gas turbines in Southaven, Mississippi. 57 of them as of last count, up from 27 in April, operating without permits, in violation of the Clean Air Act, next to homes and schools in a community the American Lung Association has already rated F for ozone pollution.

The NAACP sued in April. Earthjustice filed for an injunction. A court order to halt the turbines would be a direct power supply event for the facility Anthropic's models run on.

On June 16, the Justice Department intervened. Not to enforce the law, but to dismiss the lawsuit. The argument: Grok's operation is "critical to the economy" and national security. The DOJ, xAI, and the state of Mississippi are now jointly asking the court to make the lawsuit go away.

A judge has not yet ruled.

Anthropic's compute dependency now runs through a chain: a 90-day commercial contract, sitting on top of an unpermitted power plant, currently protected by a Justice Department intervention that exists because the current administration finds it useful. Remove any one of those three layers and the lights get complicated.

On June 12, the same day SpaceX opened on Nasdaq, the Commerce Department used export controls to order Anthropic to shut off foreign access to its two most capable models. Fable 5 and Mythos 5 went dark for all users, not just foreign nationals, because that was the only way Anthropic could comply quickly. European politicians called it a kill switch. Yann LeCun had a different read. "Dario Amodei's ridiculous fear mongering about Mythos/Fable finally pays off," he wrote on LinkedIn. OpenAI's GPT-5.5 wasn't touched. Google's Gemini wasn't touched. The directive landed on the one company that had most publicly stressed the danger of its own models.

Three months earlier, the Defense Department had designated Anthropic a national security threat, a designation never previously applied to an American company, after Anthropic declined to remove safety guardrails from a military contract. The same government that called Anthropic a threat in March called Anthropic's compute infrastructure a national security necessity in June.

"National security" is not a legal standard in this context. It is a weather vane.

The 90-day termination clause in the SpaceX contract reads like the risk. It is not the risk. It is the ceiling of the visible risk. Below it is an unpermitted power plant. Below that is a DOJ intervention. Below that is a Commerce Department with export control authority. Below that is an administration that has already demonstrated it will use all of these tools, in any direction, whenever it finds it convenient.

Anthropic signed a compute contract. It did not sign up for all of this. But here it is.

A note on what comes next

I have written before about Google's moat drying up. About SpaceX's valuation math. About what the S-1 actually says when you read the footnotes. About the retail investors holding the door open while everyone else leaves.

The thread connecting all of it is simpler than it looks. In the current AI landscape, the most important question is not who has the best model. It is who controls the dependencies.

Compute is a dependency. Regulatory relationships are a dependency. Narrative is a dependency, and Musk understands this better than almost anyone alive. The DOJ intervention on behalf of an unpermitted power plant, on the same day as the IPO, is not a coincidence. It is the dependency made visible.

Anthropic handed him one of each.

The coffee will still be good whenever Dario and I finally find a moment. I just hope by then the conversation has moved on to something less expensive.

All information in this piece is based on publicly available sources. Where I speculate, I say so. This is speculation informed by observation, not insider knowledge.