Two Clocks, One Blow-Up

Porter Stansberry wrote a long post about Leopold Aschenbrenner's July collapse. The consensus already had a simple story: a 25-year-old ran 4-to-1 leverage on concentrated AI bets and got carried out. Porter says that story is incomplete. Leverage explains speed. It does not explain the loss. Leo lost because he was wrong.

I agree with half of that. Then I disagree hard.

What actually broke

Situational Awareness LP was built on two ideas. Long the physical AI build-out: chips, memory, power, rented compute, miners turning substations into neoclouds. Short application software, because models would "sonic boom" the wrappers and make enterprise apps irrelevant.

In July the market did not crash AI. Nvidia barely moved for the month. The S&P stayed near highs. What happened was a rotation out of leveraged second-derivative compute names and into cash-generating software. Longs like Sandisk, Nebius, Bloom, CoreWeave got hit. Shorts like Adobe, Workday, Intuit, Salesforce, Veeva ripped higher. Then the whole public book was sold in a block to Citadel. Month: about minus 67%. A week earlier, LPs had been invited to wire more money after a monster first half.

That is not mysterious portfolio mechanics. That is a thesis meeting a different clock than the one it was priced for.

Porter's real point

Porter's best line is simple: leverage is a magnifying glass. It does not pass judgment.

His deeper claim is capital-structure history. When an economy builds an expensive new network, the people financing the rails often earn a poor return. Competition, overbuild, and obsolescence eat the builders. The money tends to pool in businesses that ride the network with low incremental capital, own the customer relationship, and control the standard. Railroads and express companies. Fiber and Google. Bandwidth glut and the application layer that sat on top of it.

He looks at enterprise software and does not see "code a model can replace." He sees rails: identity, compliance, validated records, switching costs so high that ripping them out is one of the most reliably disastrous projects a large company can attempt. Software priced at a few percent of the employee using it. Incumbents not getting eaten by AI so much as selling AI as an upsell on a subscription customers cannot easily cancel.

That is a serious argument. If you only trade capital cycles, it is often the winning argument for a long time.

Where the argument goes thin

Porter talks as if the capital clock is the only clock that matters.

I live on another one.

I build with AI every day. Not as a spectator of 13Fs. As someone shipping systems, redesigning workflows, watching what actually changes when models stop being search boxes and start becoming workers. From that seat, AI is not a feature cycle. It is a deep change in how work gets done, who holds judgment, and which interfaces still make sense.

So yes: we are seeing consolidation. Existing rails are winning back attention and multiples. Copilot can look brilliant on a quarterly slide. Demand for developers is still high. None of that closes the Kodak moment.

Kodak also made money while the camera was already dead.

Two clocks

Call them the capital clock and the capability clock.

On the capital clock, July made sense. The market rotated toward cash flow, compliance, and lock-in. Leo's book was long expensive second-order infrastructure and short franchises with audit trails and switching costs. Leverage turned a bad month into a near-death event.

On the capability clock, the deeper move continues. Models improve. Agents get longer horizons. Organizations start noticing that their software assumes a human in every loop, a document as the unit of work, a seat as the unit of revenue, and change control designed for quarterly releases. That stack does not merely lag AI. It actively blocks real AI development.

Copilot is AI as a sidebar on the human rail. Useful. Smart distribution. Also temporary in the architectural sense. The real shift is AI as a worker on a new rail: tools, state, permissions, evaluation, memory, continuous execution. Not a suggestion in Office. A different operating assumption about what software is for.

Elon talking about Macrohard is not interesting as brand trolling. It is interesting as an attack on the assumption that software equals seats, documents, and humans clicking through certified workflows.

Google and Microsoft are not the same bet

People flatten "Big Tech wins AI" into one sentence. It is two bets.

Google is closer to winning on substance: models, content, distribution into human attention, silicon, the stack where knowledge and compute meet.

Microsoft is closer to winning on certification: identity, Azure clearances, the enterprise permission structure, the fact that whole industries already run on their rails. That is real power on the capital clock.

It is also a false lead on the capability clock.

Certification is a time purchase. FedRAMP, GxP, Part 11, Active Directory lock-in buy years of cash and distribution while better architectures mature. They do not prove the old product shape is the end state. They prove institutions move slower than models.

Porter treats that delay like destiny. Leo treated it like it did not exist. Both mistakes are expensive in different ways.

Leo's actual error

Not "AI is fake." Not "software has no moat."

His error was portfolio construction against time:

1. Short franchise software before enterprise-grade replacement exists: permissioned, integrated, liable, validated, boring enough for a board.

2. Long the pickaxes as if they were not leverage-sensitive commodities in a capital cycle.

3. Use leverage on a thesis that can be directionally right over years and still slaughter you over weeks.

He had half a point on capability and almost no respect for capital structure. Porter has a strong point on capital structure and almost no respect for the depth of the capability shift.

Old rails cash and obstruct at the same time

This is the sentence I keep coming back to.

Old enterprise rails can cash for years while they obstruct the deeper move. They defend human-default workflows. They trap data in UIs. They price seats instead of outcomes. They force agent systems through change boards built for a slower world. That is why they look like winners in a rotation and still feel like the enemy if you are trying to build what comes next.

More developer demand right now does not refute that. A lot of current engineering is migration labor: adapters between the old world and the agent world. Temporary employment building the bridge is not proof that the old city is the destination.

How I read the board from here

I do not need Leo's trade to believe AI is profound. I do not need Porter's permanence to respect why enterprise software bounced.

If you are allocating capital, do not confuse transformative technology with good returns to the people financing the build-out. History is full of networks that changed civilization and still destroyed investors.

If you are building, do not confuse a quarterly win for certified incumbents with the end of the architectural war. The paper of the present is powerful until the day it is not.

Google-like stacks fight for the content of the future. Microsoft-like stacks collect rent on the paperwork of the present. Builders should assume certification eventually follows workflows that work, not the other way around. And still sell into the world as it is: the paperwork still opens doors.

The harder line

Leo blew up quickly because of leverage. He was vulnerable because he shorted institutional time. Porter is incomplete because he writes as if institutional time is the only time that exists.

The Kodak moment is not canceled by a software rally. It is just slower, meaner, and more expensive than the manifesto version.

Two clocks. One blow-up. Most people are still arguing about only one of them.