top of page
Search

Leopold Was Right. The Terrain Was Bigger Than the Map.



Audio cover
Leopold Was Right.

Two years ago, Leopold Aschenbrenner published Situational Awareness and made a remarkably simple argument about an extraordinarily complicated future: Follow the compute.


If AI capability continues scaling, then the infrastructure required to produce that capability becomes predictable enough to reason backward from. More intelligence requires more compute. More compute requires more chips, memory, networking, data centers and electricity.

That insight made him famous.


Then he put money behind it. And for a while, it worked spectacularly.


Aschenbrenner's fund, appropriately called Situational Awareness, built concentrated positions around the physical infrastructure of AI. By the end of June 2026, Reuters reported that the fund had returned 439% for the year. Then the AI trade turned violently against him. Leverage amplified the losses, margin pressure followed, and Situational Awareness was ultimately forced to sell most of its public-equities portfolio to Ken Griffin's Citadel.

There is an easy version of this story.


Young genius gets overconfident. Uses too much leverage. Gets humbled by Wall Street.

That's probably going to be the version most people remember.


I think they're going to miss the more important one.


Leopold may have been right about AI. What nearly killed him was everything surrounding it.


I Have Been Watching This Stack for Two Years

When Situational Awareness appeared in 2024, I wrote about it because I thought Aschenbrenner was identifying something much larger than another model cycle.

The implication wasn't simply that GPT-5 would be better than GPT-4.

It was that intelligence was becoming an industrial product.

And industrial products require factories.

I kept following that thought.

Compute meant semiconductors.

Semiconductors meant fabrication capacity, lithography, packaging and supply chains.

Compute at scale meant data centers.

Data centers meant transformers, cooling systems, substations, transmission capacity and enormous amounts of electricity.

And all of it meant capital.

That became a recurring argument in my writing because I increasingly believed that we were making a category mistake by treating AI primarily as software.

It isn't.


Software is what we see on the surface.


Underneath it is a physical machine.


Earlier this year, I described that machine another way:

We have turned sand into thought.

Silicon begins with one of the most ordinary materials on Earth. We refine it with some of the most sophisticated manufacturing technology humanity has ever built, etch billions of microscopic structures onto it, wire those chips together, surround them with memory and networking, feed them extraordinary amounts of electricity, load models onto them...

and they produce cognition.

That isn't metaphorical.

It is literally what we're doing.


But the alchemy requires a furnace.


And the furnace is expensive.



The Dependency Chain

Once you look at AI this way, the system becomes easier to see.

Sand becomes silicon.


Silicon becomes chips.


Chips become compute.


Compute consumes power.


Compute plus models produces cognition.


Cognition produces economic output.


Economic output attracts capital.


Capital builds more infrastructure.


And the cycle begins again.


That is the machine.


What's fascinating about Aschenbrenner is that he essentially attacked this machine from the opposite direction.

He started with the anticipated intelligence.

Then he reasoned backward.

If capabilities continue climbing, what infrastructure becomes unavoidable?

Who supplies it?

Who owns it?

Who benefits?

That is an exceptionally compelling framework.

And the fact that his fund subsequently got hammered does not automatically invalidate it.

In fact, one of the strangest details of the whole episode is that investors copying versions of his public positions without his leverage reportedly remained substantially positive even after his fund's collapse. Business Insider reported that the copy portfolio was still up more than 50% from its March launch.

Think about what that suggests.


The map may have been pretty good.


The vehicle wasn't built to survive the terrain.



The Missing Layer Was Capital

This is where the story gets much more interesting to me.

We talk constantly about AI constraints.


  • GPU shortages.

  • Memory.

  • Networking.

  • Transformers.

  • Power generation.

  • Grid interconnections.

  • Cooling.

  • Land.

  • Water.

  • Fabrication.


But there is another constraint underneath every one of them:

the price and availability of money.

You cannot build a gigawatt-scale AI campus with enthusiasm.


  • Someone has to finance it.

  • Someone has to tolerate the construction cycle.

  • Someone has to survive interest-rate changes.

  • Someone has to carry risk while power infrastructure catches up.

  • Someone has to finance fabs years before the chips coming out of them generate revenue.

  • Someone has to hold the assets when markets decide that yesterday's most valuable infrastructure suddenly isn't fashionable this afternoon.


Capital is not adjacent to the AI stack. Capital is part of the AI stack.


And leverage introduces something particularly dangerous into a long-duration technological thesis: a short-duration survival requirement.

You can be completely correct about what happens over the next ten years and still be forced out of the position next Thursday.

That is apparently what happened here.


Situational Awareness's public book was under enough pressure that it faced the choice of raising more capital or unloading positions. Most of those public holdings eventually went to Citadel.


The future didn't have to change. The clock did.


That's the lesson.



Apple Is Playing a Completely Different Game

There is another reason this story caught my attention.


Apple offers almost the inverse strategy. Apple doesn't need to know which frontier model ultimately wins. It doesn't even necessarily need to build that model itself.

Apple has spent years building control over the hardware beneath the experience.

Its M5 generation is explicitly designed around substantially greater AI performance, including Neural Accelerators integrated throughout the GPU architecture. And Apple's next CEO is John Ternus, the longtime hardware-engineering leader who is scheduled to take over from Tim Cook on September 1.


That's interesting.


Because while much of Silicon Valley has been fighting over models, Apple has spent more than a decade mastering the substrate.


  • Apple Silicon.

  • Unified memory.

  • Power efficiency.

  • Local inference.

  • Vertical integration.

  • Device distribution.


Apple doesn't necessarily need to correctly predict whether OpenAI, Anthropic, Google, Meta or somebody we haven't met yet owns the best intelligence five years from now.

It can buy or partner for intelligence.

What is much harder to buy overnight is twenty years of hardware architecture, silicon expertise, operating-system integration and hundreds of millions of deployed devices.

That's optionality.


And optionality matters enormously when the top layer of the stack is changing every six months.


Intelligence Is Becoming Infrastructure

This is where several things I've been writing about for the past two years start collapsing into the same picture.

In 2024, the question was whether AI capability would continue scaling rapidly enough for compute to become strategically important.

That debate is basically over.


Then the question became whether we could build enough chips. Then whether we could power enough data centers. Then whether agents could turn that intelligence into useful autonomous work.


Now we're beginning to see the entire system operating at once.


This is why I recently wrote that AI isn't a tool anymore. It's terrain. Terrain is larger than the model...Terrain includes the chips...

  • The memory.

  • The data center.

  • The electrical grid.

  • The capital markets.

  • The interest rate.

  • The supply chain.

  • The labor force.

  • The regulation.

  • The geopolitical relationships surrounding semiconductor manufacturing.

  • The agent operating on top of all of it.

  • And eventually, the economic output created by that agent.


No single layer tells you the story anymore.



Sand, Thought, Money

The more I look at this, the more I think the AI economy is best understood as a conversion engine.


We take physical resources and turn them into cognition. Then we turn cognition into economic output. Then we turn economic output back into capital. Then capital buys more physical resources. Matter becomes intelligence. Intelligence becomes money.

Money builds more machines that produce intelligence.


That's the loop. And if that loop continues accelerating, it may become one of the most powerful capital-formation engines ever created. But every dependency still matters.


  • Run out of chips and it slows.

  • Run out of electricity and it stops.

  • Run out of capital and it never gets built.

  • Misprice the risk and somebody else winds up owning your assets.


Leopold Aschenbrenner understood something important very early: if you can see where intelligence is going, you can reason backward toward the infrastructure the future will require.


I still think that's right.


What the past few weeks demonstrated is that the reasoning has to continue one layer further. Intelligence depends on compute. Compute depends on silicon. Silicon depends on infrastructure and power. Infrastructure depends on capital. And capital has its own physics.


Leverage.

Liquidity.

Duration.

Volatility.

Cost.

Time.


Ignore any one of those layers and the system eventually reminds you it exists.

That doesn't make the thesis wrong.


It makes the terrain bigger than the map.


And I suspect that the people who truly understand the next decade of AI won't be the ones staring exclusively at the models.

They'll be the ones who can see the entire machine.

From sand...

to silicon...

to power...

to thought...

to capital...

and back again.


Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital.



Source Links

Primary / Current Reporting


Leopold Aschenbrenner


Apple / Hardware


Related Writing

Comments

Animated coffee.gif
cup2 trans.fw.png

© 2018 Rich Washburn

bottom of page