Jonathan Shiller

Long horizons. First principles. Structural change.

Following the argument wherever it leads.

[Jonathan Shiller @JonathanShiller]

The optimists haven't answered the hardest question: what happens when the displacement technology is also the thing that used to absorb the displaced? https://t.co/KLQLYzaHnS

[Jonathan Shiller @JonathanShiller]

https://t.co/afkue2d8QV

[Jonathan Shiller @JonathanShiller]

Everyone assumes the AI trade is simple:
More AI demand → more data centers → more chips → more infrastructure → endless growth.
But what if the market is confusing technological inevitability with economic inevitability?
This week, reports emerged that Blackstone Inc. unexpectedly pulled back from financing one of the world’s largest AI data center campuses.
That matters.
Because AI infrastructure spending is happening at historic scale.
Hyperscalers are on pace to spend roughly $600-700 billion in capex in 2026, with the majority tied directly to AI buildout.
At the same time:
• US grid interconnection queues have exploded from ~300 GW in 2021 to over 2,600 GW today
• Transformer lead times have stretched from weeks to 18-24 months
• Specialized MLCC capacitor prices are reportedly up 20x
• Frontier model inference costs have fallen 280x in two years
And that last number is the contradiction.
The entire AI trade is funding massive physical infrastructure on the assumption that compute scarcity persists long enough to generate durable returns.
But the technology itself is rapidly destroying scarcity.
Open-weight models are compressing pricing power.
Token prices are collapsing.
Enterprise monetization remains unclear.
Even Alex Karp just warned that AI is being “completely irresponsibly oversold.”
History is full of this pattern.
Railroads were transformative. Most railroad investors got wiped out.
The internet changed civilization. It also produced a fiber glut and trillions in destroyed equity before the winners emerged.
AI may be economically transformative.
That does not mean everyone building the infrastructure survives long enough to profit from it.
The market is behaving as if technological inevitability guarantees economic inevitability.
History says otherwise.

[Jonathan Shiller @JonathanShiller]

I think the market is fundamentally misunderstanding the AI boom. Everyone still treats this as a software story, but the economics are already shifting faster than most people realize. Stanford’s 2025 AI Index found inference costs for GPT-class models have collapsed roughly 280x in two years, while open-source models continue rapidly closing the gap with frontier labs. In plain English, intelligence itself is becoming cheap.

But that doesn’t kill the AI trade — it simply shifts scarcity somewhere else. McKinsey estimates roughly $6.7 trillion in AI data center investment will be needed by 2030, while the IEA projects data center electricity demand approaching 945 TWh annually, roughly Japan’s entire power consumption.

We assumed AI would dematerialize the economy. Instead, it may force the largest infrastructure buildout in decades. The biggest winners may not be the companies creating intelligence, but the ones controlling the physical systems required to deploy it. The great irony of AI may be this: the cheaper intelligence becomes, the more valuable the physical world gets.

[Jonathan Shiller @JonathanShiller]

Everyone is debating whether AI is a bubble.

Wrong question.

The real story isn't in the models. It's in what now has to be built to power them — and who is building it.

Hyperscalers are expected to spend nearly $600 billion in capex this year alone. Global AI data center demand is projected to reach 347 gigawatts by 2030. The United States added just 15 gigawatts of new generating capacity in the first five months of this year.

The math doesn't close. Power, not chips, is becoming the binding constraint on intelligence itself.

That's what makes SpaceX's January FCC filing worth paying attention to. Not Starlink. Not consumer internet. A proposal for up to a million satellites, built for one purpose: 100 gigawatts of AI compute in orbit — roughly a fifth of total U.S. electricity consumption, none of it drawn from the terrestrial grid.

Space changes the equation. Solar arrays that never see night. Radiative cooling into vacuum instead of water-hungry chillers. No permitting queue, no transmission fight, no local hearing to lose.

But the bigger story is structural.

States used to control strategic infrastructure by default: the grid, the spectrum, the launch pad. SpaceX already controls roughly two-thirds of all active spacecraft in orbit. It's preparing an IPO that could value it north of $2 trillion. It's reportedly in merger talks with xAI. If that combination lands, one company owns the rocket, the satellite mesh, the orbital power plant, and the model — the entire stack, with no sovereign regulator positioned to say no.

We were told the digital economy would become increasingly weightless.

Instead, intelligence is becoming the most infrastructure-intensive industry humanity has ever built.

And increasingly, that infrastructure isn't being built by states.

It's being built by whoever can move fastest, raise the most capital, and escape gravity first.

[Jonathan Shiller @JonathanShiller]

AI is not replacing software engineers.

It is replacing the process that creates software engineers.
Junior developers once learned by building. Now increasingly they supervise systems they do not fully understand.

The first casualty of automation is rarely employment.
It is apprenticeship.

[Jonathan Shiller @JonathanShiller]

@alex_verem @Noahpinion The scary thing about AI isn’t that it replaces workers.

It replaces expertise.

This drone isn’t just planting trees.
It’s doing work that once required years of human knowledge.

[Jonathan Shiller @JonathanShiller]

Two years ago everyone assumed the AI race would be won by whoever built the best model.

Today even Google can’t build compute fast enough.

Turns out the real AI race may be won by whoever controls electricity, silicon, and physical infrastructure.

[Jonathan Shiller @JonathanShiller]

@FirstSquawk @azeem We spent twenty years believing software had transcended scarcity.

AI is quietly proving the opposite.

Every prompt now consumes electricity, cooling, silicon, and physical infrastructure.

The cloud was never weightless. AI simply exposed the factory behind it.

[Jonathan Shiller @JonathanShiller]

@BharatKChandar @Noahpinion Labor markets may equilibrate in aggregate while the developmental pathways that once converted ordinary effort into competence quietly disappear. The employment question and the legitimacy question aren't the same question.

[Jonathan Shiller @JonathanShiller]

Civilizations survive only so long as ordinary people believe necessity can be earned.

[Jonathan Shiller @JonathanShiller]

The immigration debate and the AI debate are secretly the same debate.
In both cases society is asking:
Should institutions protect existing people… or replace them with higher-performing alternatives?

For a long time institutions existed to serve human beings.
Increasingly human beings exist to optimize institutions.

[Jonathan Shiller @JonathanShiller]

VALUATION VELOCITY https://t.co/0kRFvxyGDL

[Jonathan Shiller @JonathanShiller]

The market keeps treating AI as inherently deflationary. But the sequence may be backwards. Before AI lowers labor costs, it is driving demand for electricity, chips, cooling systems, data centers, transmission infrastructure, and specialized construction. We keep pricing the productivity dividend. The inflation arrives first.

[Jonathan Shiller @JonathanShiller]

@rohanpaul_ai The moat isn't disappearing.

It's routing around itself.

[Jonathan Shiller @JonathanShiller]

@azeem $175 billion in revenue.

$650 billion in infrastructure that needs to earn a return on it.

The growth is real.

The gap is realer.

[Jonathan Shiller @JonathanShiller]

@sama Terra launches at half the price of GPT-5.5.

The reflexivity loop doesn't wait for permission.

(Neither, apparently, does the government.)

[Jonathan Shiller @JonathanShiller]

@michaelxpettis This is the part nobody wants to say out loud:

AI isn't a layoff event. It's a hiring freeze that compounds for a decade.

The people already inside are fine. Possibly great.

The people trying to get in are learning skills no one will pay them to develop.

[Jonathan Shiller @JonathanShiller]

The AI race was never between companies.

It's between the speed of capability deployment and the speed at which everything else — markets, law, competition, governance — can adapt.

Startups aren't being outcompeted. They're being outpaced. Those are different problems with different solutions.

That gap is the story. It keeps widening.

[Jonathan Shiller @JonathanShiller]

The optimists haven't answered the hardest question: what happens when the displacement technology is also the thing that used to absorb the displaced? https://t.co/KLQLYzaHnS

[Jonathan Shiller @JonathanShiller]

Google Maps didn't stop people from understanding cities. It stopped them from having to. The question isn't whether AI writing is useful. It's what you lose when you stop having to find the words yourself.

[Joe Weisenthal @TheStalwart]

Unfortunately, I think that in the near future, not using LLMs to write for you will be like someone refusing to use Google Maps for directions in a new city. A bizarre idiosyncratic choice that's just completely incomprehensible to the vast majority of people.