The Era Where You Need Dollars to Buy GPUs, Not Oil — After the Petrodollar, the World's New Economic Infrastructure Could Be 'Compute'

A few years back I tried to slot a GPU server into an existing rack and hit a wall on power capacity. The hardware was already sitting there. The rack just couldn't feed it.

It wasn't a server problem. It was a building problem. The facilities guy running our colo space put it plainly: "You need to free up electricity first."

I'd always thought of compute as a software thing. Turns out it's a physical thing that eats electricity. That's where this post starts.

If you had to name the one resource that drove the 20th century world economy, it was oil.

Cars ran on it. Factories needed it. Logistics ran on it, and so did wars. The bigger industry got, the more energy it burned, and getting energy meant getting oil.

And sitting in the middle of every international oil trade was the dollar. People call it the petrodollar system.

Something strange happens once you move into the AI era.

Crude isn't the only thing companies buy to stay competitive anymore. They buy GPUs. They build data centers. They lock down power. They train models.

All of it chases one thing.

Compute.

We might be at the front edge of a shift from an oil-centered economy to a compute-centered one.

Oil and compute rhyme more than you'd think

Strip post-industrial-revolution growth down to its bones and it looks like this:

Energy → machines → production → goods → money

Countries that could burn a lot of oil and electricity ran more factories and made more stuff.

The AI era adds a new piece of capital equipment to that chain. The GPU.

Feed a GPU electricity and you get computation. Feed that computation to a model and you get tokens, and from tokens you get code, documents, images, video, analysis, designs.

Short version:

Power → GPU → compute → AI → tokens → knowledge and services → money

Old machines amplified human muscle. AI is closer to a machine that amplifies brain work.

Which is why compute is getting hard to file under "IT resource." It keeps drifting toward the means-of-production column.

Hence the idea of a 'Compute Dollar'

The petrodollar story isn't just that oil is priced in dollars.

The point is that the world economy needed oil, the dollar became the settlement and financing layer for that trade, and energy dominance and financial dominance fused into one thing.

So what happens in the AI era?

Building AI takes GPUs. Building GPUs takes leading-edge silicon and HBM. Running data centers takes enormous amounts of power. Using AI services means paying for API calls and tokens.

And right now the major companies and the capital markets behind that ecosystem are wound tightly around the US.

Sketched as a stack, roughly:

Energy → Data Center → GPU / HBM / Semiconductor → Compute → Cloud / Kubernetes / AI Runtime → Foundation Model → AI Agent / API → Token → Economic Value → Dollar

Here's the interesting part of the "Compute Dollar" idea.

In the AI era, dollar demand doesn't only come from buying goods. It comes from the act of securing compute in order to manufacture intelligence.

A million tokens instead of a barrel of oil

The oil economy has an intuitive unit.

$/barrel — what it costs to produce and buy a barrel of crude.

The AI economy is settling on a different number.

$/1M tokens — what it costs to produce a million tokens.

Drop one layer down into infrastructure and more numbers show up.

Tokens per second — how fast you can manufacture intelligence. Tokens per watt — how much output you get from the same electricity. Tokens per dollar — how much AI work you can run for the same money.

The AI infrastructure race isn't a GPU counting contest. It's a fight over how much useful AI output you can extract from the least power and the least money.

Not so different from manufacturing. Factories compete to shave the unit cost of a product; AI data centers compete to shave the cost of a token.

The GPU is the machine tool, the data center is the factory, and the token is the product coming off the line.

Except compute might be scarier than oil

There's one difference.

Oil comes out of the ground in a lot of places. The Middle East, the US, Russia, South America.

The supply chain for frontier AI compute doesn't look like that. It's narrow.

GPU design. Leading-edge foundries. HBM. Semiconductor fab equipment. Data centers. Hyperscale clouds. And the power to run all of it.

Choke any one of those and everything above it shakes.

Anyone who's operated infrastructure recognizes this picture. You can build a beautiful application and it still stops when the database dies. Database is fine but storage dies, it stops. Storage is fine but the network drops, it stops again. Write enough incident reports and you notice the bottleneck was always downstairs.

AI is the same.

AI Model → AI Runtime → GPU → HBM → Semiconductor → Data Center → Power Grid

A bottleneck anywhere in that stack hits every service sitting on top of it.

That's why AI dominance isn't only a contest over who builds the better model. It's a contest over who controls the whole stack.

So is the oil age over?

Worth being cold-eyed about this.

No.

Saying the oil age ended and the compute age began is a much too clean cut.

If anything it's the opposite. The more compute matters, the more power matters.

GPUs eat electricity. Data centers eat electricity. Cooling eats electricity.

Delivering that power reliably requires generation plants and transmission and distribution networks.

Compute doesn't replace energy. It's a new industrial layer stacked on top of it.

Same way adding upper layers to the OSI model never made the physical layer disappear. The bigger the services on top get, the more the infrastructure underneath matters.

Which produces a slightly paradoxical scene in the AI era.

The most advanced industry on earth hangs on power plants, transformers, transmission lines, cooling systems, and data centers — some of the most stubbornly physical industries there are.

No matter how smart AI gets, it does nothing when the power is out. I learned that sentence standing in front of a rack.

The part of the AI race nobody looks at

When people watch the AI race, they mostly watch models.

Which model scores higher on benchmarks. Who supports longer context. Which agent writes better code.

That's the top layer, the one users can see.

The real structure goes much further down.

Power → semiconductors → GPU/HBM → data centers → networks → cloud → containers/Kubernetes → AI runtime → model → agent → service → economic activity

I think the whole stack is a better lens on AI than any single model. There's bias in that, honestly. I've spent long enough in infrastructure that I instinctively suspect the bottom layer first. But I can't remember that habit costing me much.

Real bottlenecks don't form in the flashiest spot.

Models change every few months. You cannot build a power plant or a data center in a few months. You can't scale GPU output overnight either. Grids take even longer.

The faster the upper layers move, the sharper the physical-layer bottleneck underneath gets.

And Korea sits somewhere interesting here

Look at the compute-era supply chain again and something stands out.

HBM matters a lot for AI silicon, and Korea holds serious weight in memory. Power infrastructure, manufacturing, telecom infrastructure — none of it is weak either.

But being good at making components isn't the story.

Having oil didn't automatically make a country an oil power. Being good at semiconductors doesn't hand you compute dominance.

Treat compute as an industry and you need the full chain: semiconductors → data centers → power → cloud → AI platform → model → service.

The question is this.

In the AI era, does Korea become the country that supplies parts, or the country that manufactures and sells compute itself?

The gap between those two is wider than it looks.

In the industrial age, steel output mattered. So did oil consumption, and electricity generation was a number that described a nation's industrial strength. The digital era added data center capacity to that list.

The AI era adds one more.

How many AI operations per second does this country run? How much useful intelligence does it produce per kilowatt-hour?

Those numbers might end up describing national and corporate productivity.

Seen that way, "graphics card" is too small a word for a GPU. "Server room" doesn't hold a data center, and the grid keeps escaping the category of public infrastructure. Wire them together and you get one enormous factory that stamps out intelligence.

So is the next strategic resource the ability to manufacture intelligence?

Whether the compute dollar replaces the petrodollar, I don't know yet.

Compute has no unified trading unit like the barrel. There's no giant international spot and futures market for it the way there is for oil.

Too early to declare that the petrodollar era ended and the compute dollar era arrived.

But the direction is interesting.

Countries and companies used to fight over energy. Going forward there's one more thing on that list: the computational capacity to produce intelligence.

Oil moved machines. Compute moves AI. AI in turn moves the productivity of people and companies.

Maybe the strategic resource of the future isn't GPUs or power or data. What separates the winners is how cheaply and quickly you can weave those together into intelligence.

From that angle the AI race isn't a game for software companies alone. It's a game for semiconductor companies, power utilities, data centers, cloud providers, and national infrastructure.

We might be watching the moment when a world that organized itself around securing oil turns into a world organizing itself around securing compute.

I'm not certain this picture is right. Reading it back in a few years might be embarrassing.

But here's the fun part: at the very bottom of this most advanced AI era, there's still electricity and steel and concrete. Just like the day I couldn't power on a server because of a rack limit.