In 1865, a 29-year-old English economist, William Stanley Jevons, spent the better part of the whole year mulling over a problem he couldn’t quite figure out.

New steam engines were efficient, they burned far less coal for the same work, so Britain should have needed less coal.

Instead, the opposite happened. It burned more, because cheap power got used for all sorts of jobs nobody had bothered with or thought about before.

Jevons wrote it up in a book called The Coal Question. His basic premise was that if you make something more efficient, then people use so much more of it, and total demand goes up.

Make it cheap, demand skyrockets, you end up using more.

It’s called Jevons paradox, and while it might now be a 161-year-old concept, it’s arguably as applicable today as it was in 1865.

And when you understand how it applies today, you can make a killing in the stock market thanks to recent AI fear.

On Tuesday, the Chinese lab gave away DeepInfer, free for anyone to use.

In the notes, you can read a Q&A about the Flash Inference Deployment. And there’s a section there that proves how important Jevons paradox is, and why the bears are utterly wrong.

“A single cluster currently serves 155 trillion tokens per day, with overall revenue of US$38 per GPU-hour (at valley pricing).”

Valley pricing, by the way, is the off-peak rate. Peak rate, higher.

So, every chip in that cluster takes in at least $38 an hour.

Ok, that’s handy because we know what the cost to rent a chip is.

Nvidia’s top chip, the B300, rents for around $7 to $8 an hour right now. At $7, DeepSeek keeps 82 cents of every dollar before other costs. At $9.50, it keeps 75 cents.

Imagine that you run an open source AI model, it’s so efficient that you’re serving 155 trillion tokens per cluster per day, and you can make $38 per GPU hour at around an 81% margin!

I’m sorry but the figures are insane. And anyone that’s bearish on compute, considering the cash cow that compute providers are turning into, has rocks in their heads.

Bears will say that’s just the initial take, it doesn’t factor in all the other running costs, and when you look at Anthropic burning $80 billion, then it’s clear there’s no payoff.

Well, no.

That’s Anthropic. And they might not exactly be the best model to follow. Nvidia has gone heavy into open-sourcing models, their Hugging Face acquisition is a good indicator of this.

And the idea of a handful of AI companies controlling and ruling the market doesn’t really sit well with anyone, except maybe OpenAI, Anthropic, and Meta.

But when you have these releases like DeepSeek just gave us, it proves that you can drive down the cost of tokens, increase the total usage of tokens, and still make an absolute fortune along the way.

Add to this, there are an increasing number of forecasts that suggest demand will dominate supply for at least another five years.

A recent forecast from JP Morgan has demand reaching near 550GW of compute by 2031, and supply only being able to get to around 250GW. So, supply not even half of the demand for compute.

And people are bearish on the neoclouds and compute providers. Ha.

The idea that no one is using this compute is wrong too.

On Vercel’s gateway in July, the tokens passing through grew 59%, the price paid per token fell 13.6%, and yet total spend still grew 37%.

That is very literally Jevons paradox playing out in real time.

Look, this is a volatile market, it feels like a real tug-of-war between the bears and the bulls.

And frankly, there’s enough fear to warrant some investors to sit it out. I get that.

But if you have an appetite for the risk, and you hate the idea of sitting on your hands when generational wealth is on the table, I do think the neocloud and compute market is now grossly undervalued.

And as the market proves that Jevons paradox is a real thing, because we’re seeing it live, it only solidifies my view that the most valuable resource in the world is compute.

More soon,

Lewis Maguire