The AI Trade Isn’t About Productivity. It’s About War.

Here’s the number that should worry you more than any earnings call: Nvidia and AMD have shipped roughly 20 gigawatts of AI compute.

The global knowledge workforce runs on about 1 billion brains at 20 watts each. We have already, in raw wattage terms, deployed the compute equivalent of every knowledge worker on Earth.

And productivity hasn’t moved.

The Math Doesn’t Work

Statistical models, which is what “AI” actually is under the hood, improve on a logarithmic curve. Uncertainty falls with the square root of data, not linearly with it. That’s not a temporary limitation; it’s the shape the field has always had. So when people propose 100x-ing compute to close the productivity gap, the math says you don’t get 100x the output. You get roughly a doubling.

Run the numbers on that basis and the economics stop working. A 100x compute build-out lands around $5 trillion a year in depreciated cost, call it $5,000 per knowledge worker, which sounds cheap against an $50-80k salary. But a logarithmic return on that spend gets you maybe a 2% productivity bump, which reprices the real cost at roughly $250,000 per worker. That’s not a rounding error. That’s the whole thesis breaking.

The Business Model Is Already Underwater

It gets worse on the supply side. The going market rate for AI compute is around $1.25 per kilowatt-hour. The real, fully loaded cost — chips, power, data center infrastructure, operations — runs closer to $3-5. Every unit of compute sold is being sold at a loss. That’s not a temporary land-grab discount; it’s been flat since inception, and it hasn’t closed.

There are only two explanations for a gap like that persisting: there isn’t enough real demand to support cost-covering prices, or the sellers are locked in an arms race for market position that matters more to them than near-term margin. Both are true here at once.

What This Actually Is

Which raises the real question: if the technical case for AI productivity doesn’t work, and the current pricing doesn’t work either, why is hundreds of billions of dollars still pouring in?

Because this was never really a bet on AI making knowledge workers more productive. It’s a war for the ad-supported search monopoly Google has held for two decades — and Meta is in it for the same reason, defending its own ad moat. Every “frontier benchmark” is a recruiting poster in that war, built to pull a user’s next question away from Google and onto someone else’s platform. Microsoft, Amazon, and a fleet of startups are the other side of that same fight, and everyone is spending like it’s existential — because for the loser, it will be.

That reframes the capex entirely. It isn’t investment in a durable new market. It’s munitions. Nobody in a war asks whether the ammunition will pay for itself; they ask whether they’ll still be standing when it’s over. Long-term ROI is not the operative question here. Survival is.

Where This Leaves Us

Strip away the framing and here’s the base case: AI, as a standalone productivity story, is close to a wash. Most of today’s players don’t survive the war intact and will eventually be written down to near zero. The capital being spent isn’t building a durable asset base, it’s burning as the cost of staying in the fight. And even the side left standing when the dust settles will have taken enough damage along the way to warrant a serious repricing.

This isn’t a story about whether AI works. It’s a story about who owns the ad business formerly known as search, and what it costs to find out.