Intelligence per Watt
July 2026
Watch what the frontier labs brag about now.
Two years ago, every model launch led with a benchmark table. MMLU, GPQA, SWE-bench — bar charts with your model in a slightly darker color, edging out the competition by two points. The launches of 2026 lead with something else. Price per million tokens. Latency. Efficiency curves. And since inference cost is mostly electricity plus depreciation, tokens per dollar is really tokens per joule.
Nobody announced the change. But the scoreboard is different. The race is no longer how smart. It’s how smart per watt.
That shift is worth staring at, because industries only change their headline metric when they’ve changed what they are.
Chips did this. The 1990s were the megahertz wars — raw clock speed printed on the box, bigger number wins. Then mobile arrived, batteries became the constraint, and the metric that mattered became performance per watt. Nobody puts clock speed on a billboard anymore.
Electricity did it first. The 1880s were pure spectacle — arc lamps, world’s fairs, Edison electrocuting an elephant to frighten people about alternating current. Then the spectacle wore off and electricity became cents per kilowatt-hour.
Here is the rule hiding in both stories: when a technology’s defining metric becomes output per unit of energy, the technology is confessing what it has become. An input. A commodity. A utility. Not a marvel anymore — an ingredient, priced like one.
Intelligence just made that confession.
Why now? Because the labs did this to themselves, and they had no choice.
Capability leads that used to last a year now last a quarter. Every trick discovered at the frontier leaks downmarket within months — distilled, open-weighted, replicated. The price of a fixed unit of intelligence has been collapsing roughly ten-fold a year, not because anyone wanted it to, but because five labs are competing like hell and none of them can afford to stop. When your product is interchangeable with your rival’s, and your rival’s marginal cost is the same electrons flowing through the same silicon, price doesn’t drift toward value. It falls toward cost. That’s not a strategic failure. That’s just what competition does to things that can be measured.
And intelligence, as the industry defines it, can be measured. That’s the trap. Anything with a benchmark can be matched. Anything that can be matched gets competed to cost. The benchmark culture the labs built to prove their superiority is the same machinery that guarantees the superiority is temporary.
So the labs are pouring hundreds of billions of dollars a year into building what is, structurally, a grid.
And the thing about grids: nobody got rich owning one. Electric utilities keep civilization alive and pay out like bonds. The grid electrified everything and enriched almost no one who built it.
So where did electricity’s fortunes actually go?
Radio. Hollywood. Refrigerators sold as status objects. Neon Times Square. Amusement parks strung with light bulbs. The money went to the people who took a cheap, abundant, boring input and turned it into something people wanted — which turned out to be mostly entertainment, comfort, and glamour. The frivolous layer, built on top of the serious one, captured nearly all the margin.
Telecom ran the identical script a century later. Carriers spent fortunes on fiber and spectrum and became dumb pipes. Netflix, YouTube, TikTok, mobile games — the trivial things riding the pipes — took the value. The pipe delivers bits at cost. The show on top charges whatever attention will bear.
It offends the engineer in me, but the pattern doesn’t care. The layer that touches human desire keeps the margin. The layer that delivers the electrons earns cost plus a thank-you.
The instinct is to resist this because it sounds unserious. Surely a trillion dollars of intelligence infrastructure pays for itself in cured diseases, written code, automated enterprises — utility in the noble sense.
Some of it will. But notice how each layer gets priced.
Usefulness saturates. A model that files your taxes correctly is worth exactly the price of the cheapest competent model that files your taxes correctly — and there is always a cheaper competent one next quarter. Correct is correct. Utility has a spec, the spec has a benchmark, and we already know what happens to things with benchmarks.
Delight doesn’t saturate, because it was never graded against a spec. It’s graded against your attention. The image of you and your friends rendered before you finish describing it. The companion that remembers what you said at 1 a.m. The character, the game, the feed. There is no eval for wanting more. You can watch the revealed preference in real time: companionship apps out-earn productivity assistants, image generation ships inside social feeds rather than office suites, and the labs that swore they were building scientists keep shipping entertainment. Nobody is confused about where the demand curve slopes. The superficial layer isn’t a distraction from the business. It is the business — the same way it was for electricity, the same way it was for bandwidth.
The one honest objection is winner-take-all: some lab crosses into recursive self-improvement, capability stays scarce, the grid owner becomes the everything-company, and none of this history applies.
Maybe. But every visible dynamic points the other way. Leads compress. Distillation leaks. And the binding constraint at the frontier is energy — the same electrons, the same physics, available to every competitor at the same price per joule. Intelligence per watt converges because physics is shared. Permanent scarcity of intelligence is a bet against everything currently observable, made mostly by people who need it to be true.
The datacenters are the grid. Necessary, colossal, and headed for utility returns. The scarce things are the ones capex can’t buy: distribution, attention, taste. The labs know it — that’s why every one of them is scrambling to become a consumer company before intelligence finishes commoditizing underneath them.
Intelligence will be sold by the joule. Joy will be sold at whatever the market will bear.
zguo0525@berkeley.edu · @Zhen4good