Iqbal Latif
Sma, you said one California almond uses more water than thirty-eight thousand queries to your machine. You meant to make a simple point, a query is cheap.
Grant it. A query is cheap. Cheap at producing what?
That is the question I cannot find your industry answering anywhere, and it is the only one that matters at the scale now being built.
America consumes about 4,195 billion kilowatt hours of electricity a year and produces an economy of roughly thirty one trillion dollars, which is about seven dollars forty of national output for every kilowatt hour burned. An aluminium smelter, the crudest heavy industrial use of power there is, turns a kilowatt hour into roughly twenty cents of revenue. Artificial intelligence, on the only public figures available, lands somewhere around one dollar thirty. Twenty cents. One dollar thirty. Seven dollars forty.
The comparison is imperfect and I will state the imperfection before you do. Revenue is not value added. Nobody publishes a clean split between your electricity and ordinary cloud electricity. Much of the value your systems create is captured by users and never appears in your revenue. So one dollar thirty is not a verdict against artificial intelligence. It is something worse for the present argument. It is evidence that nobody has built the statistic that would allow a verdict at all.
Your industry can tell us the chips, the tokens, the megawatts, the gigawatts, the cooling load, the capital expenditure, the training run, the accelerator count and the cost of the next building. It cannot tell us, in an agreed and auditable economic unit, what comes out.
That should matter now, because the sums have stopped being experimental. Commentators tell us technology investment is driving a remarkable share of American growth this year. Examine what that growth presently contains. Concrete, steel, copper, transformers, turbines, chips and construction labour. All of it real output, all of it real income, and there is no accounting error anywhere in it. But building an asset is not evidence that the asset will earn its keep. The national accounts can record the building of the cathedral. They cannot record one prayer said inside it. So we have a public argument treating the construction of the thing as proof of the productivity of the thing, while the series that would separate the two does not exist.
Your almond comparison has the same fault in miniature. Thirty eight thousand is not a measurement. It is the widest published almond footprint divided by the narrowest published query estimate. Take California’s actual orchard water against a narrow query boundary and it falls to around twenty thousand. Broaden the query to include the water burned generating the electricity and it falls toward the low thousands. The honest conclusion is not that some other figure is the true one. It is that nobody holds a current, independently audited, like for like measurement precise enough to justify saying thirty eight thousand at all.
But I grant you the underlying point again. The query is cheap. Now tell us what the cheap query produces.
I have spent two years working on a measure I call knowledge per kilowatt hour, and the idea is elementary. Do for machine intelligence what the world eventually learned to do for oil: stop celebrating the quantity consumed and measure the output obtained from it. Three public series would transform this argument. Value added per kilowatt hour of compute. Output per quality adjusted unit of completed work. Return against the true annual economic cost of the capital. Not tokens, not benchmark scores, not utilisation, not booked capacity, not floor space, and not another forecast of how many gigawatts the world will need. Delivered economic value.
Here is why I put it to you rather than to a statistical agency. You are in a better position than almost anyone alive to begin publishing the first two. You know the compute consumed. You know the classes of work being performed. You know how both change across model generations. You have enterprise customers capable of measuring what happened to cost, time, throughput and quality before and after deployment. You could publish a serious pilot series before any agency on earth could build one. If it shows that your machines produce ten dollars, or fifty, or five hundred of attributable value for every kilowatt hour consumed, the argument ends overwhelmingly in your favour, and I will publish that result in the same size of type as this challenge.
But if the output cannot be stated, then structurally short of compute is not yet an economic shortage. It is a statement about capacity. Short of compute for what? For whose work? In place of what? Producing how much additional value, at what cost, and how much of that value survives long enough to pay the people financing the buildings?
There is another reason this cannot wait. Every generation of hardware becomes more efficient, which is wonderful for the technology and may be brutal for the capital, because the better the next chip, the sooner the present one stops being worth running. Artificial intelligence can therefore succeed magnificently while a generation of investors fails completely, and those two statements are not in conflict. That is why I proposed the Latif Capital Survival Test. A delayed productivity argument can justify today’s capital only if today’s capital, or something valuable and appropriable created by it, survives long enough to share in the return. If the great gain arrives after the processors are obsolete, the buildings impaired and the original money written down, the later success vindicates the technology. It does not vindicate the investment. That distinction matters when trillions are being discussed.
So I am not asking whether artificial intelligence is useful. It plainly is. I am not asking whether inference is getting cheaper. It plainly is. I am not asking whether users receive benefits the accounts miss. They plainly do. I am asking the question underneath all three. What economically valuable output do we obtain for the energy, the capital and the completed work consumed in producing it?
You have spent years making intelligence measurable technically. Now make it measurable economically. Publish the series. If the numbers vindicate the build, you will have done something far more important than comparing a query with an almond. You will have given the world the missing unit.
Until then we can state the input. We cannot state, in any agreed and auditable unit, the output expected to justify it. That is not an argument against artificial intelligence. It is a demand that an industry asking the world for trillions learn to count what comes out.
