One particle or 125,000
In the first half of 2013 I was a drilling engineer trainee working a derrickman assignment, on two rigs of different design. While a rig is drilling, a big part of the derrickman's job is the mud. The mud carries the rock cuttings up from the bottom of the well, and they have to come out before it goes back down. A cutting left in goes down again and comes back smaller.
The rig takes them out with a line of machines, each one good at one size. First the shakers, which are vibrating screens. Then the desander, the desilter, the mud cleaner, and at the end the centrifuge, the only one that reaches the finest particles. The rule is that no stage should pass the next one more than it can handle.
My report for that assignment compared the two rigs item by item, down to which one had the more comfortable shaker room. It also has this sentence: it is much easier to take out one particle of 100 microns with a screen than 125,000 particles of 2 microns with a centrifuge.
The number is arithmetic, not a figure of speech. A 100-micron particle, about the thickness of a hair, is 50 times wider than a 2-micron one, so it holds 50 × 50 × 50 of them. Same rock. The screen takes it out in one piece. Let it go round the well a few more times and the centrifuge has to find 125,000.
In customer conversations this year we keep seeing the opposite order. I call it brute force. Take the most expensive model there is, put every file and all the data in the context, and let it crunch. It works, eventually. It also costs a lot of tokens, because you pay the model to read everything, including what a lookup would have caught on the first pass.
The bill is starting to show. A Gorilla Logic report on 10 September said token spend went from a technical footnote to a board-level budget line in under a year, and that agents make it worse, since every prompt carries thousands of tokens of background with it. The price of a token has been falling the whole time. What grew was the count. Thousands of tokens of background in every prompt is the whole flow going to the centrifuge.
We measured the difference once. In one of our deployments, the same problem was solved by brute force on off-the-shelf AI platforms and on ours, where tools fetch and compute what can be fetched and computed, and an ontology, a map of what each tag and table is, sits between the data and the model. Ours used six times fewer tokens.
That is the screen doing its job. The model still does the part only it can do. It just gets there with less to read.
Six is one measurement on one problem, not a constant. It will move with how much of the next problem is reading and how much is judgement.
Since 1 June, GitHub Copilot bills each interaction by the model and the tokens it consumes. Before, a prompt cost the same whether it carried one file or the whole repository. My bet, dated 27 September 2026: within a year, most platforms that sell AI to industry will bill by the token, and brute force will get more expensive with every one that switches.
Diego Mercadal started as a commissioning engineer on high-voltage motors and generators, joined an offshore drilling contractor as a rig hand, worked several positions in the drilling crew, and ended up running its AI/ML function. He is now co-founder and CEO of Wonder DataLabs.