The familiar shape of the trade
Every economy on this coast knows the pattern by heart: extract something raw, ship it out, buy back the refined product at a multiple, and call the difference the cost of not owning a refinery. For a century the raw thing was hydrocarbon. In the model economy, the raw thing is us: our text, our questions, our behavior, exported as training data and inference demand, returned as capability, priced per token.
The slogan writes itself, and officials have started saying it: compute is the new oil. I work one desk away from where these deals land, and I think the slogan is half right in a way that makes it dangerous.
What the slogan gets wrong
Oil is scarce, rivalrous, and valuable sitting in the ground. Compute is none of these: it depreciates on a schedule, its supply grows with every fabrication cycle, and a warehouse of accelerators with nothing to run is a heating bill. Buying compute is not striking oil; it is buying drills. The question the slogan skips is the only one that matters: refining into what, by whom, for which users.
The refined product of this industry is not the model file. It is deployed capability: a system that answers a Gulf user in the Gulf user’s language, inside the institution’s constraints, at a cost the institution can carry. Everything upstream of that is inventory.
The refinery, itemized
If capability is the refined product, the refinery has four units, and none of them is a chip. First, data refining: cleaned, licensed, documented Arabic corpora, the mop work we open-sourced for Kuwain because every team was reinventing it. Second, evaluation: benchmarks written from the language rather than translated into it, which is why our models travel with Tarjama-25, SadeedDiac-25, or Misraj-DocOCR attached. Third, deployment engineering: serving, guardrails, dialect handling, the unglamorous last mile where value is delivered or lost. Fourth, and slowest to build, people who can do the first three without a vendor in the room.
Notice what this list does to procurement logic. A sovereign model announcement buys headlines; a sovereign evaluation culture buys leverage over every future announcement, foreign or domestic.
Own, rent, or ally
A state does not need to own everything, and pretending otherwise wastes a decade. The useful frame is a three-way sort. Own what encodes your language and your judgment: Arabic data pipelines, dialect capability, evaluation suites, and the teams behind them, because these are cheap relative to compute and impossible to import honestly. Rent what is generic and fast-moving: frontier generality is a competitive global market, and renting keeps you current. Ally on what is capital-intensive and shared: compute and energy, where the Gulf’s real comparative advantage, cheap power and patient capital, buys a seat rather than a logo.
The failure mode is inverting the sort: renting your evaluation from the vendor you are evaluating while owning a depreciating hall of accelerators. That is buying drills and outsourcing geology.
A ledger for ministries
Five questions before the next AI memorandum
- Who evaluates the delivered system, with which Arabic benchmark, written by whom?
- What is the cost per served answer for our users, in our dialects, at our scale?
- Which data leaves, under what license, and what refined asset returns?
- How many nationals can retrain, re-evaluate, or replace this system in three years?
- If the vendor vanished tomorrow, what capability remains on our soil?
None of these requires owning a frontier lab. All of them require owning the refinery units above, and every one can be answered with a number.
The coastline, again
From my window the Gulf does not look like a metaphor; it looks like the route the crude took out and the refined goods took in, for longer than I have been alive. The region eventually learned to build refineries, then petrochemicals, then the universities that staff them. It can run the same play faster this time, because the refinery for language is mostly people and discipline, not steel.
Compute is the new oil only if we repeat the old mistake. The weights are not the wealth. The ability to make, measure, and deploy them is.
Sara Chrouf
@misc{chrouf2026oilweights,
author = {Chrouf, Sara},
title = {From Oil to Weights},
year = {2026},
url = {https://sarachrouf.com/essay-from-oil-to-weights.html}
}