
Gulf states are not buying cloud subscriptions. They are building the whole tower — chips, data centers, models, policy. That changes everything about who owns the future.
ℹ️ ბრაუზერის ხმოვანი წაკითხვა · AI სტუდიო ხმა მალე
The demo that keeps replaying in my head happened in Riyadh last month. Not a slick stage presentation with rehearsed applause — a working session in a government ministry, three engineers, a whiteboard covered in Arabic and English, and a language model running entirely on infrastructure that had never touched a hyperscaler's server. The model was imperfect. It hallucinated on a specific legal edge case. One of the engineers caught it, corrected the prompt chain, and the group moved on without drama. What struck me was not the model. It was the posture: this was a team that expected to fix their own tools, because the tools were theirs.
That scene is the most consequential thing happening in tech right now, and it is almost entirely underreported because the dominant frame for AI infrastructure is still San Jose to Seattle, with occasional detours through London.
Let me say plainly what is actually happening: a cluster of Gulf states, alongside Saudi Arabia's Vision 2030 digital agenda, the UAE's long-running compute investment program, and Qatar's quiet sovereign data push, are executing a strategy that is not about buying access to the frontier. It is about owning a sovereign AI stack — compute, model weights, data pipelines, and the regulatory architecture on top. All of it. Simultaneously.
This is not a press release strategy. The data centers are physical. The chip procurement conversations are real and ongoing. The model fine-tuning labs exist and are hiring. The policy frameworks — particularly around data residency and Arabic-language corpus rights — are being written by people who have read the European mistakes carefully and intend to skip them.
The business model question here is actually several questions layered together, and they are worth separating.
First: who pays for the build? The sovereign wealth fund model answers this in the Gulf — PIF, Mubadala, QIA are not looking for eighteen-month payback windows. They are infrastructure investors on the scale and timeline of port construction. The compute capex that would terrify a VC-backed startup is a rounding error against a fund that measures positions in decades. That is a structural advantage that the Bay Area actually cannot replicate, because the Bay Area's capital is answerable to LP quarterly letters.
Second: who pays for the run? This is where it gets interesting. Government ministries are the anchor tenants — defense, health, education, judiciary, national ID systems. That is a captive revenue base with essentially no churn risk. The commercial layer on top — enterprise services to regional businesses, export of model capabilities to neighboring markets, licensing of Arabic-language foundation models to North African and South Asian markets — is the upside that justifies the rest. If that commercial layer materializes at scale, the economics are extraordinary. If it does not, you have a very expensive national project that mostly serves the anchor tenants. Both outcomes are defensible; only the second one is a business in the conventional sense.
Third: can you build a frontier model this way? Honest answer: not yet. The weights being trained and fine-tuned in Riyadh and Abu Dhabi are downstream of architectures and base training runs that still originate in the US and, increasingly, China. Sovereign does not mean cutting-edge. What it means is that the application layer, the data, and the inference infrastructure are under local control — and that the dependency on foreign model providers is being methodically reduced, not eliminated overnight. The engineers in that Riyadh ministry were not pretending to have built GPT-5. They were running a fine-tuned derivative on sovereign hardware and treating the gap as a problem to close, not a reason to stop.
Tel Aviv deserves a mention here too, though for different reasons. Israel's AI stack strategy is less about sovereign infrastructure and more about sovereign talent density and export — building foundation model capabilities that get acquired or licensed outward. The model is closer to a research university with commercial spinouts than a national compute project. It is a different bet, and so far a productive one, but the geopolitical ceiling on that model is real and the engineers in those companies know it.
Back to the Gulf. The thing that makes this moment genuinely different from the previous wave of Gulf tech announcements — and I say this with full memory of the smart city vapor that evaporated between 2018 and 2022 — is that the current build is happening in layers that actually connect. You cannot fake a data center. You cannot fake a network of Arabic-language engineers doing RLHF on legal and medical corpora. You can fake a press conference about it, but the underlying work has a physical trace. The physical trace is visible.
The open question that nobody in the region is fully comfortable answering yet is governance of the stack once it is built. Sovereign infrastructure is only as trustworthy as the sovereign. The same data residency guarantees that make a Gulf AI stack attractive to regional enterprises also mean there is no external check on how that infrastructure gets used for domestic purposes. This is not a reason to dismiss the build — it is a reason to watch it. The European Union spent years building GDPR as an external accountability layer for private infrastructure. The Gulf is building its own infrastructure, and the accountability layer is still being designed by the same parties who own the hardware. That tension does not resolve itself.
The column I am not writing today is the one that says Gulf AI is exotic and aspirational and maybe someday. That column was wrong in 2022 and it is obviously wrong now. The column I am writing is: the sovereign AI stack is no longer a whitepaper concept, the business model is partially but not fully proven, and the governance question is the one that will determine whether this is infrastructure or apparatus.
The engineers at that whiteboard in Riyadh did not seem worried about being asked that question. They seemed like people who expected it, and were already arguing about the answer.