National data centre projects are consolidating America’s AI lead
Reported: the US and China, Gul writes, control close to 90% of compute, AI investment and notable foundation models, while data-centre electricity use is more dispersed — the IEA’s 2024 figures give the US 45%, China 25% and Europe 15%. The UAE is planning a 5GW AI campus (the largest concentration of AI infrastructure outside the US); India has assembled 45,000 GPUs and is subsidising domestic use; the EU, Saudi Arabia and Japan have mobilised billions for AI gigafactories. On 10 July the US Commerce Department upgraded the UAE’s export status to sit alongside NATO allies and approved G42 and Core42 to receive advanced equipment without individual licences — but only for 270 days unless they “become US companies”, and with Chinese kit stripped from their systems. G42 CEO Peng Xiao said cutting those ties “was not easy”. Brazil has announced two national AI computing projects, one with Huawei and iFlytek and one with Nvidia. Interpretation (Gul): hosting servers is not sovereign AI. Dispersion of hardware is not decentralisation of power. Washington is converting its remaining uncontested GPU/software-stack lead into a geopolitical bargain — access to frontier compute in return for alignment — before China can offer a substitute. The British-empire railway analogy is his: tracks on sovereign territory, a network centred on America.
Why it matters
This is a direct hit on Dealroom’s European competitiveness and industrial-capacity brief. “Sovereign AI” and “AI factory” labels will proliferate in European policy and in startup/VC decks, but the article’s test is whether Europe is building independent capability or becoming a location inside a US compute system. For Dealroom that means tagging data-centre, neocloud and “sovereign model” companies by who owns the chips, the software stack, the offtake and the export licence — not by where the building sits. It also frames why European and Gulf capital pouring into gigafactories may still leave startups renting American intelligence.
Executive takeaways
- Gul’s core distinction is operational: land, capital and electricity can buy a data centre; they do not buy OpenAI- or Nvidia-like capability. Europe’s 15% share of data-centre electricity is not a 15% share of AI power.
- The US+China ~90% claim on compute, investment and notable models is presented as the baseline; data centres are the one layer that still looks multi-polar — which is why governments are pouring money into them.
- Concrete programmes cited: UAE 5GW campus; India 45,000 GPUs with domestic-use subsidies; EU, Saudi Arabia and Japan “billions” for gigafactories. These should be tracked as capacity announcements, not as proven training clusters.
- The UAE/G42 case is the mechanism: export-licence upgrade on 10 July, G42 and Core42 licence-free access to advanced US equipment, 270-day clock unless they become US companies, and a forced dump of Chinese equipment. That is industrial policy as alliance management.
- China is described as having closed much of the model-performance gap but remaining weakest at GPUs and lacking a globally accepted substitute for the Nvidia-centred stack. Brazil’s dual Huawei/iFlytek and Nvidia projects are the early contest for that installed base.
- Dealroom implication: score “sovereign AI” startups and infrastructure funds on stack lock-in (Nvidia vs Huawei), licence contingency, and who actually trains/runs the models — not on megawatts announced.
What Financial Times may be missing
The 90% compute/investment/model figures are not sourced. The piece does not show utilisation, offtake contracts, or which models will actually be trained in the UAE, Indian, EU, Saudi or Japanese facilities — hyperscalers occupying European or Gulf halls would confirm rather than contradict the thesis, but that split is not in the article. Europe is a residual 15% in the IEA table; there is no discussion of EuroHPC/AI Factories, access rules for European startups, or the share of European sites already contracted to US clouds. China’s own data-centre build-out and Huawei’s export pipeline beyond the Brazil mention are thin. Power availability, grid queues and local political consent — the constraint in 31 August FT data-centre coverage — are absent. Software, talent, data and safety regulation, the other sovereignty layers, are out of scope.
