The latest

The US chip advantage is turning a build-out of national data centres into an American-centred AI network, even as servers spread across borders. The US and China control close to 90 per cent of compute, investment and notable foundation models, while data centres remain more geographically dispersed because countries with land, capital and electricity can build them. Mehran Gul argues that host nations gain operational capacity, but not strategic autonomy, because the underlying hardware, software and access rules remain concentrated around the US ecosystem.

Details

• Power map: In 2024, the US accounted for 45 per cent of global data centre electricity consumption, according to the International Energy Agency. China represented 25 per cent and Europe 15 per cent, producing a less concentrated map than AI investment, compute and leading models.
• National projects: The UAE is planning a five-gigawatt AI campus, the largest concentration of AI infrastructure outside the US. India has assembled 45,000 GPUs and subsidised domestic use. The EU, Saudi Arabia and Japan have mobilised billions of dollars for AI gigafactories, linking them to technological autonomy and sovereignty.
• UAE conditions: On July 10, the US commerce department upgraded the UAE under export rules, placing it alongside Nato allies. State-linked G42 and Core42 may receive advanced equipment without individual licences, but the status lasts 270 days unless they “become US companies”; Chinese equipment must also be removed.
• Choosing sides: G42 chief executive Peng Xiao said cutting ties “was not easy” and the company was effectively forced to choose sides. The UAE’s access reflects a bargain combining frontier compute with alignment, backed by Abu Dhabi’s capital and defence ties in Washington.
• China’s hardware gap: Chinese models have nearly closed the performance gap with US systems, while China has advanced in research, applications and algorithmic efficiency. Its weakness remains GPUs and the absence of a globally accepted alternative to the Nvidia-centred hardware and software ecosystem.
• Brazilian test: Brazil has announced two national AI computing projects, one involving Huawei and iFlytek and another with Nvidia. Huawei’s Ascend chips are improving, but Ascend-based gigafactories are not yet viewed by buyers as interchangeable with the US stack.
• Installed base: Countries pouring billions into Nvidia-based compute create an installed base that is difficult to displace. The US therefore needs its GPU lead to persist only long enough for organisations using that infrastructure to face steep costs if they switch.

Between the lines

Physical assets may be dispersed while technical control remains concentrated. Once governments invest billions in compute built around Nvidia hardware, replacing that installed base becomes expensive. The US advantage need not last indefinitely if it persists long enough to make switching systems prohibitively costly.

What’s next

The immediate test is the 270-day window for G42 and Core42: retaining licence-free access depends on becoming US companies and removing Chinese equipment from their systems.