Artificial Intelligence – The Middle East pt. III

Microsoft gets a licence for less than 1% of what the region needs in 2026 alone

• Microsoft has rolled out what look like big numbers in terms of investing in AI in the UAE, but the numbers it is putting out are actually pretty small and only for in-house data centres, meaning that the licenses that the UAE needs to build 5GW of compute capacity have still not been issued.  
• This means that although Xiao Peng, CEO G42, is in the photograph with Brad Smith, this announcement and investment have very little to do with G42 or sovereign AI in the UAE.
• ADIPEC is the Abu Dhabi International Petroleum Exhibition and Conference, and it is here that Brad Smith, President of Microsoft, is to be found, explaining Microsoft’s increased commitment to AI in the UAE.
• Microsoft’s investments in the UAE now include: $1.5bn invested in G42 in 2023, $4.6bn in data centre capex and $1.2bn in OPEX to the end of 2025.
• From Jan 2026 to the end of 2029, Microsoft will invest a further $7.9bn of which $5.5bn is AI-related capex and $2.4bn in OPEX.
• These investments are clearly being made in Microsoft / Azure data centres that will be used to offer AI models to the UAE (which has the highest per capita usage of generative AI in the world at 59%) and the rest of the global south.
• To be clear, I do not think that these are part of Stargate or other sovereign AI projects as the numbers are way too small.
• It is in this context that Microsoft’s chip numbers make a little more sense.
• Microsoft says that it has accumulated the equivalent of 21,500 A100 chips and has secured a license to ship a further 60,400 A100 equivalents to the UAE.
• It is a mystery to me why Microsoft is giving these numbers in chips that are now effectively obsolete, being 4 generations old, but it does make the headline numbers for the less informed look better.
• The A100 will produce 312TFLOPS at FP16, which is the equivalent of 1.2 PFLOPS at FP4, while the GB300 (which is what Microsoft is going to export) will do 30 PLOPS at FP4, meaning that it puts out 25x more compute per chip than the A100.
• Consequently, an A100 in GB300 terms means that Microsoft will ship a total of 2,416 GB300 chips, which in the NVL144 configuration will mean a total of 34 racks or cabinets.
• With each rack consuming 200kW, this means that the USA has licensed the export of 4.8MW of Nvidia GB300 compute capacity to the UAE.
• Saudi Arabia and the UAE are planning to roll out 11GW of capacity in the next 5 years or so, with 600MW already under construction and expected to come online in 2026.
• This means that Microsoft has gone to great effort to secure a licence for less than 1% of what the region needs for 2026 alone.
• Consequently, I think that the promised commitment to the region by the USA has yet to be lived up to, which is most likely due to the level of mistrust that still exists in Washington.
• Microsoft’s investment in G42 goes a long way to assuaging some of that nervousness, but until a licence is granted to export 11m chips to Saudi Arabia and the UAE, the whole strategy hangs in the balance.
• I am pretty confident that these licences will be issued, but it is taking far longer than it should, and both Core G42 and Humain need 200,000 and 400,000 chips each just to bring their under-construction data centre projects online on time in 2026.
• Following that, they will need a further 10.4m chips within the next 5 years if they are going to meet the targets that they have laid out.
• RFM research indicates that Nvidia, Micron and SK Hynix are pretty much sold out of silicon chips for AI in datacentres for the whole of calendar 2026, so unless Core 42 and Humain are already in the queue, they are going to miss their targets.
• The net result is that while I am very positive on the strategy to diversify the economies of the Middle East further away from petrochemicals via the export of low-cost compute, the practicalities are starting to get in the way.
• However, oil demand is not collapsing as many forecast, and in fact, I think it could easily continue to rise for many years to come, meaning that the outlook for the Middle East economies remains pretty secure.
• Hence, if the rollout takes 5 years longer than anticipated, this is not a particular problem and would give the region more time to ensure that the other infrastructure that this will need is in place when the data centres are ready to be turned on.
• This leads me to remain pretty positive on the outlook for this new industry in the Middle East and to expect that it will become a significant provider of compute, in the same way that it is a significant provider of energy today.