AI Newsround – The Middle East & Gemini

The Middle East – Gimme chips!

• The US has issued a licence for both the UAE and Saudi Arabia to import chips, but with only 35,000 chips each for Humain and G42, I don’t think that the Crown Prince of Saudi Arabia is going to be very happy.
• The state visit of Mohammed Bin Salman to the US has kicked off more investments and agreements between the US and Saudi Arabia, but the import of chips to the Middle East remains a problem.
• This is because the US has issued a licence for Humain and G42 to import 35,000 chips each, which is only a tiny fraction of what they need for 2026 alone, let alone the whole build-out.
• Core 42 (part of G42) and Humain both have 2 data centres each that are under construction in the region.
• Each of these data centres should be 200MW, adding up to a total of 800MW that is currently expected to come online in 2026.
• RFM’s back-of-the-envelope calculation indicates that Humain and Core42 need to obtain a licence to import 400,000 chips each making a total of 800,000 for the region.
• This licence satisfies less than 10% of the 2026 requirement, let alone the other 10.2m chips that will be needed to complete the planned build-out in Saudi Arabia and the UAE.
• I am hoping that this is a trial balloon and that much larger licenses will quickly follow because delays and disappointment will follow if they get held up.
• This is particularly relevant because American companies are increasing their commitments to the region and are investing in assets that are tied to this rollout.
• Qualcomm, for example, has committed to establishing an AI engineering centre at Humain to help with rolling out its new data centre product, while Adobe and AMD have announced partnerships to further the launch of Arabic-specific AI services.
• Other companies like Nvidia, TSMC and so on are also regular visitors to the region, but all of this is put at risk should the region be unable to access the silicon it needs to fulfil its roll-out plans.
• I am hopeful that I am making a fuss about nothing, but it only takes a small problem like this to put a serious dent in the region’s aspirations.

Gemini 3 – Competent but annoying

• Google has released its latest LLM foundation model, Gemini-3, which it claims is a big step up from 2.5 pro and while it is more competent on the benchmarks, it often clutches its pearls and refuses to do some of the more interesting things I ask of it.  
• Gemini 3 has been launched with immediate availability, and in many areas, it represents a step forward.
• For example, on the Humanity’s Last Exam benchmark, it scores 37.5% compared to 2.5 Pro on 21.6%, Claude 13.7% and GPT5.1 on 26.5%.
• The other benchmarks mostly tell a similar story of incremental improvement, which always makes me a little bit suspicious, as there is an incentive to train an LLM to perform against a benchmark rather than for real-world applications.
• In real-world applications, it fares perfectly well but not noticeably better than GPT-5, and just like GPT-5, it still makes all of the stupid mistakes that demonstrate that it can’t reason and has no grasp of causality.
• This is not unexpected as these are symptoms of a system that is based on statistics, and I don’t think it is going to change until the approach itself changes.
• Consequently, Gemini-3 is a steady improvement over 2.5, which keeps it competitive in its market, but Google’s real edge is the 3bn+ users who use at least one Google service.
• Gemini can be pushed to these users in the form of improving the service that they already use, and thereby ensure that they continue to use the Google Ecosystem rather than leave and go to OpenAI.
• However, Gemini-3 is even more annoying than its predecessor in terms of the things that it refuses to do, especially when it comes to drawing pictures.
• For example, the completely innocuous picture of Jensen as a surgeon operating on a data centre cabinet that went with yesterday’s comment on Nvidia’s FYQ3 26 results was drawn by ChatGPT.
• Gemini refused to draw the picture, and Grok-4’s attempt was so hilariously bad that it could not be used.
• This is far from being a problem but it is clear that Google is controlling Gemini much more aggressively than anyone else and there may come a time when this becomes a disadvantage.
• For example, I will always ask ChatGPT to draw a picture first and then test it on Gemini, but most of the time the Gemini nanny pops out and refuses to obey.
• The net result is a steady improvement that maintains Gemini’s competitive edge but falls way short of all of the jaw-dropping commentary that was posted on social media immediately prior to launch.