Nvidia’s free model tests what sovereign AI means in the UAE

The UAE's flagship AI data centre is planned around 400,000 Nvidia chips. Nvidia now intends to offer a free open-weight model, expected to be called Nemotron 4, by year-end—giving the country a way to run and adapt powerful software locally rather than rent every capability through an American API.

Nvidia assembled the model effort through a $7bn stake and technology-licensing deal with coding startup Poolside, and it has agreed to pay almost $13bn for Hugging Face, the main distribution platform for downloadable models. The aim is strategic: Chinese open models were downloaded roughly 2bn times this year, and governments often begin sovereign projects from Meta's or Alibaba's weights. Giving away Nemotron can increase demand for the hardware and software environment Nvidia still sells.

The UAE is divided over what sovereignty requires. The Technology Innovation Institute, which built Falcon, argues that mastery depends on developing a model from the ground up. State-backed infrastructure group G42 joined Nvidia's open-model alliance and treats inspectable, locally served foreign weights as compatible with national control. G42 had already cut Chinese ties to secure advanced US chips, and Washington placed the UAE among its closest technology allies in July.

Saudi Arabia shows why lock-in remains incomplete: it buys Nvidia hardware while commissioning an Arabic model from a Chinese laboratory. Open weights let Gulf states fine-tune, host and govern a model themselves, but the training tools, chips and upgrade path can still come from abroad. Nvidia's offer expands operational control while binding the UAE more closely to an American supplier at the layer below.