Nvidia spends $7bn to put a free model on the chips it already sold

Sometime before the end of the year, an engineer at a state-backed lab will be able to download Poolside's model for nothing and serve it on chips that are already racked. Counterpoint Research analyst Marc Einstein told Rest of World the release is aimed at buyers already running Nvidia hardware ^[2]. What Nvidia paid $7 billion for last month was a stake in Poolside and a licence for the tools Poolside uses to build models ^[1].

"The last missing piece was the model layer," Mohammed Soliman, a senior fellow at the Middle East Institute, told Rest of World ^[4]. "Nvidia already owns the chips. Now it is scaling the factory that produces the models that run on those chips, and it is releasing them as open weights," he said ^[5].

The two institutions running the UAE's AI push disagree about whether to take that offer ^[6]. The Technology Innovation Institute, the lab behind Falcon, builds from the ground up. Its chief researcher Hakim Hacid told Rest of World that the value "lies in mastering the underlying technology and retaining control across the entire AI stack" ^[7]. G42, which runs the country's AI infrastructure, joined Nvidia's open-model alliance in July as the only Gulf company in it ^[9]. Its CEO, Peng Xiao, wrote on X at the time that "frontier open models are an essential part of that future, alongside frontier closed models, with security, trust and sovereign control designed in from the start rather than bolted on later" ^[10]. G42 did not answer Rest of World's questions about whether it will use Nvidia's model ^[8]. None of the three state labs named has said publicly that it will run Nvidia's ^[18]. In Saudi Arabia, Humain introduced an Arabic model this month that it had commissioned from a Chinese lab ^[16].

Sam Winter-Levy, a senior fellow at the Carnegie Endowment for International Peace, told Rest of World that the Saudi route shows "there's little full-stack lock-in: The Saudis are using Chinese AI models to build their own models, while using U.S. chips to develop and run them" ^[17]. The dependency that bites now is the one already installed. Most of the software running on those chips is rented from U.S. companies, and those companies decide what the models may do and who may use them ^[14]. Washington approved the advanced chips after G42 sold its stakes in Chinese technology companies two years ago ^[12]. In July it placed the UAE among the countries it treats as its closest technology allies ^[13].

"Sovereignty was never about inventing the foundational model from scratch," Soliman said. "It is about the model's weights, how you fine-tune the data, and, most importantly at this stage, your ability to serve these models to users in the UAE and other markets" ^[11]. Score each workload on those three. A rented model fails the first and the third ^[14]. A free download passes all three and leaves the hardware bill exactly where it was ^[21]. Falcon passes all three and adds the ability to train the next one, which is what TII's own researchers say they are buying ^[7].

If you already own Nvidia racks, the free weights are worth taking for any workload where switching later costs a fine-tuning run instead of a rebuild. Where you have to answer for the model's behaviour yourself, in-house training still costs staff and compute. Free weights take out the licence cost and leave that bill. TII built Falcon in 2023 as a free alternative to U.S. and Chinese models, and it ranked first this year on Counterpoint's index of state-backed models ^[15].