Kamal Ghorui
Nvidia makes the most sought-after chips on the planet but due to extreme high demand, the supply is so scarce that a top executive recently admitted the company’s own engineers lack sufficient hardware because every unit is already sold out. Bryan Catanzaro, vice president, applied deep learning research at Nvidia has said that whenever his team asks for more GPUs, CEO Jensen Huang says that we have to operate with constrained hardware.
‘Jensen will say sorry…’
According to Fortune, Catanzaro oversees teams working on AI-driven graphics, speech recognition and simulation, and he made the admission at the HumanX conference in San Francisco this week.“My team uses AI very deeply in our work, and their primary complaint is they want higher limits. They want more GPUs,” Catanzaro told Fortune. He further says that when Catanzaro raises the issue with his own CEO, the answer is blunt. “Jensen will say, ‘I’m sorry, Bryan, but those are sold’. We operate within those constraints,” he said. The submission comes as a surprise because the research teams at the company building the world’s most powerful AI chips are constrained by the same scarcity that is being faced by every other organisation trying to build with AI. According to the report, one of Catanzaro’s primary responsibilities is trying to secure enough compute for his own people. “We’re all supply constrained,” he added.
‘Catanzaro saw this coming’
Fortune says that Catanzaro was among the first people to notice that AI researchers were snapping up Nvidia’s gaming GPUs to train machine learning models. That observation helped convince Huang to make a decisive pivot toward AI, investing heavily in the hardware and software infrastructure that would eventually turn Nvidia into one of the most valuable companies in history.While Catanzaro saw this problem, he also found a solution. Rather than simply waiting for more chips, Catanzaro’s team has turned the constraint into a creative force. He leads the development of Nemotron, Nvidia’s family of open-source AI models that users can freely download, study and modify. Unlike the headline AI models from OpenAI, Anthropic or Google, Nemotron is not designed to win benchmark competitions or attract consumer subscribers. It is designed to be lean — to do more with fewer GPU resources than competing models require.“In a supply-constrained world, efficiency is also intelligence,” he said.Catanzaro acknowledged that the project has been under development for a long time but only recently began attracting serious attention inside Nvidia.
