GPU crunch hits Nvidia teams
A global GPU shortage reportedly affects even Nvidia’s own research groups, highlighting how tight supply is constraining experiment velocity and model work across the industry. The report frames compute scarcity as an operational limit that influences retraining cadence, model size choices, and whether teams favour distillation or smaller specialist models. (fortune.com)
At the HumanX conference in San Francisco this week, Bryan Catanzaro said even Nvidia’s own research teams cannot get all the graphics processing units they want. The surprise is that Nvidia is the company making many of the chips everyone else is fighting over. (humanx.co) (finance.yahoo.com) A graphics processing unit is the part of a computer built to do many math operations at once. That makes it the workhorse for training artificial intelligence models, the same way a combine harvester does in hours what a field crew would do in days. (finance.yahoo.com) Catanzaro runs applied deep learning research at Nvidia, with teams working on graphics, speech recognition, and simulation. He told Fortune that when he asks chief executive Jensen Huang for more chips, the answer can be no because there is not enough supply to go around. (aol.com) (dnyuz.com) That shortage changes how research gets done. If a team cannot reserve enough machines for a full-scale training run, it trains less often, tests fewer ideas, and waits longer to see whether a model actually works. (finance.yahoo.com) It also changes what gets built. Instead of one giant model that needs a huge cluster, teams lean toward smaller specialist models or toward distillation, which means training a compact model to imitate a larger one. (aol.com) This is not just an Nvidia problem. OpenAI president Greg Brockman said last year that deciding who gets graphics processing units inside OpenAI is “pain and suffering,” which is a blunt way of saying compute has become an internal rationing problem. (finance.yahoo.com) (indexbox.io) The pressure is strongest on Nvidia’s newest Blackwell systems, which are the company’s latest high-end chips for artificial intelligence data centers. Nvidia’s February 25, 2026 earnings call said demand stayed strong enough for management to guide revenue to about $78 billion for the next quarter. (finance.yahoo.com) When a scarce tool gets more valuable, the biggest buyers usually get first access. Cloud companies and model labs can lock in giant orders, while smaller companies, universities, and even internal research groups end up competing for what is left. (finance.yahoo.com) (financialcontent.com) That is why this story is less about one awkward comment from one Nvidia executive and more about a hard ceiling on the whole field. In April 2026, the limiting factor for many artificial intelligence teams is not ideas, data, or talent, but the number of graphics processing units they can actually get onto a rack. (finance.yahoo.com)
