How Nvidia and Big Customers Finally Conquered Blackwell ‘Challenges’
Last year, big Nvidia customers struggled with the complexity of the latest generation of its AI chips and only recently began using them at scale. Future delays could cost them money and create openings for Nvidia rivals.
A year ago, Jensen Huang, CEO of Nvidia, told analysts that transitioning from an earlier generation of AI server chips to a newer one, known as Blackwell, was “challenging” for the company’s customers because of the chips’ greater complexity. “The chassis, the architecture of the system, the hardware, the power delivery, all of that had to change” to improve chip performance, he said.
Indeed, for some marquee Nvidia customers, getting Blackwell servers up and running has been a source of frustration. OpenAI, Meta Platforms and their cloud providers, for instance, struggled to set up and use these systems reliably during much of last year, according to two Nvidia employees who work with OpenAI and other large customers and a Meta employee who has managed the problems. Those same people said the Nvidia AI chips that came before Blackwell, in comparison, were easier to set up and put to work for customers within a few weeks of receiving them.