Jensen Huang is out. Rubin is Nvidia’s first real test without him

Jensen Huang has resigned from the company he co-founded in 1993, leaving Nvidia in the middle of its Vera Rubin rollout. The board accepted his resignation as chief executive, president and director on 11 November and announced it the next day. Colette Kress, Nvidia’s finance chief since 2013, will keep that job while serving as interim chief executive as the board looks for a permanent successor.

The affair that became a boardroom issue

On 8 November, Nvidia published an email to staff in which Huang acknowledged being unfaithful to his wife. He gave no details about the relationship. Reuters reported that he had an affair with Alice Huang, a former executive at Singapore cloud computing company Megaspeed whom US authorities suspect of spying for China. Beijing denied the spying allegation. Megaspeed, an Nvidia customer, said it had cooperated with US inspections and took its export control obligations seriously.

That put Nvidia’s board in the position of having to look into whether its chief executive had been targeted by a foreign intelligence effort. When it opened its independent inquiry, the board said it had no reason to believe US national security or Nvidia’s secrets had been compromised. It later brought in former US federal prosecutor Joon H. Kim and his law firm to conduct the investigation. Nvidia gave no reason for Huang’s departure in its resignation announcement. Its shares fell about 14% in early US trading on Thursday, touching $148.80 against the previous close of $173.73 before recovering some of the loss.

The hard part is getting a whole rack out the door

Nvidia said in May that Vera Rubin was ramping into full production. Its DGX Vera Rubin NVL72 puts 72 Rubin GPUs and 36 Vera CPUs into a liquid-cooled rack. The GPUs use HBM4 memory and exchange data across the rack over Nvidia’s NVLink switches. When a model is split across dozens of GPUs, how quickly those processors can exchange data helps determine how much time they spend computing and how much they spend waiting.

There’s a second networking job once a workload spans several racks. NVLink connects the GPUs inside each NVL72 rack, while links such as InfiniBand or Ethernet connect racks to one another. So a data centre can take delivery of its first racks and still have work to do before it can run the workloads it bought them for. The network needs to be ready, along with the electrical supply and the liquid cooling system that takes heat out of the racks.

That makes delivery dates matter further down the supply chain than the GPU order alone suggests. A server maker can have its processors in hand and still be waiting for switch trays or power shelves to finish a build. The cloud operator may already have booked installation crews or promised customers additional capacity. Nvidia’s new chair, Harvey Jones, said the board expects the product roadmap and customer commitments to continue without interruption. For those customers, the date that matters is when the rack can go into service.

AMD already has a foot in the door

In July, AMD and Anthropic announced plans to deploy up to 2 GW of AMD Instinct systems, with deployment of the first gigawatt due to begin in the first half of 2027. They’re also working together to tune workloads for AMD’s GPUs and develop its ROCm software. That’s a big part of making a second supplier viable. Before moving a workload, an engineering team needs to know which software already runs on the other platform, what needs tuning and how the result performs on its own models.

If Rubin deliveries slip, the orders most open to competition are likely to be for new capacity, where customers have more choice about which platform to build around. That gives Nvidia’s rivals a chance to win more work without customers having to move systems they’ve already spent time tuning for CUDA. Nvidia reports on 17 November, and one of the most telling things Kress could give the industry is a firm answer on when customers will be able to put their Rubin racks to work.