Nvidia’s Blackwell AI chips face overheating challenges

Delays in Nvidia’s Blackwell chip production raise concerns among customers.
Chinese firms are accelerating Nvidia H20 chip purchases as DeepSeek’s affordable AI models gain traction in multiple industries.

Nvidia is grappling with challenges related to its highly anticipated Blackwell AI chips. Customers have raised concerns over overheating issues in its custom server racks, which are critical for training large-scale AI models. The racks, designed to house 72 AI chips each, have undergone multiple design revisions late in the production process. Despite these setbacks, Nvidia remains optimistic about meeting its shipping deadline by mid-2024.

Dell has already begun shipping Nvidia’s GB200 NVL72 server racks to customers such as CoreWeave. Nvidia described the engineering iterations as a normal part of integrating advanced systems into diverse data centre environments. The company highlighted its collaboration with leading cloud service providers to ensure successful implementation.

Past delays in Blackwell production were attributed to a design flaw, which Nvidia’s CEO Jensen Huang openly acknowledged. The flaw, linked to low production yields, required extensive collaboration with Taiwan Semiconductor Manufacturing Company to resolve. While these issues temporarily slowed progress, Nvidia remains on track for its long-term goals.

Nvidia is set to release its fiscal third-quarter earnings on Wednesday, with analysts projecting revenue of $33 billion and net income of $17.4 billion. Although shares dipped slightly on Monday, the stock has soared by 187% this year, underscoring investor confidence in the company’s AI-driven future.