NVIDIA Blackwell AI Server Overheating and Glitching Problems Prompt Major Clients like Microsoft and Google to Reduce Orders

NVIDIA’s upcoming Blackwell AI servers are reportedly encountering significant supply chain challenges, as the company grapples with issues related to overheating and architectural shortcomings in the design.

Delays in Blackwell AI Server Deliveries Prompt Customers to Turn to ‘Hopper’ Series

The launch of NVIDIA’s Blackwell AI servers has not gone as anticipated, indicating that Team Green is currently facing a substantial hurdle. Originally projected to commence volume production by Q4 2024, it appears that the new AI architecture is beleaguered by a design flaw that has resulted in overheating issues. Although NVIDIA had previously maintained that these problems were resolved, an investigative report from The Information indicates otherwise, characterizing the Blackwell AI servers as experiencing “glitches.”

This report highlights that the initial significant shipment of NVIDIA’s GB200 AI servers has been plagued by overheating and glitching problems, predominantly stemming from issues with the “chip connectivity.”Such challenges have raised concerns among major customers, including Microsoft, Amazon, Google, and Meta, leading to a reduction in their orders for Blackwell servers, despite these companies previously placing agreements surpassing $10 billion.

The emerging situation is critical for NVIDIA and poses a potential threat to the profitability of its AI segment, as supply chain disruptions can have severe financial implications. While details surrounding the exact nature of the problem remain unclear, prior insights suggested that TSMC’s advanced packaging technology, known as CoWoS, is at the center of the chip connectivity issues mentioned earlier. NVIDIA did announce modifications to the Blackwell GPU mask produced at TSMC; however, these changes have yet to yield satisfactory outcomes.

Consequently, many customers are reverting to more reliable alternatives, such as those offered by NVIDIA’s established Hopper generation, while awaiting solutions for the lingering issues with Blackwell. As of now, the extent of the impact on NVIDIA’s financial performance remains uncertain. However, without timely resolutions, the future success of the Blackwell architecture is in jeopardy, which could pose further complications for NVIDIA moving forward.

Source & Images