NVIDIA is paving the way for a significant expansion in the AI server market with its Blackwell GB200 AI chips, anticipated to ship 500,000 units this year and upward of 2 million by 2025. This growth is expected in part due to the adoption of new packaging technology which stands to revolutionize product supply chains and meet rising demand.
The company’s previous Hopper AI servers showcased impressive economic success but also highlighted critical supply chain inadequacies—particularly in the areas of high bandwidth memory (HBM) and chip on wafer on substrate (CoWoS) supply—that led to extensive order backlogs. To circumvent these issues, NVIDIA aims to shift from CoWoS to the newer “panel-level fan-out” (PFLO) packaging technology by 2025-2026, a step already taken ahead of schedule due to the anticipated high demand for the upcoming Blackwell generation.
PFLO distinguishes itself from CoWoS by incorporating multiple integrated circuits (ICs) on separate silicon wafers and using alternative carrier materials such as laminate or glass. This innovative approach is expected to offer equivalent or superior performance and scalability compared to the CoWoS method. Though there’s a limitation on the suppliers of this new packaging technology at present, companies like Powertech and Innolux are in the running to fulfill NVIDIA’s orders.
Initial reports suggest that shipments are projected to begin with 420,000 units in the latter half of 2024. Production is set to ramp up substantially in 2025, with anticipated figures ranging from 1.5 to 2 million units. This surge is indicative of the major improvements in production capacity and potential revenue boosts that the Blackwell chips are poised to generate.
With each GPU estimated to cost between $35,000 to $45,000 and whole servers possibly reaching $3 million, the financial prospects for NVIDIA are significant. In addition to the hardware itself, advancements such as liquid cooling systems are being integrated to handle the high power consumption, which can reach up to 1200 watts for these GPUs. Industry giant Dell is expected to spearhead the deployment of these high-end servers, which could contain up to eight Blackwell AI GPUs in their configurations.
As the demand for powerful AI processing escalates, NVIDIA’s strategic innovations in production and technology are set to position the company as a leader in both the AI chip market and revenue generation. These powerful strides indicate an exciting future for the AI industry and the intensive computing tasks these chips will empower.
