NVIDIA Set to Ship 500,000 Blackwell AI Chips This Year

NVIDIA will ship 500,000 Blackwell GB200 AI chips in 2024, with projections to reach 2 million units in 2025. This comes after the success of the Hopper AI products last year, which highlighted supply chain issues, particularly with HBM and CoWoS.

Shifting to New Packaging Technology

To overcome these challenges, NVIDIA plans to switch to a newer packaging technology, the Panel Level Fan Out (PFLO) standard, by late 2025 or 2026. This shift is aimed at resolving CoWoS bottlenecks and meeting the anticipated massive demand for Blackwell products.

PFLO integrates multiple individual ICs on separate silicon wafers, using materials like laminate or glass instead of silicon. This technology is expected to enhance performance and scalability, although specific data comparing PFLO with CoWoS is not yet available.

Early Adoption of PFLO

NVIDIA has already adopted PFLO ahead of schedule due to the expected high demand. Suppliers for the new standard are currently limited, with Taiwan’s Powertech and Innolux vying for orders. NVIDIA plans to ship 420,000 units in the second half of 2024, with production figures for 2025 estimated between 1.5 million and 2 million units.

High-Cost, High-Performance GPUs

The Blackwell GPUs are priced between $35,000 and $45,000 per piece, with servers costing up to $3 million. Dell is expected to roll out servers equipped with up to 8 B200 Blackwell AI GPUs, featuring liquid cooling to handle the high power consumption of up to 1200W per GPU.

NVIDIA’s strategic shift to PFLO and increased production targets reflect the company’s efforts to address past supply chain issues and capitalize on the growing demand for AI technology.