NVIDIA is grappling with significant supply challenges for its H20 AI chips in China, with CEO Jensen Huang indicating a potential nine-month delay for new production. This comes after the U.S. government imposed new license requirements in April, forcing NVIDIA to cancel existing factory slots at Taiwan Semiconductor Manufacturing Company (TSMC). The situation has led to a scramble among Chinese cloud firms for the limited available inventory.
The H20 is currently the most powerful GPU Washington permits for export to China, developed by NVIDIA to comply with previous U.S. export thresholds. However, a new U.S. rule in April 2025 mandated a license for every H20 shipment, effectively disrupting NVIDIA's supply chain. This regulatory shift has prompted fears of further policy reversals, contributing to a frenzied demand for the H20 in China, according to the provided social media content.
Following the U.S. license requirement, TSMC, NVIDIA's primary foundry partner, reallocated its H20 production lines to other customers. Jensen Huang explained at a media event in Beijing this week that:
"restarting production is not as simple as flipping a switch." He further noted that new photomasks, wafer starts, and back-end test slots all require booking, pushing the first large lots: "into Q2 2026."
The reallocation means NVIDIA must now order fresh components and wait for the entire manufacturing process, which Huang estimates will take approximately nine months. Until these new lots are completed:
"Nvidia only has a small pile of chips it made before the rule, so supply stays tight." This scarcity forces Chinese cloud firms to rely on "aging A800 inventories," lean on domestic GPUs, or consider other cut-down parts.
To address the evolving market and regulatory landscape, NVIDIA is developing a new chip specifically for its Chinese clients. This product, named the RTX Pro GPU, is designed to be compliant with U.S. export restrictions. Expected to launch in the second half of 2025, the RTX Pro GPU aims to provide a compliant solution for AI training, although it is anticipated to train models far slower than the H20.