Nvidia has implemented strict prepayment requirements for Chinese customers ordering H200 AI accelerators due to unresolved regulatory approvals, with over 2 million chips ordered despite supply constraints.
Nvidia has instituted mandatory full prepayment terms for Chinese customers purchasing its H200 AI accelerators, according to industry sources familiar with the matter. This policy shift comes amid ongoing uncertainty regarding Chinese regulatory approval for the advanced GPUs, which are subject to U.S. export restrictions.
The new contractual terms require Chinese clients to pay 100% of the order value upfront—typically $27,000 per H200 unit—with no option for order cancellation or configuration changes post-commitment. Limited alternatives include commercial insurance or asset collateral, but these represent exceptions rather than standard practice. This contrasts sharply with previous arrangements allowing partial deposits for Nvidia hardware in China.
Regulatory approval timelines remain ambiguous, with Chinese authorities expected to potentially authorize H200 imports for commercial entities in early 2026 under specific conditions. Military, government, and state-owned enterprises will likely remain prohibited from procurement due to security concerns. Additionally, regulators reportedly require Chinese firms to purchase domestic AI accelerators alongside imported H200s, causing order processing delays.
Despite these hurdles, demand significantly outpaces supply. Chinese tech companies have placed orders exceeding 2 million H200 units against Nvidia's immediately available inventory of approximately 700,000 chips. Initial shipments will draw from existing stock, targeting delivery before the Lunar New Year in mid-February.
Fulfilling the remaining 1.3 million+ orders requires new manufacturing cycles. Each H200 involves TSMC's 4nm silicon fabrication followed by CoWoS-S packaging—a process requiring over three months per production batch. Subsequent shipments aren't expected before Q2 2024.
This prepayment strategy mitigates Nvidia's financial risk amid regulatory unpredictability while capitalizing on intense demand for AI compute. The H200, featuring 141GB of HBM3e memory and 4.8TB/s bandwidth, delivers near-doubled inference performance versus its H100 predecessor for large language models.
Anton Shilov is a semiconductor industry analyst covering processor architectures and supply chain dynamics.
Comments
Please log in or register to join the discussion