Nvidia has begun volume production of its next-generation Blackwell AI accelerators at TSMC, according to supply-chain sources familiar with the matter. Demand from cloud-service providers is said to outstrip available supply through the fourth quarter, reinforcing the company's dominant position in the AI training market.
Blackwell chips use TSMC's 4NP process and advanced CoWoS-L packaging, representing a significant step up in performance and complexity from the previous Hopper generation. The ramp is critical for Nvidia as it seeks to maintain leadership against custom silicon efforts at Amazon, Google and Microsoft.
Supply-chain checks indicate that TSMC is prioritizing CoWoS capacity for Nvidia, but shortages of HBM3E memory and complex substrate materials could limit initial shipments. Lead times for Blackwell-based systems are stretching into early 2025 for some customers.
The news lifted shares of Nvidia suppliers across Asia, including TSMC, Foxconn, Wistron and cooling-component makers. Analysts raised their calendar-2025 revenue estimates for Nvidia, citing higher average selling prices and improved supply.
MetalSemi Asia sees the Blackwell ramp as a key catalyst for the AI semiconductor supply chain through year-end. Risks include any yield issues at TSMC and potential export restrictions on advanced AI chips to China.
