NVDA Raises TSMC Orders by 25%, GB200 Demand Estimates Increase by 50%
According to reports from the supply chain, TSMC is preparing to start producing NVIDIA’s latest Blackwell platform architecture graphics processor (GPU) in the near future. In response to strong customer demand, NVIDIA has increased its chip investment in TSMC by 25%, which not only means that the AI ??market is booming Unprecedented, it also adds strong momentum to TSMC’s performance in the second half of the year and leaves the groundwork for raising its full-year outlook.
TSMC will hold a legal meeting this Thursday (18th) and is currently in the silent period before the legal meeting. As of yesterday's deadline, no response from TSMC was received.
The industry believes that with the news that TSMC has begun producing Blackwell platform architecture graphics processors (GPUs), it means that the countdown to the launch of Nvidia's AI server equipped with "the most powerful AI chip on the surface" will open a new page in the AI industry, and is expected to become TSMC's The Dharma Conference is the focus of heated discussion.
Analysts estimate that the average selling price (ASP) of the Nvidia B100 GPU built on the Blackwell architecture is US$30,000 to US$35,000, and the super chip GB200, which connects the Grace CPU and the B200 GPU in series, is priced between US$60,000 and US$70,000. The U.S. dollar is even higher, which means that Nvidia-related chips are the most expensive chips that TSMC has ever produced for terminals, and its investment in TSMC's operations is even more anticipated.
Nvidia's Blackwell architecture GPU is known as the "most powerful AI chip on the planet". It is equipped with 208 billion transistors and is manufactured using TSMC's customized 4nm process. The GPU die is twice the mask size through a 10TB chip per second. The chip interconnect connects into a single, unified GPU and supports AI training and real-time large language model inference, with models scalable to 10 trillion parameters.
According to industry insiders, major international manufacturers such as Amazon, Dell, Google, Meta, and Microsoft will introduce Nvidia's Blackwell architecture GPU to build AI servers, and the capacity exceeds expectations. To this end, Nvidia has increased its order for TSMC by about 25%.
As Nvidia expands its Blackwell architecture GPU wafer production, in terms of the number of terminal server cabinets, shipments of GB200 NVL72 and GB200 NVL36 server cabinets have increased simultaneously, from the original expected combined shipment of 40,000 units. increased to 60,000 units, an increase of up to 50%, among which the total number of GB200 NVL36 reached 50,000 units.
According to industry estimates, the average selling price of the GB200 NVL36 server cabinet is US$1.8 million, and the price of the GB200 NVL72 server cabinet is as high as US$3 million. GB200 NVL36 has 36 super chips GB200, 18 Grace CPUs, and 36 enhanced B200 GPUs; GB200 NVL72 has 72 super chips GB200, 36 Grace CPUs, and 72 B200 GPUs.
Before the handover of the leadership in June, former TSMC Chairman Liu Deyin predicted that the demand for AI applications is now more optimistic than a year ago. "This year is another year of great growth." Current Chairman Wei Zhejia also revealed that AI applications have just begun. , "I'm as optimistic as everyone else."
AI
Market Data 
Markets 
