律动BlockBeats
律动BlockBeats|11月 19, 2026 04:43
[NVIDIA Blackwell Contributes 74% of New AI Inference Capacity; GB300 Single Card 25 Times That of H100] Beating AI News: According to the latest estimates from research firm Epoch AI, NVIDIA's Blackwell series chips contributed 74% of the global increase in AI agent request processing capacity over the past year. Specifically, a single GB300 card, under specific test conditions, can handle approximately 25 times the number of agent requests as the H100 and 8 times as many as the H200. Programming agents require repeatedly reading code, retaining historical dialogues, and calling tools. In addition to computation, the chip also needs to store large amounts of context in its GPU memory and quickly read data. GPU memory capacity, read/write bandwidth, and communication speed between multiple chips all affect the number of agents that can run simultaneously. Measured by traditional computations per second, the peak computing power of the GB300 is only about 2.5 times that of the H100. However, considering these factors, Epoch AI estimates its agent request processing capacity to be approximately 25 times greater. From Q2 2025 to Q2 2026, the cumulative number of agent requests that AI chips sold globally could simultaneously handle increased 4.6 times, while peak computing power only increased 2.6 times during the same period. Nvidia's interim CEO, Colette Kress, confirmed in the earnings call on November 17 that the company has stopped shipping licensed Hopper series data center chips, including the H200, to China. Huawei plans to launch the Ascend 960DT in Q1 2027, with announced FP8 peak computing power specifications close to that of the H200. [Original Link]
+4
Mentioned
Share To

Timeline

HotFlash

APP

X

Telegram

Facebook

Reddit

CopyLink

Hot Reads