Beating AI News Flash: Research firm Epoch AI's latest estimate shows that Nvidia's Blackwell series chips contributed 74% of the global increase in AI Agent request processing capacity over the past year. Among them, under specific test conditions, a single GB300 card can simultaneously handle about 25 times the Agent requests of an H100 and 8 times that of an H200.
Coding Agents need to repeatedly read code, retain historical conversations, and call tools. In addition to performing computations, chips also need to use VRAM to store large amounts of context and quickly read data. VRAM capacity, read/write bandwidth, and the communication speed between multiple chips all affect how many Agents can run simultaneously.
Measured by traditional calculations per second, GB300's peak computing power is only about 2.5 times that of H100. But after factoring in these elements, Epoch AI estimates that its Agent request processing capacity reaches about 25 times.
From the second quarter of 2025 to the second quarter of 2026, the cumulative volume of AI chips sold worldwide can simultaneously handle 4.6 times the original number of Agent requests, while peak computing power during the same period only increased to 2.6 times.
Nvidia interim CEO Colette Kress confirmed on the November 17 earnings call that the company has stopped shipping licensed Hopper series data center chips to China, including H200. Huawei plans to launch the Ascend 960DT in the first quarter of 2027, with published FP8 peak computing power specifications close to those of H200.
