NVIDIA Blackwell, in its first hardware benchmark for intelligent agents, can support 64,200 concurrent agents per megawatt of power consumption, with an efficiency more than 20 times that of H200.

This figure is not just about scores. It signifies that the inference costs for AI Agents are being significantly reduced at the hardware level.
Inference providers like Together AI have already deployed Blackwell first, providing real-time inference for programming agents like Cursor.
As agents move from demos to commercial closed loops, the efficiency differences in underlying hardware will directly determine who can run economic models.
But the flip side is: AMD MI355X lags noticeably in the same test, despite the evaluation agency noting it was not deeply optimized.
If NVIDIA’s hardware advantage in the Agent era continues to expand, the cost structure of the entire AI application layer will become more dependent on a single supplier.
For the crypto market, the narrative around AI Agent-related tokens needs to be re-evaluated—once the computing power bottleneck loosens, genuine demand growth on the demand side may finally materialize.
$ai #ai #Blockchain #加密市场 #Crypto Circle