NVIDIA Open Agent Safety Platform 拆解：OpenShell 与 Sentry
NVIDIA 新平台含开源运行时 OpenShell（跑在 Vera CPU）和 BlueField-4 上的看门狗 Sentry，可在毫秒级隔离失控 agent。
$NVDA have moved AI safety into the chip. Today they launched the Open Agent Safety Platform. It has two parts: > OpenShell: an open-source runtime that sets hard limits on what an agent can do, running on Vera CPUs > Sentry: a watchdog on BlueField-4 chips that runs separately from the agent, monitors it continuously, and can quarantine it within milliseconds More than 100 partners have signed on, including Anthropic, $MSFT, $CRWD, $PANW, JPMorgan and $PLTR. What this means for the AI thesis: 1. Trust is the bottleneck for agents. Enterprises won't let agents touch money, data or code unless they can govern them. Compute isn't what's holding back the next leg of adoption. Permission is, and Nvidia is going after that bottleneck. 2. The market Nvidia sells into keeps getting bigger. An agent doesn't need just a GPU. It also needs a CPU to run the agent, a DPU to watch it, and networking to connect them. Every safety layer is more silicon. 3. The moat is getting deeper. The software is open, but it runs best on Nvidia hardware. That's the CUDA playbook applied to agents. 4. The trade is moving from training to deployment. Safety infrastructure only gets built when agents are about to go into production at scale. The first phase of AI was about building the most capable models. This phase is about making them trusted enough to deploy. $NVDA clearly wants to sit on both engines here. Never endless, this is very bullish AI in general and for $IREN and $NBIS. -BP Not financial advice.
这条是引用着另一条写的：x.com/jensenhuang/status/2104499465055023424