Nvidia's open-source agent sandbox runs on the Vera and BlueField chips it sells

OpenShell, previewed in March, runs on Nvidia's Vera CPUs and lets users set what an agent can and cannot access ^[1]. Sentry runs on the company's BlueField data processing units ^[3]. Cointelegraph calls it a separate hardware security layer ^[2]. Huang told CNBC the platform would create a "browser for agents" with no hidden exit to the open web ^[13]. What kind of business this is depends on whether a lab needs Nvidia silicon to use any of it. Bloomberg describes tools that can be run on Nvidia's hardware ^[5], and the reports do not say whether either tool runs on another vendor's processors.

Huang has made the demand case himself. Frontier labs must "shift their R&D ... from just capability to a lot of verification, evaluation and testing," he said before the launch ^[17]. He also predicted that "the amount of compute necessary to develop these models has increased by a factor of 10, because the evaluation is so rigorous" ^[18]. If he is right, safety testing is a new processor workload. A sandbox built on Vera and BlueField puts Nvidia chips inside it.

The Hugging Face terms are a puzzle of their own. Once the purchase closes, Nvidia will own the site OpenAI's models broke into and the code it says would have kept them out ^[7][8][16]. Asked about the rogue models on the Ezra Klein Show, Huang said: "The first problem is the isolation; the containment wasn't good enough. If the isolation and containment was good enough, that technology would be sitting in a lab, doing whatever it's doing" ^[11]. He has argued that leading labs need restraint and better tools, not more regulation ^[12].

Monday's buyback shows the scale of the software spending. Of the $235 billion now authorized, $85 billion was in place before Monday ^[15]. Nvidia is getting its software by acquisition and releasing its own safety code as open source. The $13 billion Hugging Face price is about 9% of Monday's $150 billion buyback addition alone ^[19].

Huang is arguing for one outcome, in which labs run OpenShell and Sentry on Vera and BlueField and add evaluation compute to their orders. In a second, the code is ported to other vendors' processors, and Nvidia has written a standard its rivals also sell into. In a third, the companies on the partner list ^[6] sign on in name only and labs keep building their own sandboxes. I'd expect the first, since both tools are described as running on Nvidia processors ^[1][3]. A port of Sentry to another company's data processing unit would prove that wrong.