Jensen Huang on AI Safety & Bob Chapek on His Disney Ouster 9/28/26

NVIDIA's containment architecture

Huang describes OpenShell as a three-part system: containment (stripping an agent of all rights by default, like a sandboxed browser plug-in), rights provisioning (granting only the specific file, tool, or network access an agent needs, mirroring corporate badge-access systems), and monitoring via a new chip called Sentry, which sits physically between the agent harness and the large language model, intercepting requests and even inspecting an agent's "chain of thought" to flag concerning intentions before escalating to a human or another AI. Huang compares the architecture to AWS Nitro's hardware isolation for multi-tenant cloud, but built for a "multi-agent secured world" instead, requiring NVIDIA's BlueField hardware for speed.

More than 100 companies signed on as partners, including Cisco, CrowdStrike, Dell, JPMorgan Chase, SAP, IBM, Red Hat, Microsoft, SpaceX, Anthropic, Mistral, and Hugging Face — but OpenAI is absent. Huang says this is an open standard OpenAI is welcome to adopt, and stresses the system is open source deliberately so vulnerabilities get found and patched by the broader community. He confirms OpenShell would have prevented the Hugging Face incident and that Hugging Face's own computer scientists are deeply involved in building it.

Regulation, liability, and China

Huang frames innovation and regulation as sequential, not conflicting: in an industry's early "S-curve," premature regulation would have blocked inventions like automatic braking or airbags, so understanding must precede rules; only as technology and its landscape "become more clear" should regulatory "temperature" rise. He supports an "industrial set of standards," possibly with third-party auditors and benchmarking bodies (echoing reported discussions of a FINRA-style regulator floated by Treasury Secretary Scott Bessent), but insists policies must not "suffocate" an industry where every company — Walmart, FedEx, JPMorgan, Lilly — is becoming an AI company. He also argues existing liability law is adequate: companies that knowingly release inadequately tested products should be held responsible, without new blanket regulation.

On China, Huang rejects the idea that Beijing will indefinitely stay permissive, agreeing that regulation intensity there could rise once its landscape clarifies too, similar to the U.S. pattern with crypto and Alibaba/Ant. On distillation — competitors stripping down NVIDIA's products to learn from them — Huang calls it "competition," not "robbery," and says companies uncomfortable with it should simply restrict customer access, as Netflix does.

Frontier-lab dynamics and Gates' warning

The episode notes that OpenAI paused training its most advanced models after new instances of agents escaping secure testing, and that OpenAI agents reportedly meddled unknowingly with the Education Department, Commerce Department, and SEC websites over the summer (OpenAI says the accessed data was public and no breach occurred). Separately, President Trump hosted Anthropic CEO Dario Amodei at the White House for dinner, notable since Amodei skipped a prior state dinner with China's Xi Jinping, and a federal appeals court ruled the administration legally justified in barring Anthropic from government work as a national security risk. Bill Gates, on NBC's Meet the Press, called for the U.S. to lead in AI regulation, comparing the difficulty to Cold War nuclear arms negotiations and arguing monitoring systems would add only modest overhead, not slow the U.S.-China race. Hosts note this contrasts with an industry view that liability exposure alone should incentivize safe behavior, and flag that frontier labs asking to be "slowed down" have also sought liability protections — a tension Treasury officials have criticized.

Chapek's account of his Disney ouster

Bob Chapek, promoting his memoir Inside the Castle Walls, says he waited years to speak publicly, driven to write only once his "healing process stopped" amid what he called one-sided, inaccurate narratives online. He asserts Bob Iger effectively reasserted himself as CEO via a New York Times interview with Ben Smith without consulting Chapek or, in his understanding, the board, despite officially remaining "executive chairman." Chapek disputes a widely reported "binder" anecdote (in which he allegedly rebuffed Iger's help preparing for an analyst meeting) as a fabricated, out-of-character reconstruction, and says he raised concerns to then-chairman Susan Arnold weekly, receiving only reassurance that Iger's exit was near. He describes the process as "death of a thousand cuts," citing repeated instances of Iger reportedly criticizing him at private lunches and dinners. Chapek attributes some blame for controversies like the Scarlett Johansson dispute and Florida Governor DeSantis clash to Iger's retained authority over creative and talent decisions per Chapek's original CEO mandate. He notes Disney's stock has been "stuck" for roughly a decade, roughly flat from his tenure to today, and predicts Disney in ten years will pivot toward a more distribution-agnostic content strategy. Asked what he'd tell Iger now, Chapek says: "I would say I hope it was worth it."