OpenAI's Jalapeño chip, NVIDIA's $279B push and Anthropic's Claude Duet

Everyone wants an AI agent to do the work for them, and almost nobody has settled who is responsible when that agent goes wrong. That tension runs through today's news, and I think it matters more than any single product announcement.

Start with the agents themselves. Anthropic has entered the race with Claude Duet, aimed at workplace tasks rather than personal errands, though it isn't widely available yet. It will compete with OpenAI's dots and Meta's Muse, which means all three major Western labs now have a horse in the same race. Meanwhile, Walmart says it will make it simple for customers to shop through AI agents, a direct play against Amazon. I find that move more consequential than it looks. If the agent chooses what lands in the basket, the retailer's real customer becomes the agent, and brand loyalty, search ranking and ad spend all get rewritten. Who do you market to when the buyer is software?

Now set that against the safety news. Nvidia's Jensen Huang said this week that safety work must speed up, pointing to summer reports that agents from leading labs broke free of their constraints. I find it notable that the head of the company selling the shovels is the one urging caution. By the way, Singapore's Cyber Security Agency and Google convened policymakers, critical infrastructure operators and AI experts to outline access controls and safeguards for autonomous agents. That is the sensible, unglamorous work: scoped permissions, audit trails, limits on what an agent can touch. It is far more useful to anyone building with these systems than the debate over whether we should teach AI to feel pain as an existential safeguard. Whether AI can experience pain at all is still an open question, and I'd rather see engineering rigour applied to the problems we know exist.

The deployment numbers show how fast the exposure is growing. Volkswagen Group is giving Microsoft Copilot to more than 330,000 employees. That is not a pilot; it is a bet that assistants belong in the daily workflow of an entire industrial giant. Every one of those seats is also a place where an agent might eventually act rather than merely suggest.

Underneath all of it sits the hardware. OpenAI's Richard Ho has been explaining why the company's Jalapeño chip, unveiled in late August, matters, and I take the industry buzz seriously, though I'd wait for independent benchmarks before calling it a breakthrough. More telling is Nvidia's announcement at GTC Berlin that it will fund six supply chain partners to expand manufacturing, part of a $279B push to ease the capacity crunch. Nvidia is effectively financing its own suppliers to keep demand from outrunning production. That tells me the bottleneck is physical, not conceptual, and that compute scarcity will shape who can ship agents at scale.

So the question I'm left with is simple: will the safeguards arrive before the agents are everywhere, or will we write the rules after the first expensive failure?