Jensen Huang's invisible hand is tied on AI safety

Oracle

29/09/2026 13:00

The author is a Reuters Breakingviews columnist. The opinions expressed are his own.
By Robert Cyran
NEW YORK, Sept 28 (Reuters Breakingviews) - Economist Adam Smith noted that the butcher, the brewer and the baker don’t provide food out of benevolence, but from their own self-interest. The idea has taken a modern twist with Nvidia unveiling tools to enhance AI safety. Although a welcome development as the industry grapples with existential questions about rogue agents, the chipmaking titan also has plenty of selfish reasons for introducing them.
Like many nowadays, Nvidia NVDA.O boss Jensen Huang frets about bot hacking campaigns. The company recently agreed to pay some $13 billion to buy Hugging Face, a hub for developers that was compromised by OpenAI agents this summer. More broadly, if usage of these autonomous brains takes off, it will require even more data centers stuffed with Nvidia gear.
The trend is already accelerating. Meta Platforms META.O earlier this month rolled out Muse, a personal AI assistant that soared to the top of download charts. There’s plenty of skepticism, too. Americans ranked highest among nations in a Gallup poll last week saying the technology worries them.
Nvidia's two new products sound promising. OpenShell limits what AI agents can do and access, and runs on hardware from multiple manufacturers. Sentry, meanwhile, monitors and enforces what happens in OpenShell, shutting down agents if they drift from their intended tasks or constraints, but only uses Nvidia processing units.
AI adoption will be broader if it's considered less dangerous, helping Nvidia and its shareholders. Sell the safety features, and the $5.6 trillion company will benefit even more. Nvidia's addition of a record $150 billion to an existing $85 billion stock buyback plan provides clear evidence of the financial incentives involved.
Although Huang reckons the market itself can do even more protecting and is urging companies to get their houses in order, it's unlikely to be enough. Microsoft MSFT.O, Oracle ORCL.N and Intel INTC.O are featured partners in Nvidia’s security announcement on Monday, but OpenAI, Alphabet GOOGL.O and Meta Platforms aren’t. Self-regulation is notoriously hard to coordinate.
The far bigger risk: inflated valuations. Model makers Anthropic and OpenAI are mulling giant initial public offerings while others in China and elsewhere chase AI riches, too. Investing in such companies only makes sense if progress will be fast and their software gets used widely. The incentives, therefore, are to push the limits to get a leg up. Tech entrepreneurs also have made clear their preference for asking forgiveness, not permission. In this case, however, there may be no time for absolution.
Follow Robert Cyran on Bluesky.
     
CONTEXT NEWS
Nvidia on September 28 unveiled tools it says will prevent AI agents from acting in an unauthorized manner and could have prevented the Hugging Face attack by OpenAI agents.
One product is OpenShell, which limits actions AI agents can take and what they can access. It is open source and runs on hardware from both Nvidia and other manufacturers. The other is Sentry, which Nvidia says monitors what happens in OpenShell and shuts down an agent quickly if it is drifting from its intended task or operating constraints. Sentry runs on Nvidia Bluefield processing units.
Nvidia also said on September 28 that its board authorized a $150 billion increase to its existing share repurchase program, bringing it to $235 billion. The company plans to complete the buybacks by the end of its 2028 fiscal year on January 31, 2028.

(Editing by Jeffrey Goldfarb; Production by Pranav Kiran)
((For previous columns by the author, Reuters customers can click on CYRAN/robert.cyran@thomsonreuters.com; Reuters Messaging: robert.cyran.thomsonreuters.com@reuters.net))