September 16, 2026
Nvidia's Jensen Huang told TechCrunch this week that AI doesn't need external regulation because it isn't some "alien mind" -- it's hardware and software, and safety can be built in by whoever makes the product. That's a comforting message if you're the company selling the chips that everyone else's AI runs on. It's also a convenient one: the fewer rules governing how AI gets deployed, the faster Nvidia's customers buy more silicon. I don't think Huang is wrong that safety has to be engineered at the product level -- it does. But treating that as a substitute for any outside guardrail assumes every company will make the same safety tradeoffs Nvidia would, and history with software security, cars, and finance suggests that's optimistic at best.
The same day, TechCrunch reported that OpenAI, Anthropic, and Google DeepMind have been in safety talks for weeks -- even as the Trump administration reportedly waves off safety concerns in favor of racing China. That's a striking contrast to Huang's framing. If safety were purely a self-contained engineering exercise, competitors wouldn't need to coordinate with each other on it at all. The fact that three rival labs are talking suggests even the people closest to the technology think some shared norms matter more than pure market competition. For a business buyer, this is the tell worth watching: when the vendors start comparing notes on safety without being told to, that's usually a sign the risks are real, not theoretical. It's also worth remembering our own take a few days back that when Altman and Amodei started talking caution, the words mattered less than whether the labs actually followed through.
While the safety debate stays theoretical, the physical cost of AI is becoming very concrete. TechCrunch reported that US data centers could burn through more natural gas than Germany and Japan combined by 2035, and that the data center boom is now colliding with cities like Philadelphia that already carry scars from heavy industry -- in this case, a neighborhood near a defunct oil refinery. This is the argument regulation skeptics tend to skip: safety isn't just about model behavior, it's about who bears the environmental and community cost of running these systems at scale. Local officials and residents don't get a vote on Nvidia's chip roadmap, but they're the ones living next to the power plants and substations that make the roadmap possible. That tension isn't going away, and it's a much harder one to engineer your way out of than a jailbreak.
Most companies evaluating AI tools aren't thinking about gas turbines or Senate hearings -- they're thinking about whether a given product is reliable and won't blow up their operations. That's exactly why the self-regulation debate matters practically, not just politically. If you're choosing between vendors, the ones voluntarily coordinating on safety and publishing real practices are giving you more to evaluate than the ones simply asserting they've got it handled. It's the same reason we think transparency about data handling belongs in the Trust Center and Incident Response process rather than just a marketing claim -- trust has to be demonstrated, not declared. Businesses adopting AI agents for things like workflow automation should be asking vendors the same question regulators are implicitly asking labs: show your work, don't just tell us to trust the hardware.
Do you think AI companies policing themselves is enough, or does the data center backlash in Philadelphia change your answer?
Sources