Independent AI research firm SemiAnalysis argues making increasingly powerful models safer may itself consume substantial computing power, which may calm fears that Nvidia Corp. (NASDAQ: NVDA) may face the fallout from slower artificial intelligence development.
Safety Work Eats Compute
Speaking on SemiAnalysis's podcast Tuesday, analyst Max Kan said the market may be overlooking two factors.
First, AI companies still want more computing capacity than can physically come online.
"Demand will continue to outstrip supply," Kan said.
Second, safety itself requires compute. AI labs increasingly use models to monitor other models, test safeguards and check the reinforcement-learning environments used to improve their systems.
Kan said people "underestimate how much compute is required for safety, alignment" and monitoring. He also pointed to interpretability, research aimed at understanding how models reach their answers, as another compute-intensive workload.
Another host said the "nuanced take" is that labs could actually "spend a lot more compute" as they devote more resources to making increasingly capable systems safer.
Anthropic CEO Dario Amodei has made a similar point. His proposal to pace frontier AI does not call for halting training, but for devoting more resources to alignment, interpretability, testing and operational safeguards.
Cybersecurity Cuts Both Ways
SemiAnalysis also argued that advanced AI could reshape cybersecurity by forcing companies to fight AI with AI.
The panel pointed to Greg Brockman's description of a continuous security system at OpenAI that would use the latest models to constantly find and patch vulnerabilities. As attackers deploy autonomous agents at scale, defenders may need equally capable AI systems running continuously to detect and stop them.
One speaker called it potentially "the most impactful shift in cybersecurity," arguing that companies capable of delivering security outcomes rather than simply selling tools could capture enormous value.
The hosts warned that major cybersecurity firms could themselves become prime targets for increasingly capable AI attacks, naming Okta Inc. (NASDAQ: OKTA), CrowdStrike Holdings Inc. (NASDAQ: CRWD), Palo Alto Networks Inc. (NASDAQ: PANW) and Zscaler Inc. (NASDAQ: ZS).
Near-Term Pause Still Bites
OpenAI paused frontier reinforcement-learning training after its Hugging Face incident, and its largest planned frontier run remains on hold while smaller training and evaluations continue.
President Donald Trump on Monday called fears of runaway AI a "hoax" and questioned why an industry would seek rules that could bankrupt it.
Polymarket traders give a U.S. AI safety law a 16% chance of passing before 2027, with about $123,000 traded. The contract covers measures including restrictions on AI training or use and requirements for human oversight.
For Nvidia, the question is whether slower frontier training cuts more GPU demand than safety, monitoring and cybersecurity add back. Shares recovered 0.6% Tuesday.