AI & Tech Brief — July 16, 2026

The Big Story

OpenAI is making a regulatory bet that sounds counterintuitive: instead of waiting for a single federal AI rulebook, the company is advocating for what it calls "reverse federalism" — letting states build their own AI safety frameworks first ^[1]. The theory is that state-level experimentation can inform a stronger national standard, rather than the typical approach of Washington mandating rules top-down.

This matters because it signals how the AI industry sees governance happening over the next few years. Right now, there's no unified US AI safety law. Europe has its AI Act. China has its own approach. The US has mostly relied on voluntary commitments from big labs. OpenAI's proposal suggests the company believes state action — not federal action alone — will shape the rules of the road.

At the same time, OpenAI is also unveiling GPT-Red, an automated red teaming system that uses self-play to stress-test its own models ^[2]. Red teaming is the practice of trying to break AI systems — testing for biases, jailbreaks, prompt injection attacks, and other failure modes — before they go public. The twist here is automation: instead of hiring teams of humans to poke holes, OpenAI is using AI to find vulnerabilities in AI. This is a concrete technical answer to the safety problem that governance frameworks are trying to address.

The two announcements together paint a picture of OpenAI's strategy: legal and technical guardrails, moving in parallel. One is political (state governance); one is computational (automated red teaming). Neither alone is enough.

What Else Moved

IBM's AI Problem Meets a Mainframe Reality Check

IBM announced preliminary results that missed expectations, and the software market reacted negatively ^[3]. But the real story, according to analysis, is specific to IBM itself: the company has a legendary mainframe franchise — its oldest, most profitable business — that's insulated from competition but also largely insulated from AI opportunity ^[3].

Mainframes are the massive computers that banks, insurance companies, and governments have relied on for decades. IBM's mainframe business is so entrenched that customers are locked in; they can't easily switch. That moat (a term for a competitive advantage that's hard to cross) is real. But it's also a drag on AI growth. While the rest of the tech industry is racing to build and deploy AI, IBM's core business is in a mature market that doesn't drive the kind of AI revenue that investors want to see. IBM has AI problems — plural — and a mainframe business that doesn't solve them.

DeepSeek, Kalshi, Bonsai on the Horizon

DeepSeek, a Chinese AI lab, is planning an IPO ^[4]. Kalshi is launching compute markets (platforms where people can trade contracts based on computing power prices) ^[4]. Bonsai has developed a phone-optimized AI model ^[4]. These moves suggest the AI infrastructure market is expanding — not just who builds the biggest language models, but how compute gets priced, traded, and deployed on smaller devices.