OpenAI and Broadcom unveil AI inference chip for large-scale data centres

The agent revolution is accelerating on multiple fronts, and we're now seeing the infrastructure race catch up with the capability race. That matters because it suggests the bottleneck is shifting from "can we build this?" to "can we deploy this at scale?"

Start with Google's move. Gemini 3.5 Flash now has native computer use—meaning it can see screens, click buttons, and navigate interfaces the way a human would. This isn't theoretical anymore. What strikes me is the practical implication: agents can now operate across legacy systems without requiring API integrations or custom bridges. That's a massive force multiplier for enterprise adoption, especially in industries drowning in poorly connected software. By the way, Google's decision to bake this directly into the model rather than bolting it on afterward suggests they learned from earlier missteps.

But capability means nothing without infrastructure. OpenAI and Broadcom's announcement of Jalapeno—a custom inference chip built for LLM workloads at scale—addresses the real constraint now: cost per inference at massive throughput. The economics of running agents at hundreds of thousands of simultaneous tasks only work if you can push the hardware efficiency far beyond what consumer chips allow. I find it telling that OpenAI is vertically integrating here. It suggests confidence that inference margins, not just model capability, will define competitive advantage.

Trase's $107 million funding round lands in this context too. A company building an operating system for agents in healthcare and high-stakes industries wouldn't attract that kind of capital unless investors believed the unit economics were real. The shift from "chatbot interaction" to "delegated long-horizon task" is profound—it changes what we're asking AI to do and how we measure whether it's working.

Then there's the safety question. Cambridge's warning about AI falling into criminal and state hands isn't new in substance, but the urgency feels different now. When agents could soon be operating autonomously across business systems, the surface area for misuse expands dramatically. A compromised agent running procurement workflows or medical diagnostics has orders of magnitude more impact than a compromised chatbot. I'm not convinced the safeguards are keeping pace, and that gap is real.

Microsoft's move to ship MAI-Code-1-Flash more broadly across Copilot tiers shows the commodification accelerating too. Better models are becoming a baseline feature, not a differentiator. The race is on to embed agents into workflow, and whoever connects them most seamlessly to actual work will win.

We're at an inflection point where the architecture for agentic AI is crystallizing—better chips, better models with native computer use, operating systems to orchestrate them. The question now is how quickly safety, governance, and business models adapt to follow.