Friday, 26 June 2026

The semiconductor industry is having a moment where the physics is finally catching up with the ambition. IBM's announcement of sub-one-nanometer chip technology feels significant not because of the marketing—we've heard this before—but because the constraint that's been limiting AI inference performance is genuinely shifting. Smaller transistors mean denser computation, lower power consumption, and the ability to run larger models on edge devices. That matters for every company trying to reduce their cloud compute bills. By the way, this also matters geopolitically in ways that often get undersold. The ability to manufacture advanced chips at scale is becoming as strategically important as oil was fifty years ago.

What strikes me more immediately is the convergence happening in custom silicon for AI workloads. Broadcom and OpenAI's Jalapeño chip is not trying to be general-purpose. It's built specifically for LLM inference, which means every design decision is optimized for what these models actually do at scale. This is the opposite of the era when companies tried to squeeze AI onto GPUs designed for gaming. When you have the volume and the leverage that OpenAI does, building custom hardware makes sense. I'd expect more of this—not everyone can do it, but the companies that can will gain a meaningful cost advantage.

The more interesting pattern, though, is how AI agents are becoming the vehicle for enterprise adoption. Salesforce's Help Agent for customer service is straightforward: prepackaged, ready to deploy, handles a specific workflow well enough that companies don't need to hire ML engineers to make it work. That's the unsexy but crucial part of AI's real impact—not the frontier model that everyone reads about, but the agent that actually answers your customer service call better than the previous system. Research from Pacific Northwest National Laboratory showed something similar: an agentic AI interface that lets scientists talk to robots made experiments faster. None of this is shocking, but it validates something important: agents aren't just a research curiosity anymore. They're becoming the default interface between people and complex systems.

I find it worth noting that GitHub's Copilot agentic harness is showing strong results across multiple benchmarks while maintaining token efficiency. That's the metric that matters for actual users—not peak performance on a one-off task, but consistent competence with reasonable resource consumption. The gap between what cutting-edge models can do and what you can actually deploy profitably is closing, and that gap-closing is where the real economic value lives.

One thing I keep thinking about: as these systems become more autonomous and more widely deployed, the safety conversation can't stay in the nonprofit briefing circuit. When agents are making decisions at scale in customer service, supply chains, and laboratories, the question of how they fail—not whether they fail—becomes a business problem, not just an ethics problem.