When AI starts designing its own chips

For decades, software has run on hardware designed by humans. Now, AI is designing the hardware that powers it.

Considered a productivity upgrade, it’s the first visible step towards a recursive technological loop, where intelligent systems are improving the physical substrate on which they run.

The real bottleneck is hardware

AI progress has been driven by larger models and more compute. But as models scale, a constraint becomes clear: hardware.

Training and running advanced systems require enormous computational resources. Data centres are constrained by power, cooling and availability of materials like silicon. Meanwhile, designing high-performance chips remains a slow and deeply complex process.

Chip design involves solving vast combinatorial problems, placing millions or billions of components while optimising performance, energy efficiency, heat and manufacturability. Historically, this has relied on expert heuristics and iteration cycles that can take years.

When model development progresses in months, but hardware development progresses in years, friction emerges.

From assistance to optimisation

Traditional electronic design automation tools assist engineers. What is happening now is that AI systems are directly optimising chip layouts, tackling placement and routing problems that once depended heavily on human intuition.

Rather than generating text or code, they navigate enormous design spaces on a scale that humans simply cannot explore manually.

AI has already demonstrated that it can produce chip layouts comparable to those designed by experts. The next step is broader orchestration, using advanced models to coordinate larger parts of the design pipeline. That marks a change from tool-assisted engineering to AI-mediated industrial optimisation.

Closing the loop

AI models required better hardware, and that required a structural change. Engineers design chips to support those models, but now AI is working on the chip design process itself. This creates a feedback loop:

1. Better models require better hardware.
2. AI helps design better hardware.
3. Improved hardware enables better models.
4. The cycle repeats.

For the first time, the system is improving the substrate that enables it.

This is recursion at a technical level: a system contributing to the optimisation of its own physical foundation.

The acceleration question

Early evidence suggested AI can design chips. But what happens when hardware and model iteration cycles align? Shorter hardware cycles mean:

• Faster experimentation with new architectures
• Tighter coupling between models and hardware
• More specialised accelerators

Technological acceleration does not require consciousness. It requires feedback loops. When a system starts improving the substrate on which it runs, progress can begin to compound.

For years, we have said that chips are the fuel of AI. Now AI is beginning to refine its own fuel.