The guy who architected the chips that made Apple a trillion-dollar company just left to build the infrastructure for everyone else's AI boom.

The Summary

The Signal

Gerard Williams III didn't just work on Apple chips. He was the chief architect behind the A7, Apple's first 64-bit mobile processor, through the M1, the chip that proved ARM could kill x86 in laptops. That's the entire arc of Apple's silicon revolution. Now he's taking everything he learned about power efficiency and architectural innovation to NUVACORE, a company explicitly targeting the AI data center market.

The timing matters. AI inference, not training, is becoming the real infrastructure bottleneck. Every company spinning up ChatGPT wrappers or building agent platforms needs compute that doesn't melt servers or bankrupt budgets. Williams is betting that energy-efficient CPUs, not just GPUs, will win the AI infrastructure race. He watched Apple do it for consumer devices. Now he's doing it for the agent economy.

"The real AI chip race isn't about raw power anymore. It's about performance per watt when you're running inference 24/7."

This isn't Williams' first rodeo outside Apple. He co-founded NUVACORE after a brief stint at another chip startup, but this time the market conditions are radically different. In 2026, every serious AI company is either:

  • Burning cash on Nvidia H100s for training
  • Scrambling to figure out inference costs at scale
  • Realizing that running agents continuously is a totally different cost model than batch training

Williams spent years working alongside Johny Srouji, Apple's chief hardware officer, during the period when Apple went from buying everyone else's chips to making the industry's best. That experience, designing for performance-per-watt in battery-constrained devices, translates directly to data centers where power and cooling are the biggest operational expenses.

The Implication

If Williams is right, the companies that win the agent economy won't be the ones with the biggest training clusters. They'll be the ones running inference cheaply enough to make AI agents economically viable at scale. Watch NUVACORE's first chip specs. If they can deliver CPU performance that competes with current inference accelerators while using a fraction of the power, the AI infrastructure stack reshuffles completely. For builders, this means the cost curves for running persistent agents, real-time inference, and always-on AI services might drop faster than anyone's modeling.

Sources

Bloomberg Tech