Who is speaking, and why it matters
In an episode of the Power On podcast hosted by Mark Gurman for Bloomberg, reported by IT之家 on November 13, 2026, Gerard Williams III spoke: an engineer who led the CPU architecture of Apple's A- and M-series processors. His résumé includes participating in the development of the A7, the company's first 64-bit mobile chip, and the M1, the processor with which Apple moved its Macs from Intel chips to the Apple Silicon platform.
The experience gained on that path, based on internally designed silicon and tight hardware-software integration, is the premise for the arguments presented in the interview. After leaving Apple in 2019, Williams co-founded NUVIA, a company later acquired by Qualcomm, where he led the Oryon CPU architecture; in 2026 he launched a new venture, NUVACORE, active in AI-oriented data center chips.
The CPU behind the scenes of AI
According to Williams, the industry's focus on NVIDIA GPUs must not overshadow the CPU's role in data centers. While GPUs and accelerators perform most of the parallel computing required by language models, the CPU is left with coordination tasks: memory management, security mechanisms, operating system execution, and serial activities that do not lend themselves to parallelization.
On this point his position is clear-cut: the CPU is the brain that holds the infrastructure together, not an accessory. NUVACORE translates this conviction into a design choice called Core First: the core is developed without initially tying itself to a specific instruction set, whether x86, ARM, or RISC-V, and only later is it adapted to the target ecosystem.
From the phone battery to the data center power bill
The most concrete passage concerns energy. Williams argues that the almost obsessive attention to power consumption in the mobile world—where every milliwatt counted for battery life—can be transferred to server processors. Because AI data centers consume power on the order of gigawatts, savings of a few milliwatts per core, multiplied by millions of cores, could allow, with the same current and available space, a computing capacity two to three times higher.
Williams also indicated a possible effect of the new tools on design: the initial performance modeling phase, which traditionally takes seven to twelve months within a two-and-a-half- to three-year development cycle, would have been reduced to two or three months thanks to AI. He described it as an extension of the team's capabilities, useful for exploring more ambitious ideas, but specified that models are not always reliable: final verification remains entrusted to experienced engineers.
What remains to be verified
The figures on energy savings and on the acceleration of project timelines are estimates and statements by Williams, not results measured on available products. NUVACORE is a recently founded company, and the material does not indicate whether it has already produced silicon, customers, or independent benchmarks. Even the comparison between data center consumption and mobile chip consumption, although coherent in units of measurement, is not accompanied by actual consumption data.
