
Nov 14, 2026 · 47 min
Chip veteran says efficient CPUs still anchor the AI era
Weekend Listen: Apple’s Former Chief Chip Architect on the AI Chip Race
As AI accelerators dominate attention, Gerard Williams III makes the case that efficient CPUs remain essential to building practical systems and bringing useful AI to everyday devices.
- 1Efficient CPUs still coordinate memory, security, and workloads around specialized AI accelerators.
- 2AI is compressing early chip-design timelines while experienced engineers remain central to the process.
- 3Smaller language models could move useful AI from data centers onto phones, laptops, and desktops.
Don't miss
Williams explains how AI reportedly shortened NuvaCore’s early CPU performance-modeling work from seven to twelve months to roughly two or three.
The brief
Gerard Williams III, Apple’s former chief CPU architect and Nuvia founder, traces his path from the A7 era to building NuvaCore amid the AI boom.
Williams argues that AI accelerators do not eliminate the need for powerful, efficient CPUs: they still orchestrate memory, security, and surrounding workloads.
His clean-sheet strategy keeps the core central while leaving room to connect CPUs and accelerators through chiplets or complete data-center systems.
AI is already changing chip design at NuvaCore, where early CPU performance modeling reportedly fell from seven to twelve months to roughly two or three.
The conversation ends with a consumer-facing question: whether smaller language models can make AI useful on ordinary devices instead of concentrating it in data centers.
What was said on this episode
20 statements · 17 positive · 1 mixed · 2 neutral
Saving milliwatts per CPU materially reduces power across large data centers.
“if you can save milliwatts in every CPU that you constructed, by the time you multiply that out by tens, hundreds, thousands, if not millions of cores, it's a material change that you can make.”
Listen at 7:15
Lower CPU power can enable two- to three-times more cores in a data center.
“If I can lower the power and I can suddenly squeeze 2, 3x the number of cores into that data center building, I've won.”
Listen at 8:02
Empowered engineering teams can produce exceptional results.
“if you empower a group, they can do some pretty amazing things.”
Listen at 9:40
Apple’s silicon organization continues delivering and developing year after year.
“Amazing organization year on year. They're delivering and developing.”
Listen at 12:04
High-performance CPUs are necessary for large language models to operate.
“those high-performance CPUs are necessary in order to enable these large language models to do their work.”
Listen at 19:06
CPUs orchestrate memory, security, and fundamental workloads for language models.
“They orchestrate memory, they orchestrate the security aspects, and do some of the fundamental work.”
Listen at 19:15
A well-designed CPU can scale across computing applications.
“the CPU that you build has the ability to scale.”
Listen at 20:24
NuvaCore aims to target its CPU principles across x86, ARM, or RISC-V.
“make sure the CPU's fundamental principles that we build are held so that they can be targeted at any one of the three at the time that we want to do our final.”
Listen at 22:56
Smaller language models on personal devices can expand AI experimentation for students and engineers.
“I think that is a huge opportunity for The engineering community, including the up-and-coming engineers, those students that are in elementary, high school, and university level, to be able to understand and experiment with those technologies.”
Listen at 25:19
AI models extend chip-design teams rather than replacing them.
“these models, it does not replace your design team, but it's an extension of your design team.”
Listen at 26:35
Traditional clean-sheet CPU development typically takes two and a half to three years.
“Without that event of AI, that development cycle in its best typically was around two and a half to three years.”
Listen at 30:16
NuvaCore reduced early CPU performance modeling from seven-to-twelve months to two-to-three months.
“what would have taken seven months to 12 months to get to a point where you are ready to start the design process, we've actually collapsed that down to a two or three month window,”
Listen at 30:45
AI can explore CPU state spaces and improve verification testing.
“You can apply it directly there because it is what I call a very brute force technique where you can have it explore the state space of a CPU, understand what to test, how to test it, and then release it onto the design to basically find out, is it working? Yes or no?”
Listen at 31:39
AI could compress CPU development schedules from two-and-a-half-to-three years toward two years.
“With the end goal being that you compress that entire two and a half to three year schedule into something that may look like a two year type of design point.”
Listen at 31:59
Full clean-sheet CPU development may eventually be compressed substantially.
“there's a possibility for that to become true.”
Listen at 32:45
NuvaCore could develop CPU chiplets integrated with accelerators.
“that could turn into chiplet form of silicon with many CPUs on a die that's co-integrated with accelerators.”
Listen at 33:50
NuvaCore could eventually produce a complete blade-based data-center compute product.
“It could turn into a full-scale end product like you see in classical data center and hyperscaler, which is a blade-based compute device.”
Listen at 33:59
NuvaCore expects full silicon development to take two to three years.
“Full silicon is somewhere between a two to three year journey. in order to create that silicon.”
Listen at 35:20
Within several years, everyday users will experience LLMs through CPU and compute technology.
“what we're going to discover is this utilization of LLMs and CPU and compute technology that every end user will get to experience”
Listen at 40:04
AI should improve people’s working and daily lives rather than primarily create fear.
“my hope is that that actually is what happens with all the end users, not this fear that you see in the industry wherever, oh, AI is going to do this or do that. No, it's about can it make my working life that much better, my daily life.”
Listen at 40:33
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.