CharlieKey takeaways from your favorite podcasts
← 9 October 2026a16z

Pat Gelsinger on Why This is the Best Time to Build Hardware

Pat Gelsinger argues that AI has made chip design easy, so the bottlenecks move to fabrication time, memory, power and energy, which is why hardware now has a renaissance of opportunity.

Watch on YouTube ↗

Charlie's summary

  1. When AI makes chip design easy, the bottleneck moves to manufacturing: 3 months to design, 9 months to use

    Pat says AI tools could produce a strong chip design in three months, but fab takes about three months at best, advanced packaging another three, and rack-scale integration the rest. “Nothing's a chip anymore. It's a rack.” By then a year and a half may pass and the workload assumptions are stale. Pat wants new lithography and flows that avoid $50 million mask costs before prototyping, so that nine months compresses to a month or two.

    ▶ 9:08
  2. Memory, power modeling and thermals are the next bottlenecks after fabrication

    Pat calls HBM “a hideous memory”: poor bit density, limited shoreline bandwidth, and heat that DRAM dislikes. He says AI is a memory-compute workload, so memory has to be fixed. Power simulation is also weak: with poor 3D modeling of hotspots, designers guard-band nearly the whole power envelope and lose about 40% of their power. Better power modeling and dissipation are needed.

    ▶ 11:40
  3. The hundred AI chip startups will converge because specialization is unsustainable at scale

    Pat expects consolidation for three reasons. Workloads keep shifting (prefill, decode, reasoning models that look like CPUs, HPC-style 64-bit precision returning). Scale requires capital and workloads, so not a hundred vendors can win. And big buyers like OpenAI, Nvidia and Anthropic will pick winners. He says heterogeneity will be hidden by abstraction inside big players, citing Nvidia absorbing Groq.

    ▶ 14:42
  4. Memory innovation for the first time in 30 years is nigh upon us

    Pat has been involved with at least five new memory architectures, including Optane, and knows of close to a hundred over 30 years, yet only DRAM, SRAM and flash survive. Memory was a brutal commodity business, but the big three vendors are now among the top 20 most valuable companies, adding about $2.5 trillion in market cap. He thinks that capital plus AI demand will fund breakthroughs: stacked memory and compute, and new materials like ferroelectrics.

    ▶ 22:26

Give your email to unlock the rest of the key points — for every episode.

Do more with Charlie

Free — sign in with your email, no password.