Former Intel CEO Pat Gelsinger: AI Hardware Renaissance Underway, Optical Interconnects to Hit Inflection Point by 2028-2029

In an a16z podcast released on October 9, Pat Gelsinger, former CEO of Intel and general partner at Playground Global, said AI hardware is entering a renaissance. He argued that while AI has made chip design easier, manufacturing, memory, and energy bottlenecks now present the biggest constraints.

Gelsinger emphasized that in the AI digital era, energy capacity equals economic capacity. He warned that U.S. energy capacity has been nearly stagnant for 15 years, growing only about 4% annually, which mismatches the demand of million-GPU clusters. He predicted more data center project defaults due to insufficient power. He proposed bringing 800V DC back to data centers and using vertical GaN technology for single-step voltage conversion, along with liquid cooling upgrades.

On memory, Gelsinger called HBM a "hideous memory" but the best currently available, citing poor bit density, bandwidth limits, power and thermal issues. He is optimistic about new memory innovations based on ferroelectric materials, which could enable non-capacitive, high-density, stackable memory, and said he has invested in a stealth startup in this area. For 3D packaging, he believes 2 to 4 layers of memory stacking is the sweet spot due to yield challenges.

He also predicted that optical interconnects will reach an inflection point in 2028-2029, with Nvidia planning NPO/CPO adoption around that time. He supports shifting network architectures toward optical circuit switching (OCS), as AI workloads feature large-scale, predictable traffic that resembles circuit switching rather than packet switching.