a16z Podcast: Pat Gelsinger on Why Hardware’s Golden Age Is Just Starting

SPEAKER_1 (0:01)

When AI races ahead, the real bottleneck is often not intelligence at all, but power, packaging, and time.

Today we're diving into a fascinating conversation with Pat Gelsinger, former Intel CEO and now a general partner at Playground Global, who believes hardware may be entering a new golden age. Though he has a pretty stark warning about what could limit that golden age. He puts it bluntly: "In an AI digital age, energy capacity equals economic capacity." His argument is that data centers, GPUs, entire projects will hit a wall if the power simply isn't there. Right, and he makes this really interesting point about scale. He says, "Nothing is a chip anymore; it is a rack." Modern hardware has to be designed as part of a much larger system now.

That's such a shift from his early days. Gelsinger actually describes winning a scholarship as a teenager, joining Intel young, and moving from technician to architect on the 286, 386, and 486 He calls that era magical because students and industry were learning from each other in real time. And that collaboration helped launch the modern EDA industry. But fast forward to today, AI is beginning to reshape chip design itself. He points to tools like Jalapeno as examples of first-principles AI in design. The catch though?

The physical timeline still dominates everything. A chip may take three months to design, then nine months to fabricate, plus months more for packaging and integration. You can't AI your way around physics. Speaking of breakthroughs, what did he say about memory? That's been stagnant for decades. About 30 years actually. But Gelsinger thinks memory innovation may finally be near, with new materials and structures under discussion. He's skeptical of some approaches, but believes investment and market conditions are finally strong enough to support real breakthroughs. He's pretty direct about AI accelerator startups too. There are more than 100 right now, but he doubts they'll all survive. His reasoning? Workloads change too fast, capital is too expensive, and the biggest AI companies will likely pull hardware and software into tighter co-evolution. The market will consolidate. Then there's the next layer of physical limits: stacking, optical I/O, and power delivery. He says three to five layers may be the manufacturing sweet spot. Copper is running out of runway, and optical systems will become critical around 2028 or 2029

And he says, "Power is cool again," as the industry rebuilds how electricity is delivered and converted inside data centers.

Looking forward, Gelsinger believes virtual machines are coming back in a new form, but this time built for agents, not just humans.

That means new security, new management, and what he calls "vMotion for agents." The takeaway is clear: the next hardware era will be shaped by energy, physics, and a very different kind of software stack.