Pat Gelsinger held 100 meetings in 100 days after stepping down as Intel CEO in late 2024. He told WIRED he wanted to narrow his options until he knew his next move.
By March, he had joined Playground Capital as a general partner. The firm invests in deep tech and new science. Gelsinger aims to help semiconductor startups restart the growth predicted by Gordon Moore decades ago. That prediction held true for a long time: the number of transistors on a chip would double roughly every two years. Now, physics limits further shrinking of atomic-scale transistors. It has become prohibitively difficult and expensive.
Gelsinger believes lithography is the solution. This involves etching chips using nanometer-scale beams of light. The leading technology comes from Dutch firm ASML. It uses light with a 13.5-nanometer wavelength. The logic is that smaller features allow for more powerful processors.
When he joined Playground, Gelsinger took a board seat at xLight. The portfolio company develops novel lithography techniques and recently received investment from the US government. During interviews, he repeatedly quoted scripture: “God said, ‘Let there be light!'”
Venture capital firms are shifting toward deep tech as artificial intelligence upends the software industry. Gelsinger argues he is better equipped to spot winners than most.
The interview
WIRED met Gelsinger in early July at the RAISE Summit in Paris. They discussed vetting deep tech founders, the US government’s stance on AI and semiconductors, and why breakthroughs in those fields come together.
This conversation has been edited for length and clarity.
Departure and decision
WIRED: Roughly four months separated your departure from Intel and arrival at Playground. What was going through your head during that period?
PAT GELSINGER: My wife said, “You’re not done yet.”
I was looking at government roles, university roles, CEO roles, private equity, venture. It was really a deductive process. I decided I didn’t want to do [any more] public earnings calls. I didn’t see myself as a politician. It came down to private equity or venture.
I want to do things that matter—if they succeed, make a difference—with people I enjoy.
Why did you decide against private equity?
Private equity writes bigger checks, but it’s not as focused on the tech. At this phase of my career, do I want to write big checks and worry about financial returns, or do I want to do cool tech?
We’re at the edge of science, proving things out. That’s always the kind of person I’ve been. I love tech.
Investment shifts
A bunch of VCs are shifting toward deep tech in response to the disruption of the software industry by AI. What’s your take on how AI is changing where the next investment opportunities appear?
The door has blown wide open.
In 2024, the semiconductor industry aimed to hit a trillion dollars by 2030. Now, we’ll hit a trillion dollars next year. I don’t need my companies to win the market to get extraordinary returns. I just need them to win a decent percentage. That’s what AI has done to deep tech venture.
The good news is that a lot of venture firms are swinging in that direction. The bad news is that, for the most part, they’ve forgotten how to do deep tech—how to pick the winners and losers.
Assessing founders
What’s your process for assessing the credibility of deep tech founders, when sometimes the physics behind their inventions hasn’t yet been proven?
Our investment team is deeply technical—they’re engineers, PhDs, professors, etc. Then we go through a rigorous tech diligence process: lots of interviews, and background checks. What’s the hard problem? Can they articulate it in depth?
We’re looking to fund the best team—not a team—on a given topic.
But to what extent are those checks even possible, when the ideas behind these startups are sometimes pressing the limit of scientific understanding?
Generally, when you see deep tech things emerge, there’s usually two or three companies gravitating to that idea. Very rarely do you find the dodo bird. Then you’re asking, ‘Am I picking the best one?’
The xLight bet
Tell me about the bets you’re making at Playground. You’ve taken a board seat at xLight.
One of the reasons I joined Playground was because of xLight. Those are the kind of companies I want to work on, because I’m deeply invested in making the semiconductor industry the future—waking Moore’s law from its nap.
If we solve light, that is the hardest problem. Can I move past 13.5-nanometer light? That’s the next breakthrough. With free-electron lasers, we could go [smaller], to 5-, 4-, 3-, 2-nanometer wavelength light.
xLight is not versus ASML, it’s with ASML. The first thing we want to do is hook our light source up and make ASML machines better. It doesn’t get better than that.
Lithography versus design
You think lithography is the key to reawakening Moore’s law, more than processor design?
Light is the most important thing. God said, “Let there be light.” We’re going to harness that as far as we can take it.
There’s a variety of things percolating in the space—new material structures, superconducting, ferroelectric materials—but all of them need lithography. It’s always been the center of semiconductors. If I wake up lithography, that’s thrilling.
It’s a pretty exhilarating period of human history. For a technologist, it doesn’t get better than this.
AI inference and memory
What do you make of the wave of semiconductor startups trying to challenge Nvidia in AI inference—and this vision of a future where processors from a multitude of vendors harmonize inside a single system?
We train models once, we use them many [times]. There will be a swing toward inference. And it’s very clear we could do a lot better.
Obviously, training has been the heartland of GPUs. But even Nvidia recognized that its GPUs were not a great fit for all inferencing. [E.g. the Groq deal.]
My job at Playground is to make AI 10,000 times better, not 10 times. That will happen on chips that don’t look like today’s GPUs.
I think that model vendors are trying to abstract themselves from the underlying hardware, too. They’re trying to make it easier to have a heterogeneous hardware structure underneath.
What about memory?
High-bandwidth memory is a problematic technology. It’s the best we have right now. But by the end of the decade you’ll start to see stacked memory architectures will become much more dominant. d-Matrix, Fractile, and Cerebras are breaking the boundaries of what memory architectures will look like.
I believe that some of these hardware innovations are going to be 10 to 100X better. That means one gigawatt produces 10 gigawatts worth of tokens. I’ll take that deal any day.
Energy constraints
You’ve previously described energy as one of the main bottlenecks to progress in AI. Where does investment need to be made in order to rectify that problem?
In the US, we’ve had low-single-digit expansion in energy capacity in the last decade. That’s despicable. In a digital AI age, energy capacity is economic capacity. I don’t want to be anti-climate, but we were so consumed with climate that we forgot about capacity.
We have to start looking at how to turn on energy expansion. But it’s really a conundrum: New gas turbines have an eight-year supply chain; nuclear takes a decade to build; solar [depends on] Chinese supply chains.
One of my companies, Alva Energy, is doing nuclear upgrading. Let’s take the fast path—harvesting more value from today’s nuclear footprint—but also let’s reignite the nuclear build.
This is an area that needs innovation, because fundamentally the winners and losers in the AI age will be those with the energy capacity to build their systems.
The final piece is to make AI a lot more efficient—the chips, power distribution. We have a number of companies working in voltage regulation and conversion. It’s about looking at the whole stack.
Regulation and leadership
When you joined Playground, you said you wanted to extend US leadership and ensure the benefits of AI are evenly distributed. Recently, there have been signs—in the form of chip export controls and interventions in the release of AI models—that the US administration is willing to wield its leadership as leverage. Against that backdrop, I wonder whether those two ambitions might become mutually exclusive?
There’s yin and yang on these types of topics. But fundamentally, I don’t view those in dissonance.
If I had the choice of the US or China having leadership in foundational models, which would I pick? The US, of course. I want the models to be based on our values, to enhance human experience, to solve many of the world’s hardest problems. This is a race we want Western nations to win.
It seems like the US administration can’t decide whether to be maximally hands-off in regulating AI or maximally hands-on, dictating which models can be widely released.
This is just moving so fast. A major foundational model is being released every four weeks. Against that backdrop, do I need to regulate? What do I need to regulate? What are the quality and security requirements for these models? These are valid questions.
We’re figuring it out in real time, because things are moving so rapidly. One of the things that gives me solace is that we’re debating it.
But what’s your stance? Should models be reviewed by an American government body before release?
Models need to have integrity of process and visibility of the testing that was done on them. I want to know what proprietary foundational models are trained on. I want vigorous benchmarking. I want to know they’re not just the first to do something, but they do it with the appropriate security requirements and values alignment.
One of two things needs to happen: Either the industry does that review, or the government has to step in to make sure it gets done.
