Pat Gelsinger isn't done. The former Intel CEO, who left the company in late 2024 after a turbulent turnaround attempt, has landed at Playground Global as a venture capitalist hunting for "hard technology" and "hard physics" investments. In a 2026 interview with Dr. Ian Cutress, Gelsinger says he's working with about 10 portfolio companies, sitting on most of their boards, and calls this phase of his career about doing "things that matter" with people he enjoys. For example, Mario Zechner recently joined Earendil to launch coding agent pi.
What caught my attention is Gelsinger's claim that AI inference efficiency needs to improve by 10,000x. He frames the future as what he calls the "trinity of computing": classical, AI, and quantum systems working together. And he's putting money behind that thesis. One of Playground's portfolio companies, Groq, uses a spatial dataflow architecture that ditches traditional cache hierarchies entirely. Instead of the von Neumann bottleneck that plagues most chips, Groq's Language Processing Unit streams data through functional units with compiler-controlled static routing. The company says it's hitting over 500 tokens per second on large language models.
Gelsinger is also learning nuclear energy through Alva Energy, superconducting and Josephson junctions for quantum computing, and bioengineering applications of semiconductors. The quantum stuff, he says, required turning his brain "to a 90 degree angle" after 45 years as "a digital boolean guy." When he jokes that his only disappointment is not being 35 years younger, you can tell he means it. Someone doesn't sign up to learn about Josephson junctions at this stage unless they're genuinely excited.
Gelsinger's 10,000x efficiency claim sounds wild. But Groq's early numbers show the dataflow approach is a credible path. Spatial architectures that eliminate the von Neumann bottleneck aren't theoretical. They're shipping now, and they're fast. He's betting on the right problem: AI inference costs will dominate, and the current chip paradigm won't solve it.