Pat Gelsinger, the former CEO of Intel, has shifted his focus from manufacturing chips to identifying the next generation of hard technology bets, and he's convinced that AI inference efficiency represents one of the industry's most pressing unsolved problems. After 45 years leading major semiconductor and infrastructure companies, Gelsinger joined Playground Global as a venture investor and board member, where he now evaluates roughly 10 companies he believes could reshape the computing landscape. His new vantage point offers a rare perspective on where the industry is headed, particularly regarding the computational bottlenecks that remain largely invisible to the average consumer .

What Changed When a Semiconductor Legend Became a Venture Investor?

Gelsinger's transition from running Intel to evaluating startups represents a significant shift in how he approaches problem-solving. Rather than managing product roadmaps and manufacturing timelines, he now spends his time meeting with founders, asking probing questions, and assessing which technical ideas have the depth and leadership to matter. He explained that after leaving Intel, he spent 100 days interviewing venture firms, private equity companies, CEO roles, and government positions before deciding that Playground Global aligned best with his goal of making a difference in the future of science and technology .

What makes Gelsinger's new role particularly valuable is his ability to connect decades of experience in classical computing architecture with emerging challenges in AI acceleration and quantum computing. He noted that he is learning about entirely new domains, from nuclear energy upgrades to superconducting Josephson junctions, stretching his understanding of what computing can become. His perspective is not limited to incremental improvements in existing systems; instead, he is thinking about fundamental shifts in how we approach computation itself .

Why Does Inference Efficiency Matter More Than Most People Realize?

One of Gelsinger's core arguments is that inference, the process of running trained AI models to generate predictions or responses, still needs to improve by orders of magnitude. This is a critical distinction from training, which is the computationally expensive process of teaching an AI model. While much of the industry's attention has focused on training larger and more capable models, the real bottleneck for practical AI deployment is making inference faster, cheaper, and more efficient. This efficiency gap affects everything from data center costs to the feasibility of running AI applications at scale .

Gelsinger's emphasis on inference efficiency aligns with a broader industry recognition that as AI models become more capable, the cost of running them in production becomes the limiting factor. Companies deploying large language models (LLMs), which are AI systems trained on vast amounts of text to generate human-like responses, face enormous bills for the computational resources required to serve users. Improving inference efficiency directly translates to lower operational costs and faster response times for end users .

How to Think About the Future of Computing Architecture

• Classical Computing Foundations: Traditional CPU and GPU architectures will remain essential for general-purpose computing tasks, but they are not optimized for the specific demands of AI workloads, particularly inference at scale.
• AI Acceleration Specialization: Dedicated AI accelerators, such as specialized inference chips and dataflow machines, are becoming necessary to handle the computational patterns that AI models require, moving beyond the von Neumann architecture that has dominated computing for decades.
• Quantum Computing Integration: Quantum systems will eventually enable computation of problems that are mathematically intractable with classical or AI-only approaches, creating a "trinity" of computing modalities that work together rather than in isolation.

Gelsinger describes this convergence as the "trinity of computing," a fusion of classical, AI, and quantum effects working together. He argued that almost everything that can be expressed mathematically will eventually become computable through this combination of approaches. This is not a prediction that quantum computing will replace classical systems, but rather that the most powerful solutions will leverage all three modalities in concert .

Gelsinger

"I want to do things that matter. If they succeed, it makes a difference with people I enjoy. Playground is a place I can do that," said Pat Gelsinger.

Pat Gelsinger, Venture Investor at Playground Global

What Does Gelsinger See as the Next Pressure Points in Computing?

Beyond inference efficiency, Gelsinger identifies several emerging pressure points that will define the next phase of computing infrastructure. These include precision requirements for AI models, resilience in distributed systems, optical interconnects for data center communication, and the search for architectures that move beyond traditional von Neumann designs, which separate memory from processing in ways that create bottlenecks for modern workloads .

The industry is still early in the real buildout of agentic AI workloads, which are systems that can autonomously plan and execute tasks, and scientific AI applications that tackle complex research problems. These emerging use cases will place new demands on inference systems that current architectures are not optimized to handle. Gelsinger's focus on identifying companies working on these next-generation problems reflects his belief that the biggest opportunities lie not in incremental improvements to existing systems, but in fundamental rethinking of how computation should work .

At 65 years old, Gelsinger remarked that the only disappointment he has at this phase of his career is not being 35 years younger, describing the current moment as "the greatest time to be a technologist in human history." His shift from running one of the world's largest semiconductor companies to evaluating early-stage ventures suggests that the real innovation in computing may not come from established players optimizing existing approaches, but from startups willing to challenge fundamental assumptions about how inference, acceleration, and computation should work .

Gelsinger