Pat Gelsinger on AI's energy crisis, the 'siliconomy,' and benchmarking AI for human flourishing

Summary

Pat Gelsinger — former Intel CEO, now deep tech investor at Playground and head of technology at faith-tech startup Glue — is building what may be the first rigorous, values-based benchmarking framework for AI foundation models, and argues that energy, not chips or capital, will be the binding constraint on AI's future.

Benchmarking AI for Human Flourishing

Gelsinger's team at Glue is applying a landmark academic study — a five-year, Harvard, Baylor, and Gallup research effort spanning 22 nations and 50-plus cultural groups — to create a new class of AI benchmarks measuring whether models produce outcomes aligned with human flourishing. The framework assesses model outputs across dimensions including relationships, character development, finances, and spirituality. The current corpus stands at roughly 1,200 to 1,300 questions, updated on a rolling basis every few weeks, with results published at glue.com.

Existing AI benchmarks measure performance metrics — time to first token, cost per token, total throughput — but none assess whether a response is genuinely good. Gelsinger's framing is direct: the absence of bad does not demonstrate the presence of good. His benchmarks attempt to close that gap quantitatively.

Early findings are pointed. Models score well on finance-related queries and poorly on faith. Gelsinger attributes this to training data composition — a reflection of who builds these systems and what content dominates the internet. Glue plans to train its own foundational models specifically to avoid embedding problematic content from the start, arguing that values-aligned outputs require values-aligned inputs.

Major labs — OpenAI, Google (Gemini), Anthropic (Claude), DeepSeek — are being tracked and benchmarked. Gelsinger says scores have improved modestly since engagement began, though he acknowledges the framework is still early-stage. The next development priority is multi-turn conversation testing, which he considers more revealing of underlying model weights than single-shot prompts. He specifically cites the need to detect and respond to suicidal ideation across extended conversations as a concrete safety gap the current benchmark does not yet adequately address.

Energy as the Hard Ceiling on AI Capex

On AI infrastructure, Gelsinger's position is unambiguous: the sector is not chip-limited, not land-limited, and not capital-limited. It is energy-limited. He frames the scale of proposed data center buildouts in visceral terms — a 1-gigawatt data center equals roughly one nuclear reactor in power draw. The US has built approximately one to three nuclear reactors in the past 25 years, while China currently has roughly 60 under construction.

He argues the US lost a decade of energy capacity investment by focusing on renewables that were not yet economically viable at scale, and that AI-driven capex is now forcing a necessary correction. Even if some capacity forecasts prove overstated, Gelsinger views the investment wave as a net positive — excess energy supply benefits electrification broadly and puts downward pressure on consumer power prices.

His deep tech fund, Playground, is now leading an investment in the nuclear space directly. He is also invested in Snow Cap, a portfolio company targeting 100x improvement in AI power efficiency through superconducting technology — reducing a gigawatt-scale data center footprint to roughly 100 megawatts of cryogenic capacity while delivering 10x more inferencing throughput. He views this class of efficiency gain as industry-reshaping, with transformational impact concentrated toward the end of the decade.

Cycle Awareness and CEO Discipline

Gelsinger draws explicit parallels between current AI capex enthusiasm and the optical infrastructure buildout of the early internet era — a stunning run that nonetheless failed to extrapolate past a decade. He expects the next two to three years to remain relatively stable, with significant economic and capacity dislocations arriving by the end of the decade as inferencing costs, performance, and power consumption undergo structural shifts.

His advice to founders is characteristically blunt: a CEO's job is to dampen both the highs and the lows, not amplify them. He distinguishes between CEOs and cheerleaders, warning that too many companies are currently led by the latter. The practical implication is capital discipline — raise cash when conditions are favorable, because access disappears precisely when it is needed most.