Eric Schmidt: Human Expertise Still Matters in the Age of AI

If you want to make money in AI right now, Eric Schmidt has a blunt answer: don’t invest, build. Found an agentic AI company. Not one that designs agents, but one that builds an agent to do something specific. This is the agentic period, and for the next year or two, everyone will be building agents. Build a better one than anyone else and you can own an incredible company.

A Davos Conversation on AI’s Real Economy

At Imagination In Action’s AI Summit in Davos, former Google CEO Eric Schmidt and Deloitte Global CEO Joe Ucuzoglu sat down with billionaire David Rubenstein for a wide-ranging conversation about where AI stands today. What emerged is a portrait of an industry at an inflection point: the technology is undeniably powerful, the infrastructure demands are staggering, and the gap between Silicon Valley’s capabilities and enterprise reality remains stubbornly wide.

Schmidt came armed with data and strong opinions; Ucuzoglu brought the enterprise perspective of running a 500,000-person professional services firm. Their views complement each other: Schmidt sees the raw power and competitive landscape, while Ucuzoglu lives in the messy reality of deploying AI inside large, bureaucratic organizations.

80 Gigawatts and the Infrastructure Race

Schmidt, who recently started a data center company, laid out numbers that put the AI infrastructure buildout in perspective. AI is currently contributing more than 1% to US GDP, largely because of data center construction. The hyperscalers (Google, Microsoft, Amazon) each need 1, 5, even 10 gigawatts of capacity. The best study Schmidt has seen indicates the industry needs 80 gigawatts in the next 3 to 5 years.

For context: a single nuclear power plant produces about 1.5 gigawatts. So the industry is asking for the equivalent of roughly 53 new nuclear plants’ worth of power capacity in under five years. The economics of AI right now are being most felt in this infrastructure buildout, not yet in the downstream applications.

China: Six to Twelve Months Behind

Schmidt’s assessment of the US-China AI competition has shifted notably. Historically, there was a view that China was well behind. That’s no longer true.

Three Chinese models, Kimi, Qwen, and DeepSeek, now regularly appear at the top of benchmark rankings. Americans complain about “distillation,” claiming Chinese teams used American models to train theirs. Schmidt’s response: that may or may not be true, but the important thing is the Chinese models are really good.

More significantly, China has decided to be completely self-sufficient in AI technology. They’re building the entire hardware and software stack: their own chips (the SN930), their own software layers. And their strategic approach differs fundamentally from the US: China isn’t going for AGI. They’re going for AI used everywhere in society, with a massive national program focused especially on edge deployment.

“They are in my view six to 12 months behind, which is not much.”

The Agentic Moment

Schmidt framed the current AI landscape as a transition from “language to language” (think ChatGPT-style interactions) to “language to action” through agents. When pushed on what the next big technology wave will be after AI, he pushed back: AI isn’t done. The wave is just getting started.

“We’re going from language to language, which is think of it as ChatGPT, to language to action through agents.”

He pointed to reasoning models (OpenAI’s R3, DeepSeek V3) as evidence. These systems perform graduate-level MIT/Stanford reasoning at your fingertips. The investments in data centers, new algorithms, and much larger models are entirely focused on pushing this further.

His concrete advice for making money: don’t try to catch the Magnificent 7 (too late for those gains) or get into hyper-valued startups. Instead, build an agent that does something better than anyone else’s agent. It’s the most direct path to creating value right now.

The Enterprise Gap

Ucuzoglu offered a grounded counterpoint to the technology hype. Nobody doubts the power of AI or its applicability across every sector and business function. The issue is actually bringing it to life inside large, complex, established organizations.

“It’s really hard to actually bring it to life in a large complex established bureaucratic enterprise.”

He used the baseball analogy: we’re in the top of the first inning. And history tells us there’s a lot of stickiness in the system. The diffusion from the Silicon Valley lab to the actual factory and office where people work takes time and effort.

Schmidt agreed, calling this the “technology overhang,” where the industry keeps producing things faster than organizations can absorb them. Where they diverged was on the pace: Schmidt sees concrete evidence of job loss in customer service (now highly automated) and entry-level software programming, while Ucuzoglu maintained Deloitte’s workforce grew even as AI adoption expanded.

Domain Expertise as the Key Differentiator

Both speakers converged on a crucial point: domain expertise isn’t being replaced by AI; it’s becoming more valuable. Ucuzoglu described the failure mode he’s seen repeatedly: a piece of technology developed in a tech laboratory gets force-deployed into an enterprise. It generally fails. You need people who understand the business function and the industry, paired with technologists on cross-functional teams.

The AI output problem Ucuzoglu identified is particularly sharp: the technology produces results that look impressive on the surface. To an average recipient, the output appears very convincing. But to someone who’s actually an expert in the area, some of it is gibberish. It takes the human with domain knowledge to pick between substance and garbage.

This insight led to a practical recommendation for education: Schmidt argued that the most important new course MIT and Stanford don’t yet offer is a freshman class on prompt engineering. Every student’s entire career will involve working with AI assistants. The same applies to any 40-year-old professional trying to stay relevant.

Ucuzoglu expanded this: he doesn’t want only STEM graduates. He wants “best available athletes,” people who are deep thinkers, who can understand societal implications, and who can adapt AI usage in an ethical fashion. The generalist still has a big role to play.

AI in Military and Society

Schmidt drew on his experience advising the Pentagon to trace AI’s military evolution. It started with Project Maven about 10 years ago, which was essentially using AI for surveillance: watching drone video feeds that soldiers previously monitored manually. The next phase is autonomous drones, accelerated by lessons from Ukraine. Secretary Hegseth has announced major programs around drone adoption across land, sea, and air.

On the broader societal level, Ucuzoglu flagged polling data showing a stark divide: people who actually use AI are far more positive about its potential impact than those who don’t. He worries about the “powder keg” moment, where populist concerns about the future collide with fears of job disruption. Their number one job as leaders, he said, is to go out and demonstrate how AI improves the quality of human life.

A Moment That Will Be in History for Thousands of Years

Schmidt closed with a perspective that transcends the investment advice and enterprise strategy. We are living through a moment that will be remembered for thousands of years: the arrival of non-human intelligence as a competitor to us.

“It is the moment in history when a non-human intelligence arrived and it was a competitor to us. How we use that, how we shape it, how we guarantee alignment will determine the outcome.”

His final distinction was characteristically precise: the best name for this technology is “artificial intelligence” because it is intelligence and it is artificial. It doesn’t have human values. It doesn’t think the same way. It doesn’t think at all. It’s a different kind of intelligence.

“When we work with it, we will be more powerful. When it runs us, we will be very unhappy.”

Some Thoughts

This conversation works because it pairs two genuinely different perspectives. Schmidt brings the builder-investor lens: raw numbers, competitive dynamics, concrete advice. Ucuzoglu brings the deployer lens: the messy reality of making AI work inside organizations where change is hard.

A few threads worth pulling:

• The 80 GW figure is staggering. Roughly 53 nuclear plants’ worth of new power capacity in under five years. This isn’t a prediction about AI’s potential; it’s a description of committed capital expenditure happening right now.
• Schmidt’s China assessment is the most useful framing available. Not “China is winning” or “China is far behind,” but a specific, verifiable claim: 6-12 months, with a fundamentally different strategic approach (AI everywhere vs. AGI).
• The domain expertise argument has teeth. It’s easy to dismiss “humans still matter” as comforting rhetoric, but Ucuzoglu’s specific failure mode (impressive-looking AI output that’s actually garbage to domain experts) is something anyone using these tools has encountered.
• Schmidt’s investment advice is unusually actionable. Not “invest in AI” but specifically: found an agentic AI company, one that builds agents to do specific tasks. The window is the next year or two.
• The education gap is real. Both speakers agree that banning AI in schools is the worst possible response, yet neither has a clean solution for how to teach with it. The prompt engineering course idea is a start, but the deeper challenge of assessing genuine learning remains unsolved.