During a recent episode of the a16z podcast, technology investor Gavin Baker sat down with partner David George for an in-depth conversation. Baker revealed that he spent the summer touring the industry, asking one single question to everyone he met: "Can you show me even one quantitative data point in your business that is getting worse? Just one." The result, he said, was that he couldn't find a single one.

While the prevailing concern is an overexpansion of AI infrastructure, Baker argues that market estimates for AI demand are profoundly underestimated. The risk that deserves real attention today, he suggests, is underbuilding. Currently, the number of heavy AI paying users worldwide may be under 10 million, contrasting with 1.5 billion knowledge workers globally. Demand diffusion is just beginning, and supply is already severely constrained.

Fundamentals Are Accelerating While Stock Prices Decline

Baker has noticed a clear divergence: the fundamental trajectory of the AI industry accelerated through July and August, yet AI-related public equities suffered significant drawdowns during that same period. "Overall, AI accelerated in July and accelerated again in August," Baker said. "What's strange is that public stocks have fallen all over the place over the last two months."

He specifically highlighted several companies' momentum: OpenAI is clearly accelerating, open-source models are accelerating even faster, and Grok has seen a "quite dramatic acceleration" after launching Grokbot. He noted that Anthropic is in a quiet period ahead of its potential IPO, so its growth data is temporarily opaque, but the rest of the sector "is all accelerating."

Baker used a vivid metaphor: "You know the saying, a river with an average depth of only two feet can still drown you. There hasn't been much movement at the index level, but some AI stocks have already seen fairly substantial pullbacks—while the fundamentals are broadly accelerating."

Fewer Than 10 Million Heavy Users; Diffusion Is Just Starting

Baker believes the tens of billions of dollars in AI revenue currently being generated is supported by an extremely niche user base. "These companies currently have around $80 billion in revenue. How many heavy-paying users are behind that? I'd guess no more than 30 million, maybe even less," David George said. Baker's estimate was more conservative: "Perhaps it's actually under 10 million."

He cited internal data from his firm, Atreide: from March to August, internal token consumption grew 100-fold. With just two people using the Grokbot enterprise version, token consumption is projected to increase another 10 to 20 times within a month. "There are 1.5 billion knowledge workers globally. On the demand side, we seem to be just getting started, and we're already severely constrained by supply," Baker said.

He also observed that AI-native companies now spend over 10% of their labor compensation costs on tokens each month, while the best-performing traditional enterprises have reached 1%. "When I look at the supply-demand characteristics, the supply-side question is 'is this sustainable,' but combined with the demand side, I think it's quite clear."

The Real Risk: Underbuilding

This is where Baker diverges most sharply from the mainstream market narrative. "Everyone is worried about oversupply; I'm more worried about severe undersupply," Baker said. "If that's the case, what you might see isn't a decline in AI access costs, but a significant price increase."

David George further pointed out that this supply shortage could persist until 2028, and political resistance could further delay planned construction projects. Baker cited a "sounds absurd" view from Marc Andreessen—that token costs could rise tenfold. "It sounds absurd, but we do live in a supply-demand-driven world. If demand expands dramatically and supply can't keep up, the entire premise changes."

What worries him even more are the social consequences of supply scarcity. "That could ironically lead to real compute inequality—large companies and wealthy individuals can afford compute," George said. "And those 'data center degrowth advocates' will complain about this two years from now, without realizing it's precisely what they caused." Baker responded directly: "Exactly because they won't let us build data centers."

Compute Investment Payback Under One Year; Economic Logic Unusually Strong

Baker believes the current economic logic of compute investment is extraordinarily rare in his investing career. He cited data disclosed by Nebius and CoreWeave: bringing 1 gigawatt of compute online costs about $50 billion. Customers can prepay 50% to 60%, or $25–30 billion, with the remainder recovering faster through the spot market. The overall payback period is approximately 9 to 10 months.

"In my investing career, it's rare to see opportunities where companies can deploy tens of billions of dollars and achieve a payback period of under a year," Baker said. "This is quite exceptional."

He also pointed out that Nvidia GPUs can be financed at very attractive rates, with institutions like Blackstone, KKR, and Apollo participating at low costs. "As models improve and useful lifespans keep extending, monetization per gigawatt keeps rising—the true equity payback period could be well under a year."

Baker used Microsoft as a cautionary example: after Satya Nadella committed $80 billion in capital expenditures at Davos last year, the company slowed down, "and now they regret it." In contrast, OpenAI chose an aggressive investment path, "and the high returns clearly prove that was the right decision, both short-term and long-term."