In a recent installment of the a16z podcast, technology investor Gavin Baker engaged in a deep-dive conversation with partner David George, challenging the prevailing anxiety surrounding AI infrastructure buildout. During his summer travels across the industry, Baker posed a singular question to everyone he met: "Can you identify even a single deteriorating quantitative data point in your business? Just one." His findings were striking — he could not uncover a single one. While the market obsesses over overcapacity, Baker contends that the actual risk lies in a severe underestimation of AI demand and, consequently, underbuilding. With global heavy AI paying users potentially under 10 million against a backdrop of 1.5 billion knowledge workers worldwide, demand diffusion appears to be in its infancy, while supply is already critically constrained.

Accelerating Fundamentals Amid Declining Stock Prices

Baker observed a clear divergence: AI industry fundamentals continued to accelerate through July and August, even as public AI-related stocks experienced significant pullbacks. "Overall, AI accelerated in July and it accelerated again in August," Baker noted, puzzled by the market's sharp declines during this period of fundamental strength. He highlighted specific momentum at OpenAI, even faster acceleration in the open-source segment, and a "rather dramatic acceleration" for Grok following its Grokbot launch. He suggested Anthropic's growth figures are temporarily opaque due to its pre-IPO quiet period, but described the rest of the sector as "all accelerating." Drawing a vivid analogy, he remarked, "You know the saying, a river with an average depth of only two feet can still drown you. At the index level, there hasn't been much movement, but some AI stocks have seen substantial drawdowns — while fundamentals have broadly accelerated."

Fewer Than 10 Million Heavy Users: Diffusion Is Just Starting

Baker argues that the billions in AI revenue currently being generated is supported by an extremely niche user base. George speculated that the roughly $80 billion in annual revenue for leading AI companies might be backed by perhaps 30 million paying users, or possibly fewer. Baker's estimate is even more conservative: "Maybe it's actually under 10 million." Citing internal data from his firm Atreide, he noted that internal Token consumption grew 100-fold from March to August. With just two people using the Grokbot enterprise version, Token consumption is projected to surge another 10 to 20 times within a month. "There are 1.5 billion knowledge workers globally, and it feels like we are just at the starting line on the demand side, yet we are already severely constrained by supply," Baker stated. He also noted that AI-native companies now allocate over 10% of their payroll costs to monthly Token spending, with traditional companies excelling in this area reaching the 1% mark. "When I look at the supply-demand dynamics, the supply-side question is 'is this sustainable,' and when you factor in the demand side, I believe it's quite clear."

The Real Risk Is Underbuilding

This is where Baker diverges most sharply from mainstream market narratives. "Everyone worries about oversupply, but I'm more concerned about a severe supply shortage," he said. "If that's the case, what you might witness is not falling AI access costs, but significantly higher prices." David George elaborated that this supply crunch could persist until 2028, further delayed by political resistance to planned projects. Baker referenced Marc Andreessen's seemingly "absurd" notion that Token costs could rise tenfold. "It sounds outlandish, but we do live in a supply-and-demand-driven world. If demand expands dramatically and supply can't keep up, the entire premise changes." He expressed deeper concern about the societal consequences of shortages. George painted a stark picture: "That could ironically lead to genuine compute inequality — where large corporations and the wealthy can afford it, while the 'data center degrowth advocates' will complain about this two years later, without realizing they caused it." Baker responded directly: "Precisely because they won't let us build data centers."

Sub-Annual Payback Periods: Rare Economic Logic

Baker believes the current economic logic of compute investment is exceptionally rare in his career. Citing data from Nebius and CoreWeave, he outlined that deploying one gigawatt of compute costs approximately $50 billion. Customers can prepay 50% to 60% of that amount, or $25 to $30 billion, with the remainder recouped faster in the spot market. This yields an overall payback period of roughly 9 to 10 months. "In my investing career, it's rare to see companies deploy tens of billions of dollars and achieve a payback period of less than a year," Baker said. "That's quite exceptional." He also highlighted that Nvidia GPUs can be financed at very favorable rates, with firms like Blackstone, KKR, and Apollo participating at low costs. "As these models get better and their useful lives extend, and monetization per gigawatt keeps rising, the actual equity payback period is likely well under a year." Baker cited Microsoft as a cautionary tale: after Satya Nadella's $80 billion capex promise at Davos last year was scaled back, "they now regret it." In contrast, OpenAI's aggressive investment strategy is, in Baker's view, "clearly vindicated by high returns, both short and long term."