The last time @GavinSBaker and I sat down to record a podcast, it was almost exactly a year ago and we started with the immediate question: is AI a bubble? Gavin's answer focused on utilization and returns (and perhaps unsurprisingly, his answer was: No). Unlike the dark fiber of 2000, there were, and still are, no idle GPUs. The largest buyers of compute were funding the buildout from some of the strongest balance sheets in the world, and their AI investment was already producing real returns. I sat down with Gavin again last week, and to chart just exactly where we are in the cycle. He has spent the summer asking operators for one quantitative measure in their business that is getting worse and has yet to find one. At the same time, public AI stocks have gone through meaningful drawdowns. Still, if you look at the way people are using AI today, there’s a good chance that demand diffusion has barely begun. AI revenue rests on fewer than 10m heavy users, against roughly 1.5b knowledge workers. At the most AI-native startups, token spend is approaching or exceeding 10% of human compensation, which suggests that there’s a long way to go before even the earliest adopters fully integrate agentic capabilities. Within Atreides, Gavin said token consumption rose 100x from March through August, and even further once the team started using products like GrokBot. As Gavin put it, once people begin approving automations, token consumption starts to feel “sort of endless.” This dynamic would be reason enough alone to believe that we’re massively undersupplied at the moment. But there are other reasons on the supply side too, namely that there are a near-unbounded number of potential winners in the space: - Frontier labs can keep winning because on the highest-value tasks, marginal improvements in intelligence are worth far more than marginal differences in price. - Open models illustrate that most work doesn’t require frontier performance, creating a much larger market for intelligence that is cheaper and customizable. - Nvidia, hyperscalers, neoclouds, and inference providers can simultaneously win because every additional token still requires physical compute, even if models become more efficient. - Enterprises can win as proprietary data, workflows, and institutional knowledge become more valuable. - Application companies can win by capturing services budgets and owning the customer relationship. This doesn’t suggest everyone will be successful, but it does mean the market itself is positive-sum and we will look back on zero-sum thinking as far too limiting. Better models create better products, better products create more users, more users create more token demand, and more demand supports continued investment in models and infrastructure. Check out the whole conversation below: @a16z