AI is disrupting services. What does that mean for power demand?

The ground he argues from is a bet his firm made about three years ago, around 2023 and shortly after GPT-3.5, that open-source models would become a dominant way AI gets delivered, because enterprises would want something cheaper, more private and more customizable than a frontier model call. What followed was Together AI, which sells inference and training for open-source models and is therefore a heavy consumer of compute and data centers. He names Baseten as another player in the space he likes, without saying he has invested. What interests him is what the Together investment turned into. It began as software, optimizing training and inference, became infrastructure, and then infrastructure in the physical sense. He treats that blurring as general rather than a quirk of one company, and it is also why he thinks his read on deployment is grounded: he can see what traffic actually runs through both businesses.

What he sees is that almost nothing has happened yet. He borrows Demis Hassabis’s phrase about the foothills of the singularity, declines the larger claim by saying he is not smart enough to judge whether the singularity is near, and keeps the narrower one: we are in the foothills of enterprise deployment, particularly for larger and less technology-forward companies. He splits the capital three ways. The vast majority of AI dollars went into developing models; a shift toward the infrastructure to run them has come over the past 12 to 18 months; and the third bucket, deploying the technology across the country, is paltry relative to where it needs to go. The error he attributes to his own industry is treating capability as deployment, imagining a single binary threshold after which everything changes at the speed of light. The real world does not work that way, and it will take a lot longer to deploy: 80% of the American economy is services-driven, so until a way is found to get AI into services businesses, he says, this is going to take a while. That is what makes today’s compute, in his phrase, infinitesimally low relative to the ultimate need, and why this year’s hyperscaler capital spending could look small in a few years. Demand for tokens is the part he is confident about; how much physical infrastructure is needed to deliver them as models get more efficient is the part nobody can answer.

The bulk of the conversation is the thesis underneath that deployment claim. An AI-native service, a term he says he made up and concedes is not well understood, is any service deliverable with AI faster, better and/or cheaper than the incumbent, with, for many of them, a highly paid and highly valued human on top to check the work and put their name to it. The obvious way to build one is to buy existing firms for their distribution and add AI, which he says ignores how hard behavior change is: roll-ups are difficult anyway, and this one requires convincing the people you did not fire to work differently inside a new conglomerate. Building from scratch is easier, though still exceedingly difficult, because you can build the general ledger for one use case, build agents against it, run it yourself with a few customers, then hire senior practitioners to sit on top and train it. His example is Hanover Park, a fund administrator for private equity that built its ledger and its agents from scratch and now serves clients faster and at higher quality than the legacy providers. The reversal is the part that matters. For the best versions of an AI-native service, he says, you generally are not demand constrained, because a faster, better or cheaper version of an existing service sells itself; the question is only whether you can deliver it primarily with AI. The failure mode he names is mirage product market fit: you sell plenty, customers are satisfied, and you got there by hiring people, which makes you an ordinary services business that took the wrong form of capital. The best operators, Hanover Park included, periodically stop selling to let the platform catch up to 80 or 90% of the work. He is blunt that this is harder than software, because you have to build Stripe and McKinsey at once and make them fit together.

On energy his worry is political rather than physical, and this is where his position and the host’s sit side by side unreconciled. AI has become genuinely unpopular for reasons all over the map; electricity pricing is one, and that one he calls imminently addressable: make the grid more flexible, make it cheaper to build, and in most cases data centers should lower energy prices rather than raise them. He adds that the industry has explained AI’s benefits poorly, beyond a vague promise to cure cancer someday. Johnson then puts the structural case, that demand from air conditioning, vehicles and data centers outruns supply so prices rise regardless, with AI the obvious villain. Saper’s reply is a hedged concession, that this is probably true and the pinch point would probably have arrived anyway; he does not withdraw his own claim, and neither man puts the two together. He then offers a future he himself labels Pollyanna and too rosy, in which the buildout deliberately supports the communities hosting it, noting that a calm, non-screamy policy position is a centrist one and centrism is out of fashion. Robotics runs the same way: he is a bull, having invested in Bedrock Robotics, whose self-driving construction equipment does a lot of data center work and, when that works at scale, should let data centers be built faster and more cheaply, which he hopes ultimately translates into cheaper power. On labor he insists on being granular rather than ham-fisted, since excavator operators are already in shortage and the immigration crackdown deepened it, so many robots deploy before any current worker is displaced. His overall position is that the best outcome is a medium pace of deployment, too slow and the country is outrun by China and others, too fast and the unemployment shock arrives before anyone can retrain. The current pace feels medium if not slow, and what he fears is that not enough people are planning for the tails.