The 'time to power' frame is missing AI's real fault line
Benedict Evans argues the industry's focus on GPU and power waitlists skips the real question: whether foundation models become commodity infrastructure or hold pricing power.
Benedict Evans argues the industry's focus on GPU and power waitlists skips the real question: whether foundation models become commodity infrastructure or hold pricing power.
For most of 2026, AI discourse has fixated on "time to power": how many gigawatts the build-outs are wiring up, how many GPUs the labs are buying, how long the waitlist for inference capacity will last. Benedict Evans argues the framework answers the wrong question.
In a new essay, Ways to think about token pricing, Evans frames the supply crunch currently reshaping AI economics not as an infrastructure story but as a value-capture question. The crunch is the mechanism, but the answer it will eventually produce is whether foundation model providers end up with durable strategic leverage, or whether AI settles into the economics of steel, electricity, or cloud compute: low-margin commodity infrastructure where pricing follows capacity and competitors are interchangeable.
Three facts anchor the essay. More than $1 trillion of data center capex is now in the pipeline, with another layer of semiconductor capex behind it. Inference efficiency is improving rapidly, even as new models vary widely in how many tokens they need to produce a given answer: some far more efficient than their predecessors, some far less. And most pointedly: "the current crunch is concentrated in coding workloads rather than spread across use cases."
That last line is the load-bearing one. The H1 2026 demand spike is not a general-purpose AI story. It is a software-development story, driven, in Evans' framing, by sudden product-market fit in a "relatively small field." Capacity has been constrained since 2022, but the present crunch is narrower than the discourse implies. AI has not eaten the economy in 2026. It has eaten some developer hours, and the rest of the economy has noticed because those developer hours route through the same APIs everyone else is trying to access.
If the demand side is narrow, the long-run equilibrium looks different than it would under broad-spectrum adoption. Narrow demand means the supply build-out, once it lands, may overshoot more dramatically. It also means the model providers most exposed to coding workloads carry the most pricing pressure, in both directions: they get the spike, but they also absorb the deflation risk when the build-out catches up.
Evans does not predict AI costs in the essay. He names the structural fault line that the dominant discourse sidesteps. "Time to power" can tell you when the crunch eases. It cannot tell you whether the providers coming out the other side will have pricing power, strategic leverage, or both.
The analysts who do engage that question fall into two camps. One side argues foundation models retain pricing power through differentiation, distribution moats, vertical integration, or enterprise lock-in. The other leans toward the commodity outcome: a future in which models are interchangeable, margins are thin, and competitive advantage lives in the application layer or the infrastructure layer, not in the model itself. Evans' visible-dynamics read leans toward the commodity outcome, though he stops short of declaring it.
Nobody has settled this yet, because the variables are still in motion. Supply is building. Demand is concentrated. Efficiency is rising unevenly. Model design choices are still bouncing; every training run is an experiment in token economics as much as in capability. A new equilibrium will shake out over the next few years, and the providers' balance sheets over that window will reveal which side of the fault line they ended up on.
The watch item is not the waitlist for H100s. It is whether model providers can hold token prices flat while their build-out converts from capital expenditure to depreciation, or whether the deflation that arrives with the next hundred gigawatts of supply turns the foundation-model business into a business that runs on volume and tolerates thin margins. The next twelve months of inference pricing data on workloads that are not coding will start to separate the two outcomes.