The Clay Jar and the Data Center

Jensen Huang describes Nvidia's business in a single phrase: electrons in, tokens out.¹ The phrase is doing more work than it appears to. It implies that the production of intelligence is, at its core, the conversion of electrical energy into useful computation, and that the rate of that conversion is what matters: the tokens per joule, the joules per dollar, the dollars per token.

I think this framing is correct. I also think it points to a problem that most of the current AI infrastructure buildout is structured to avoid rather than solve.

The dominant response to the AI energy problem, judged by where capex is flowing, is to treat energy primarily as a procurement problem. Sign longer PPAs.² Restart nuclear.³ Buy your own gas turbines.⁴ Microsoft and Three Mile Island,⁵ Amazon and Talen,⁶ Meta and Clinton,⁷ and the surge of new gas turbine orders backing up data center campuses across Texas and the Mountain West are each rational responses to real constraints such as scarce firm capacity, long lead times, and the strategic need to lock in resources before competitors do. But they share an implicit assumption, that AI workloads are continuous, massive, and inflexible, and therefore demand continuous, massive, inflexible supply.

I think the procurement view is correct as far as it goes, and incomplete in a way that will matter a great deal in the next few years. The reason is that the demand side of an AI data center is not flat, and treating it as flat leaves most of the available economic value on the table.

Orchestrating energy end to end, from how it is procured and stored through how it is allocated to workloads in real time, will be one of the most strategic layers of the inference era. This holds regardless of the generation mix. Renewables-plus-storage, nuclear-powered, and gas-fired campuses all face the same underlying problem: a demand side that varies through the day, and a supply side whose value depends on the moment. Orchestration doesn't go away when the generation mix is firm.

Why inference is not flat

There is an assumption embedded in the procurement view that deserves to be made explicit: that AI workloads are continuous and inflexible, and therefore the right job is to procure continuous and inflexible supply. The first half of this is becoming less true every year.⁸

A training run is a single coherent computation, fragile and tightly synchronized, that cannot be paused without significant cost.⁹ Inference is not. Inference is millions of independent requests with wildly varying latency requirements. Some are real-time chat, where every additional second of latency degrades the product, but some are batch document processing that can wait six hours without anyone noticing. Some are agent workflows where the user has already gone to lunch and will not check the result until a few hours later. Some are scheduled analyses that could run any time before the morning report. In aggregate, inference has substantial temporal flexibility, and the proportion of inference that is genuinely latency-bound is shrinking as agentic and long-running workloads grow.

The market is already pricing this. Both OpenAI and Anthropic offer a 50% discount on their batch APIs, which run with a roughly 24-hour turnaround.¹⁰ That is the going rate for temporal flexibility in inference today, and the fact that the two largest model providers converged on the same number is not an accident. It is the market revealing that a meaningful share of inference does not need to run immediately. Agentic workloads are accelerating the shift: Gartner projects that 40% of enterprise applications will integrate task-specific AI agents by the end of 2026, up from less than 5% a year earlier,¹¹ and inference workloads, broadly, are projected to rise from roughly a third of total AI compute in 2023 to two-thirds by 2026.¹² The category that is growing fastest is the category with the most temporal flexibility.

This flexibility is, today, almost entirely unexploited at the energy layer. The systems that schedule inference jobs don't see the price of electricity. The systems that manage electricity don't see the value of the next token. The two control loops run in separate buildings, optimized for separate objectives. A facility that could see both at once would, on most days, find itself with hours of cheaper electrons and hours of more valuable workloads, and the gap between them is pure margin that nobody is currently capturing.

The clay jar

Ten thousand years ago, food was not scarce. Hunter-gatherers were often well-fed. The constraint was time, not production. Everything had to be consumed almost immediately, because there was no reliable way to carry today's abundance into tomorrow. Without storage, there was no surplus. And without surplus, there was no progress.

The invention that broke through that ceiling was not farming. It was the clay jar. The sealed grain pit. The granary. These technologies decoupled eating from finding, so that a good harvest could carry a community through a bad season. People could settle. Individuals could stop farming and become potters, soldiers, priests, traders. The jar was the unlock, but the jar alone wasn't enough. What mattered just as much was the knowledge of when to open it. The granary needed a granary-keeper.

The AI energy problem has the same structure. Solar and wind are now genuinely cheap. The global weighted average levelized cost of utility-scale solar fell roughly 90% between 2010 and 2024, from $0.417/kWh to $0.043/kWh.¹³ Nuclear, where it can be brought online, provides something different and also valuable: firm power at scale, indifferent to weather and time of day. Natural gas, the actual workhorse of most new AI campuses in the United States, sits between them, dispatchable and relatively cheap but exposed to fuel price volatility and emissions constraints. These are not competing solutions. They are different inputs to the same machine.

What none of them gives you, on its own, is the matching of supply to a demand curve that is itself highly variable. In 2024, ERCOT curtailed more than 8 TWh of wind and solar, energy that existed, was generated, and was thrown away because nothing could use it in that moment.¹⁴ A nuclear-powered data center, meanwhile, runs its reactors flat through hours when global inference load is at its trough and the marginal token is worth a fraction of what it would be worth six hours later. A gas-fired campus has a different problem: the turbines cannot ramp fast enough to follow GPU demand swings, and forcing them to do so shortens their life and increases emissions, because the equipment is being used in a way it was not designed for.¹⁵ These are three forms of the same waste: generation and consumption that are not coordinated through time.

Storage is already infrastructure

A modern AI data center already runs at least six distinct storage technologies simultaneously, across timescales spanning ten orders of magnitude. Ceramic capacitors on the GPU board absorb microsecond voltage spikes. Rack-level supercapacitors handle millisecond fluctuations. Nvidia's Vera Rubin NVL72 ships with 400 joules of storage per GPU, six times the prior generation, cutting peak current demand by up to 25%.¹⁶ UPS systems bridge seconds-long outages. Facility-level battery energy storage systems buffer the volatile power signatures of GPU clusters from the stability requirements of the grid. Storage is no longer insurance. It is infrastructure, embedded in the machine itself.

What does not yet exist, at any data center I am aware of, is the layer that coordinates across all of it. The capacitors are governed by the GPU firmware. The UPS is run by the facility team. The grid-scale battery, where one exists, is dispatched by an energy trading desk optimizing for day-ahead prices. The workload scheduler is owned by the inference team, which has no visibility into any of the above. Each control loop is locally sensible and globally incoherent.

The interesting object is the system that closes this gap, that can see generation, storage, grid prices, thermal limits, workload demand, and revenue per token at the same time, and make second-by-second decisions about all of them. Storage is already part of the stack. What is missing is the coordination on top of it. This is a software problem, not a hardware problem.

The orchestration layer forecasts energy prices and grid conditions, tracks the state of charge of every battery, the temperature of every rack, and the priority of every queued workload, and decides in real time which inference job runs where, which battery discharges, and which token gets produced now versus which one gets queued. The pattern has analogues. High-frequency trading firms built it for markets, cloud providers built it for compute, grid operators built it for reliability and stability. Nobody has built it for the intersection of all three.

When a passing cloud cuts solar output across West Texas, an orchestration layer would not just issue a dashboard alert. It would shift non-urgent agent workloads to a region with cheaper power, draw down on-site batteries to buffer the local GPUs, and preserve the most economical electrons for latency-sensitive inference. The point is not backup power or conventional demand response. It is the joint optimization of energy and computation as a single economic system.

ERCOT as the leading indicator

Texas is the most instructive case because the dynamics are visible there earlier than anywhere else. The state has more data center pipeline than any other, the most aggressive battery buildout in the country, the largest queue of new gas turbines, and a deregulated market structure that surfaces the tension between them in real time.¹⁷

ERCOT entered 2026 with 13.9 GW of installed battery storage, nearly doubling from 7.8 GW at the start of 2025, and Modo Energy projects 40 to 55 GW operational by 2029.¹⁸ The data center pipeline looks larger on paper. ERCOT is processing roughly 226 GW of large-load interconnection requests, 72.9% from data centers, up 270% in 2025 alone.¹⁹ Only 5.3 GW of large loads were actually energized as of November 2025.²⁰ A more conservative range for real buildout is 40 to 65 GW by 2030.²¹

The conventional reading is that storage is dramatically undersupplied. It isn't, because batteries don't need to power data centers around the clock. They need to firm an underlying generation mix that is already, on most hours of most days, more than capable of meeting demand. ERCOT's nameplate capacity sits well above its all-time peak demand of 85.5 GW (set in August 2023) and its summer 2025 peak of 83.9 GW.²² The state is not short of electrons. It is short of electrons in the shape a data center needs them: smooth, instantaneous, and reliable.

The hardware is not the bottleneck. The bottleneck is that batteries and data centers are nominally in the same market and operationally in different worlds. What Texas makes visible, every other grid will eventually surface in its own way.

Looking ahead

By 2028, I think at least one major AI lab will be publicly reporting that a meaningful share of its inference cost is being managed by a combination of storage orchestration and workload routing to regions and times of day where firmed power is cheapest. By 2030, I think the gap in $/token between the best and worst data centers will depend as much on energy orchestration as on the underlying silicon. And I think the unified control plane across procurement, storage, dispatch, and workload will become a recognized layer of the AI infrastructure stack. The granary-keeper.

I could be wrong in specific ways. If inference demand turns out to be much more latency-sensitive in aggregate than I expect, the temporal flexibility I am leaning on shrinks, and the procurement view holds up longer. If grid-scale storage costs stall, whether from a chemistry floor, supply chain shocks, or policy changes, the economics of coordinated systems become less compelling against simply buying more firm supply.

What I am more confident about is the underlying shape. Energy is not solved by procurement alone, regardless of whether the supply is nuclear, gas, or pure renewables. The demand side of an AI data center is variable, the supply side has options, and the value of a joule depends on the moment it is used. A system that doesn't coordinate across these is leaving money on the table every second it operates. By 2030, the gap between the facilities that coordinate well and the ones that don't will be a major source of variance in $/token.

I would extend Huang's framework to include the middle term. Electrons in, orchestration, tokens out. The middle term is not just a scheduling problem. It is the control layer that decides, second by second, when an electron is worth more stored, when it is worth more as grid revenue, and when it is worth most as intelligence.

I think the ceiling on what AI can do for the world in the next decade will be set less by how good the models get and more by how many tokens we can afford to spend on the problems worth solving. Cheaper inference is not just a margin improvement. It is the difference between intelligence that gets rationed and intelligence that gets applied. Every joule we coordinate well is a token that gets to do useful work somewhere it could not before.