The price of compute, the price of power

In this article

The next decade may give us much cheaper computation without giving us much cheaper power where and when computation needs it.

That is the dual thesis. One price falls as chips, models, and software improve. The other is tied to land, wires, generation, financing, and the hour of the day. The two prices meet inside a data center, but they do not obey the same curve. I think the gap between them will become one of the more interesting markets of the 2030s.

There is already reason to take the collision seriously. The International Energy Agency's 2025 base case has global data-center electricity use roughly doubling to 945 TWh by 2030, while its cases vary with adoption, efficiency, and energy bottlenecks. That is a scenario, not a booked load. In the United States, Lawrence Berkeley National Laboratory estimated 176 TWh of data-center use in 2023 and a wide 325–580 TWh range for 2028.1IEA, Energy and AI: Energy demand from AI (2025); Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report.

Two prices, two units#

When I say the price of compute, I mean the cost of completing a specified useful task at a specified quality, latency, and reliability. A GPU-hour is an input; a completed inference or training run is closer to what a buyer wants. Stanford's 2025 AI Index found that the price of inference at one fixed benchmark threshold fell more than 280-fold from late 2022 to late 2024. That is striking evidence of improvement on that measure, not a rate I would compound for another ten years.2Stanford HAI, AI Index 2025: State of AI in 10 Charts, inference cost at a fixed MMLU threshold.

The price of power needs a different definition: the all-in cost of delivered, usable electricity at a particular site and time. The cost of building a solar plant is not that price. IRENA reports low 2024 global average costs for new onshore wind and solar generation, but those levelized generation costs do not include every wire, storage hour, backup arrangement, or queue delay a data center faces.3IRENA, Renewable Power Generation Costs in 2024 (2025). Its generation LCOE is distinct from a site's delivered tariff.

A commodity is something buyers can specify well enough to compare suppliers. Compute moves in that direction when workloads are portable and service terms are legible. It does not mean all accelerators are interchangeable. Memory, interconnect, data gravity, security, and latency still matter. The claim is narrower and more useful: an increasing share of compute can be described, priced, scheduled, and bought as capacity rather than as a relationship with a particular machine.

There is a second distinction hidden in the word price. The engineering cost of a useful task can fall while the rental price of a scarce cluster rises. The first reflects how many chips, joules, and operator hours the task consumes. The second clears a market between buyers who may value the same hour very differently. A routine batch inference and a time-sensitive frontier training run can therefore move in opposite directions without either price being a mistake. The table below models the former's cost, not a universal spot price for GPUs.

A ten-year model, with the assumptions exposed#

To make the intuition testable, set both 2026 prices to an index of 100. These are separate indices, not dollars and not a conversion between compute and electricity. Assume a representative useful compute task costs 85 units in hardware, capital, operations, and margin, plus 15 units in electricity in 2026. Let the nonpower component fall 25% per year as performance and utilization improve. Let electricity used per task fall 10% per year, while delivered power's unit price rises 2% per year in the representative constrained market. None of those rates is a measured forecast; they are assumptions for the thought experiment.

Power index: P(t) = 100 × 1.02^t. Compute index: C(t) = 85 × 0.75^t + 15 × (0.90 × 1.02)^t, where t is years after 2026.

Rounded values. The modeled compute task is held at constant quality and service terms. The 2026 cost split and annual rates are illustrative. A different workload, utilization, or electricity contract gives a different path.

The number I care about is less the 11 than the changing composition. Efficiency makes each task require less electricity, yet electricity becomes a larger share of the remaining cost because the other components fall faster. The direction of delivered power is also uncertain. Hold every other assumption fixed and vary only its annual price change:

These are mechanical sensitivity cases, not probability-weighted forecasts. The annual power changes refer to a hypothetical delivered-price index, not to the generation cost reported by IRENA.

The model is sensitive to the power path once hardware gets cheap, but it is still more sensitive to how fast the nonpower component actually improves.

Now add quantity. Suppose completed tasks grow 20% a year because lower prices and new uses expand demand, while energy per task falls 10% a year. Total energy for those tasks grows by 1.20 × 0.90 = 1.08 each year, or about 2.16 times over a decade. This is an illustrative demand response, not a data-center electricity forecast. It shows why efficiency and rising total load can coexist. If task growth were below about 11.1% a year under the same efficiency assumption, total energy for those tasks would instead fall. That threshold is a useful thing to measure.

There is no law saying lower cost is passed through immediately. When demand for firm, low-latency capacity exceeds what is available at the required place and hour, the clearing price includes a scarcity premium. When capacity is idle, a supplier may price near marginal operating cost to recover something on a depreciating asset. A decade can contain both situations, sometimes on the same day.

The price of a particular hour#

The annual index hides the part a trader would actually need to know. Power prices are distributions across hours and locations. A data center's electricity bill depends on its own load shape against that distribution, not on the arithmetic average of a region's prices. A flat, uninterruptible load buys expensive hours along with cheap ones. A movable batch job can select hours. A serving cluster with a hard latency promise often cannot.

Consider a simplified 100 MW facility with a batch workload that can pause for four expensive hours and finish later without missing a deadline. If the all-in price of usable power is $200/MWh during those four hours and $50/MWh when it resumes, shifting 400 MWh saves $60,000 before the costs of stopping, restarting, moving data, or reserving spare capacity. If the work cannot be made up, there is no $60,000 saving: the operator has traded away the output too. This is why flexibility belongs in a contract rather than in a slogan.

For an incremental job, the operating choice can be written plainly: run now when the value of the completed work exceeds the nonpower cost plus the power needed multiplied by the delivered power price. If the job can wait, compare that net value with running it later, including the chance that later capacity will be unavailable. The price at which a site declines an optional new job is usually lower than the price at which it shuts down a job already promised to a customer. Turning off a live cluster may violate service terms, strand synchronized work, or impose restart costs. The two thresholds should not be conflated.

We can already see a crude deadline market in compute services. OpenAI's Batch API offers a lower price in exchange for a turnaround window of up to 24 hours. That discount is evidence that timing has value to both buyer and seller; it is not proof that the provider schedules every request around wholesale power prices. The physical and contractual bridge between the two markets is the thing to build.4OpenAI, Batch API documentation, including its price and completion window.

The map is part of the model#

Texas makes the distinction between a cheap input and a usable megawatt concrete. The EIA found that natural gas at the Waha hub traded below zero on 42% of trading days in 2024 because production outran pipeline takeaway. A negative gas price at a hub does not mean a data center can buy round-the-clock electricity for less than zero. It still needs generation, a connection, transmission access, and a way through the expensive hours. ERCOT publishes prices that vary by location and time, so a Texas site has a real exposure to congestion as well as energy. The opportunity is strongest for jobs that can move; the risk is strongest for a buyer that values uninterrupted service but budgets at a cheap-hour average.5EIA, Waha gas prices in 2024; ERCOT, market prices and data.

Texas also has a denominator problem. A request to connect is an option on a possible project, not a meter reading. ERCOT's large-load process now batches applications and studies the grid upgrades needed to serve them; its planning documents explicitly adjust proposed loads for the fraction likely to appear. Counting every requested megawatt as inevitable demand would overstate the constraint, while counting none would miss the projects that do get built. For the next decade, the useful series is the progression from requested load to approved load to energized load to measured hourly use. Each conversion rate changes the forward view of both power and compute capacity.6ERCOT, Batch Zero large-load process (2026) and adjusted load forecasting (2025).

Alberta tests whether flexible demand can be matched with new supply quickly enough. Its system operator said in 2025 that 29 proposed projects sought more than 16 GW of connections while an interim process could connect up to 1.2 GW of large loads through 2028. The gap describes applications and a connection policy, not a forecast of what will be built. In 2026 AESO also explored a process for large loads that bring their own generation. This makes the central trade sharper: a developer can wait for firm grid service, finance generation alongside a site, or accept limits on when it can draw power. Each path changes capital cost and the value of the compute contract it can honestly sell.7AESO, interim large-load approach (2025) and bring-your-own-generation proposal (2026).

Québec shows why abundant low-carbon generation does not automatically become a cheap, unlimited tariff for every new buyer. Hydro-Québec proposed a separate rate for large data centers in 2026, subject to regulatory approval, with the stated aim of reflecting the value of its electricity and protecting other customers from the cost of fast-growing demand. A model that plugs the historical hydro generation cost into a data-center P&L misses the decision about who receives the next block of power and on what terms. Allocation policy can be as material as fuel price.8Hydro-Québec, proposed data-center tariff (February 2026).

Northern Virginia and PJM provide a fourth comparison: a mature data-center cluster is valuable for network access, customers, and existing infrastructure even if power becomes harder to secure. PJM's 2026/27 capacity auction cleared at its approved price cap, and PJM attributed part of the load-forecast increase to data centers. That capacity price pays for reliability over a future delivery year; it is not the same thing as a spot electricity price. A buyer choosing between Virginia and a cheaper-power location is therefore comparing latency, network effects, capacity charges, and time to energize, not simply two $/MWh figures.9PJM, 2026/27 capacity auction results (July 2025).

The map can change. A new line, a pipeline, a tariff decision, or a generation project can erase yesterday's advantage. Location is a state variable in the model, not a permanent ranking of states and provinces.

What gets traded#

Electricity already has spot markets, forwards, and congestion instruments. Spare cloud capacity already has a price that varies by instance type and location; AWS Spot is a familiar example, with interruption risk built into the bargain. Compute derivatives are beginning to take shape too: CME and Silicon Data announced futures tied to H100 and B200 rental indices in 2026, subject to regulatory review. A chip-rental index is a meaningful benchmark, but it does not deliver a completed job or hedge every constraint in a buyer's workload.10Amazon EC2, Spot documentation; ERCOT, Congestion Revenue Rights; CME Group, compute futures announcement (August 2026).

I expect the market to develop in layers. The first is procurement: a buyer reserves a specified cluster and pays for availability, whether or not every chip is used. The second is scheduling: a buyer grants the operator a window to complete work, and receives a discount for that flexibility. The third is financial hedging: a buyer or supplier fixes some exposure to a published rental or electricity index. The fourth, and hardest, is a contract for output: a task completed to a quality standard by a deadline. Each layer settles a different risk. Treating them as one generic “compute price” would make the market look more liquid than it is.

Consider a seller that promises firm capacity in Texas for a year. A forward on a regional power index can reduce exposure to broad energy-price moves. It cannot guarantee the site's connection, eliminate the difference between the index and its own node, or replace a failed transformer. A GPU rental-index hedge can offset a move in market rental rates. It cannot guarantee that a particular model fits memory, that a large training run has a synchronized cluster, or that a customer's data may cross a border. These residual differences are basis risk. They are where careful underwriting, and eventually better contracts, matter.

The most interesting optionality may belong to the workload. A firm job pays for the right to run through a bad power hour. A flexible job sells the operator the right to choose its hour. A portable job adds a choice of grid, but pays in latency, transfer cost, and legal complexity. The economic value of each option is the avoided cost in the hours when it can actually be exercised, less the cost of preserving that flexibility. A cheap-hour average cannot tell you its value. You need a distribution of power prices, a distribution of compute demand, and a realistic model of which jobs can move.

That is also why market structure will develop unevenly. H100 rental hours can be indexed sooner than a guaranteed answer from a changing model. Batch inference can move more easily than a tightly synchronized frontier training run. A data center can curtail some optional workloads without becoming a fully dispatchable power plant. My bet is that procurement and scheduling become routine, while standardized output contracts arrive only where the job, service terms, and failure remedy can be stated precisely.

What would change my mind#

This thesis weakens if electricity per useful task falls faster than demand for tasks grows, if firm generation and transmission arrive quickly enough to tame the expensive hours, or if workloads remain too differentiated to compare across suppliers. It also weakens if customers will not pay enough for the data centers now proposed. A ten-year index cannot settle any of those questions.

I would watch a few observable series instead of defending the model's point estimate: task-level prices at fixed quality; GPU rental rates by chip and region; electricity per completed task; requested versus energized megawatts; the distribution of hourly delivered power prices; and the spread between firm and interruptible compute contracts. If compute costs keep falling but the firm premium disappears across regions and hours, the trading thesis has lost much of its force. If the premium widens while queue conversions stay low, physical access has become the scarce product.

Neel Somani's Power 2026 and Rahil Mittal's The Price of Now helped frame the power-market side of this question. What I want to keep testing is the joint market: how a workload's deadline and mobility turn a megawatt-hour into a compute contract, and when the cost saved by moving it exceeds the value of having it done immediately.

I think compute deserves to be treated as an economic commodity in the making: something to meter by useful output, reserve in advance, relocate when possible, and price against its physical inputs. The hardware story is only half of it. Over the next decade, the valuable judgment may be knowing which computation must happen here and now, and which can wait for a cheaper hour on a different grid. That is where the price of compute meets the price of power. It is also where trading begins.