Skip to main content
Skip to content
Agentic Finance

Can an AI Batch Wait for Cheaper Power? A Deadline and Restart-Cost Model

Compare running now, waiting or checkpointing into a cheaper power window. Model deadline fit, restart overhead and the cost a real tariff passes through.

A violet computing rack framed by an amber clock arc and a lime power ribbon illustrates scheduling work within a limited electricity-price window.
Reading options: audio, text size & sections

Decentralised News · 2027 planning series · Reviewed 21 November 2026

A cheaper electricity window is useful only if a workload can reach it, complete useful work inside it and still satisfy its deadline. Waiting can save money. Checkpointing, retained capacity and an incomplete result can also consume the entire saving.

The practical decision is a schedule: run now, start later, or complete part of the job now and resume during another tariff window. This article compares those three choices for one fixed batch. It does not assume that a tokenized energy balance gives its holder access to a wholesale electricity price.

What Matters

Use the price the workload’s actual contract passes through. Then compare the available hours with useful runtime, checkpoint capture, restore time and the deadline. A low-rate window can be too short for the job. A paused schedule can consume more kilowatt-hours while costing less, and the lowest arithmetic cost can belong to a schedule that finishes too late.

The Signal

Google Cloud’s sustainability guidance identifies geographical and temporal shifting for flexible workloads. AWS Batch’s Spot guidance discusses shorter jobs, checkpoints and retries. These documents establish practical mechanisms and constraints. They do not establish that every AI job can move, that every lower wholesale price reaches a customer, or that a delay will be economical.

DN Alpha Thesis

Workload flexibility has value when it converts an available tariff difference into a completed, accepted result. Treat the deadline and recovery overhead as capacity constraints before treating the price difference as savings. The useful output is a feasible schedule with a cost boundary, not an abstract spread between an energy token and a GPU quote.

Which electricity price reaches the job?

The US Energy Information Administration distinguishes real-time wholesale costs from the prices most consumers pay. It explains that many consumers pay rates based on seasonal averages, while some utilities offer time-of-day pricing. A chart of hourly wholesale prices therefore cannot be inserted automatically into a customer’s savings calculation.

An owned machine on a qualifying time-of-use tariff may have a direct link between its electricity consumption and its bill. A colocation contract may pass through a defined energy charge. A rented GPU or hosted model API may bundle electricity into another unit. In the last case, moving a job to an hour with cheaper grid power need not change the customer’s quoted rate.

PurchasePrice needed for this decisionCommon mistake
Metered equipmentApplicable tariff and meter boundarySubstituting a wholesale market chart
ColocationActual pass-through and demand-charge termsAssuming every power component varies hourly
Rented computeBillable capacity, storage and transfer termsAdding electricity already included in the rate
Hosted APIThe service’s billable unit and contractTreating an estimated server kWh as the invoice

The worksheet below addresses a declared electricity-cost boundary. For a bundled service, use it only when the relevant energy component can actually be isolated. Otherwise compare the service’s complete prices using DN’s finished-compute-job cost model.

What can move without changing the result?

Start with one accepted batch: the same input, model or program version, output requirements and hardware performance. Its useful runtime is the time needed to produce that result, excluding the additional checkpoint and restart work introduced by the proposed schedule.

A background embedding job, a rendering batch or a nonurgent training stage may have scheduling flexibility. An interactive request with a short response commitment has a different deadline. Even a flexible job may depend on an upstream dataset arriving, a license being available or a downstream process starting on time.

Google’s guidance discusses flexible, noncritical workloads for temporal shifting. The operational inference is to identify flexibility at the job level. Calling a fleet “AI” or “batch” does not establish that each member can wait.

The available cheap window must also contain actual compute capacity. A tariff is not a reservation. Preserve the permitted machine, memory, software and capacity interval when comparing schedules; a hypothetical idle GPU is not an available execution slot.

Three schedules, one finite deadline

DN’s model uses hours from a common observation. Let H be useful runtime, P the declared metered power in kW, T the alternate window’s opening time, W its duration and D the completion deadline. The window has one electricity rate; time outside it has another.

Run now: execute continuously from hour zero to H. If this run overlaps the alternate window, those overlapping hours receive that window’s rate. The model does not incorrectly charge the whole run at the outside rate. Completion requires H ≤ D.

Start later: begin the whole job at T. It may extend beyond the alternate window and then incur the outside rate. Completion requires T + H ≤ D. A cheap first portion cannot make a late final result eligible.

Pause and resume: perform some useful work now, capture a checkpoint, stop, then restore at T and complete the remainder within the available window. The time available after T is the smaller of W and D − T. Restore time consumes part of that capacity.

If the window can support b hours of useful resumed work after restore, the earlier useful work is a = H − b. The earlier block plus checkpoint capture must finish by T. Both useful blocks must be positive; if the entire job fits later, the start-later schedule handles it without inventing a checkpoint.

This is a comparison of three defined schedules, not a general cluster scheduler. The paused candidate maximizes useful work inside this one alternate window. It does not search multiple windows, move work between machines or model uncertain runtimes.

A synthetic batch that can save money only by pausing

Consider an invented six-hour job with an 8 kW metered boundary. The outside electricity rate is 0.30 currency units per kWh. A four-hour window opens at hour eight with a rate of 0.10. The result is due at hour twelve.

Checkpoint capture takes 0.25 hours before the pause; restore takes another 0.25 hours inside the later window. Both use the same assumed 8 kW. The paused schedule has one additional currency unit of separately entered cost, such as a declared checkpoint-storage charge. These values are synthetic, not a utility tariff, cloud quote or measured checkpoint test.

ScheduleCompletion hourModeled active and recovery energyModeled costDeadline result
Run now6.0048.00 kWh14.40Fits
Start at hour eight14.0048.00 kWh8.00Misses hour twelve
Pause and resume12.0052.00 kWh10.20Fits mathematically, with no margin

The paused plan performs 2.25 hours of useful work now and spends 0.25 hours capturing its checkpoint. It stops at hour 2.50. At hour eight it restores for 0.25 hours and completes the remaining 3.75 useful hours. The two useful portions sum to six.

Its outside-window electricity costs 2.50 × 8 × 0.30 = 6.00. Its later four hours cost 4 × 8 × 0.10 = 3.20. Adding the separate unit gives 10.20, a modeled saving of 4.20 against running now. That saving is conditional on the stated checkpoint capability and available capacity.

The apparently cheaper 8.00 start-later row is ineligible because it finishes after the deadline. The paused plan finishes exactly at the deadline, leaving no allowance for queueing, verification or delay.

Calculate the price and overhead budget

The paused electricity cost is the early block, including checkpoint capture, multiplied by the outside rate, plus the restored block multiplied by the window rate. Add the separately declared paused-schedule cost once.

paused cost = P × (a + checkpoint hours) × outside rate + P × (b + restore hours) × window rate + extra cost

For the example, the maximum separately entered additional cost before tying the run-now schedule is 14.40 − 6.00 − 3.20 = 5.20. The entered extra cost of 1.00 leaves 4.20 of headroom. If keeping capacity or transferring checkpoints adds another 5.00, the schedule is no longer cheaper.

Holding the timing and extra cost fixed, the window rate can rise to approximately 0.23125 per kWh before the paused plan ties the original 14.40 baseline. That threshold uses the example’s run-now cost, which does not overlap the later window. If the run-now schedule also consumes electricity inside that window, its cost changes with the same rate and the comparison must be recalculated.

Bring the deadline forward to hour nine. Only one hour remains in the window, leaving 0.75 useful resumed hours after restore. The early useful portion rises to 5.25 hours. With the same assumptions, paused cost becomes 15.00, above the 14.40 run-now cost. The electricity spread did not change; usable scheduling capacity did.

Cheaper is not the same as lower-energy or lower-carbon

The original paused schedule models 52 kWh rather than 48 kWh for useful work, capture and restore. It assumes zero additional waiting consumption at the chosen boundary. It is cheaper within that assumption because more active consumption occurs at the lower rate. Reporting an energy-efficiency improvement would contradict those quantities.

A paused machine is not necessarily powered off. The worksheet does not calculate idle energy before starting or between capture and restore. Add any known incremental waiting-electricity charge to the relevant extra-cost field, without counting active electricity again. Its energy display remains an active-and-recovery subtotal; use a fuller meter model for total consumption.

For example, if retaining the paused machine adds 3 kW throughout its 5.5-hour wait, while the run-now alternative can release that load, the extra 16.5 kWh costs 4.95 at the invented outside rate. Adding that charge to the existing one-unit extra cost raises paused cost to 15.15, above 14.40. This is a counterfactual assumption, not a measured idle-power claim.

Carbon is another calculation. The electricity rate alone does not establish the emissions associated with either interval. A carbon-aware scheduling objective needs the relevant intensity data, timing and boundary; a financial objective needs the applicable bill. Neither substitutes automatically for the other.

DN’s energy-per-accepted-task guide addresses measurement denominators. Its renewable-electricity claim record addresses attributes, ownership and retirement. Those decisions remain separate from choosing when a particular job runs.

An energy certificate or token also cannot establish the schedule’s software compatibility, available GPU capacity or completed output. Keep the purchased right and the workload’s physical and billing arrangements explicit before using the word arbitrage.

Demand charges and reserved capacity can dominate

The US Department of Energy’s demand-flexibility guidance explains demand charges tied to a customer’s highest demand during a billing cycle. Moving a job into an inexpensive energy interval can still increase a facility’s chargeable peak if it overlaps other equipment.

The relevant comparison is the incremental billing consequence of the proposed schedule. Allocating the entire monthly demand charge to one job would usually answer a different question. Establish the tariff’s interval, existing load and proposed overlapping demand before entering any resulting incremental charge.

The worksheet keeps power constant and does not reconstruct a building’s demand profile. Put a known additional monetary consequence in the schedule’s extra-cost field; leave an unknown consequence unresolved in the wider decision. A zero means no extra cost is being assumed, not that demand charges have been investigated.

For another explicitly synthetic sensitivity, suppose a documented billing counterfactual attributes a one-kW increase in chargeable facility demand to the paused schedule, at an invented eight currency units per kW. That adds eight units to the schedule, lifting its 10.20 cost to 18.20. The original 4.20 electricity saving no longer covers the incremental bill. This example is a billing assumption, not a published utility rate.

Compute reservations can create another cost while a job waits. If pausing does not release the billable resource, the electricity saving may coexist with several hours of unused capacity charges. The cheaper-energy result then describes only one component of the actual purchase.

A retry is not proof of a usable checkpoint

AWS Batch’s retry documentation describes rerunning failed attempts under a configured policy. That mechanism does not by itself establish that an application resumes from its last useful state. A repeated job might start from the beginning, repeat an external side effect or discover that its checkpoint is unavailable.

A paused schedule needs application-level evidence: the captured state must survive the pause, be readable by the intended replacement process and restore the correct task position. The comparison also needs the time spent capturing and restoring it. An empty checkpoint file created quickly is not a successful recovery.

A controlled qualification run should compare the restored job’s accepted output with an uninterrupted reference and account for repeated work. For stochastic training, the acceptance rule need not require identical bytes, but its permitted differences must be defined before looking at the result. DN has not performed such a workload test for this article.

Planned pauses also differ from provider interruption. EC2’s interruption notices are best effort; stop and termination normally have a two-minute notice, while hibernation begins immediately. Do not treat that notice as a guaranteed checkpoint budget. The worksheet’s deterministic plan does not simulate an unannounced interruption.

Moving across regions is a separate extension

Time shifting on one compatible machine avoids some of the complications of relocating a job. A geographical move can change hardware performance, capacity availability, software permissions, data-handling requirements and the route back to the output consumer.

Google Cloud’s network-pricing documentation distinguishes transfer categories and locations. Data movement can therefore be a chargeable part of a migration, even when the destination’s compute offer looks cheaper. This article uses no quoted transfer rate because the applicable product and route are not specified.

Record both transfer time and cost. A cost that looks tolerable can still consume the deadline margin. Conversely, keeping a reusable copy near the alternate machine may change the economics of later batches. That is a repeated-workload placement decision, beyond this one-job, one-machine model.

Use the correct power boundary

The model’s P is a declared average power for the chosen metered boundary during useful work and overhead. A rack-level reading, a whole-machine reading and a GPU-board reading are different quantities.

NVIDIA’s management documentation describes board-power readings and device-specific support. Such a reading should not silently become a facility bill. Include supporting equipment when the tariff applies to that broader meter, and avoid adding a facility multiplier twice if the input already includes it.

If checkpointing uses a different power profile, this equal-power model needs refinement. Its purpose is to expose timing and cost trade-offs, not to manufacture a meter reading from a hardware specification.

Use the DN Batch Window and Deadline Model

Enter one job, one power boundary, an outside rate, one alternate window and the deadline. Add capture, restore and schedule-specific costs. The worksheet displays deadline eligibility separately from cost and energy. Its checkpoint field begins unresolved; a mathematically attractive paused plan remains unqualified until that assumption is supported.

DN Batch Window and Deadline Model

Change the illustrative assumptions below. This worksheet runs locally in your browser. It records your declarations; it does not verify evidence, connect to accounts, submit instructions or grant approval.

Paused schedule has lower modeled cost; checkpoint capability is unresolved

Run-now cost, currency units14.40
Run-now completion, hours from now6.00 — fits
Start-later cost, currency units8.00 — late, ineligible
Start-later completion, hours from now14.00 — misses deadline
Conditional paused cost, currency units10.20
Paused completion, hours from now12.00
Run-now / paused active and recovery energy, kWh48.00 / 52.00
Early / resumed useful work, hours2.25 / 3.75
Conditional paused saving versus run now4.20
Paused extra-cost ceiling versus run now5.20
  • Three deterministic schedules for one batch, one machine and one alternate tariff window. This does not verify capacity, a tariff, checkpoint behavior or the legal right to any energy price.
  • Run-now and start-later active electricity costs include their actual overlap with the alternate window. Additional schedule costs are separate from that electricity arithmetic.
  • The kWh display covers useful work and capture/restore only. Waiting consumption is not calculated. Include any known incremental idle-electricity charge in the relevant extra-cost field; do not add active electricity twice. A complete energy total needs a fuller meter model.
  • A deadline-fitting result is not a guaranteed completion. Queueing, output delivery, random interruptions, variable runtime, demand-charge reconstruction and multi-region transfer timing are not modeled.
  • Start later misses the deadline; its displayed cost must not be chosen merely because it is lower.
  • Paused candidate: 2.25 useful hours now, 0.25 capture hours, stop at hour 2.50, restore at hour 8.00 for 0.25 hours, then 3.75 useful hours.
  • Checkpoint and restore support remains unresolved. The paused arithmetic is conditional and is excluded from declared eligible schedules.
  • The paused candidate finishes exactly at the deadline. There is no margin for delay or validation.
  • Common costs that do not change across schedules cancel from this comparison. Include all actual incremental costs before treating it as a full financial decision.

The initial numbers are synthetic. Unknown evidence remains unresolved. Downloads contain the values you enter; use aliases and avoid private or identifying information.

Methodology and limitations

Primary documentation was reviewed on 21 November 2026. DN compared this scope with its existing completed-job cost, energy-measurement, electricity-claim and scarce-power articles. This contribution is a finite schedule model and synthetic worked example. No utility account, GPU provider, electricity market or paid workload was accessed or tested.

The model assumes fixed useful runtime and active power, one machine, one alternate tariff window, continuous capacity where declared and deterministic capture/restore durations. Waiting consumption is excluded from its kWh subtotal; any incremental waiting charge must be supplied separately. It excludes uncertain queues, stochastic interruptions, model-quality changes, multi-region transfer timing, demand-charge reconstruction and revenue from selling electricity or tokens. Declared extra costs must cover applicable omissions before a full financial decision can be made.

Use absolute, timezone-aware timestamps when translating the relative hours into an operating plan, especially across tariff changes or daylight-saving transitions. Retain time for output validation and delivery in the actual deadline. A schedule finishing exactly at D is mathematically feasible and operationally without margin.

Sources and evidence register

Primary sources reviewed on 21 November 2026. The article distinguishes documented facts, DN analysis and synthetic examples. Recheck changeable rules and configurations before acting.

  1. Google Cloud: Optimize resource usage for sustainability — Temporal and geographical shifting for flexible workloads; not a workload-specific savings guarantee.
  2. AWS Batch: EC2 Spot best practices — Job size, checkpointing and retry considerations; no universal checkpoint budget inferred.
  3. US EIA: Electricity prices and factors — Wholesale costs versus retail rates and time-of-day pricing; US explanatory scope.
  4. US Department of Energy: Federal demand flexibility — Demand charges and peak-load consequences; actual tariff applicability remains contract-specific.
  5. AWS Batch: Automated job retries — Retry attempts and failure handling; retries are distinct from application-state recovery.
  6. AWS EC2: Spot interruption notices — Best-effort notices and the hibernation exception; no guaranteed advance warning.
  7. Google Cloud: Network pricing — Data-transfer categories and location-specific charging; no live rate quoted in the example.
  8. NVIDIA: System Management Interface — GPU-board power-reading scope and device support; not a whole-facility meter.

Frequently asked questions

Does a lower wholesale electricity price reduce my GPU bill?

Only if the applicable contract or tariff passes that change through. A bundled GPU or API price may not vary with the grid price at all.

Can a schedule use more energy but cost less?

Yes. The synthetic paused schedule models 52 kWh of useful work and recovery instead of 48 kWh, while moving enough consumption into the cheaper interval to reduce cost. Waiting consumption is excluded unless its charge is supplied separately; this is not an energy-efficiency claim.

Why is the cheapest start-later example rejected?

It completes at hour 14, after the hour 12 deadline. Its 8.00 cost is arithmetic for a late schedule, not an eligible way to complete the required task.

Does enabling automated retries establish checkpoint support?

No. A retry may restart the application. Capture, persistence, restore behavior, repeated work and output acceptance need application-level evidence.

Does the worksheet include demand charges or retained GPU reservations?

It does not reconstruct those billing rules. Enter known additional schedule costs separately and investigate unresolved costs before treating the energy comparison as a complete bill.

Is the result an energy-arbitrage recommendation or a guaranteed deadline?

No. It compares declared deterministic schedules for one workload. It does not trade energy, value a token, guarantee available capacity or predict provider interruptions.

DN research note: This article is original editorial analysis supported by the linked sources. Calculations and example cases are illustrative unless explicitly described otherwise. The worksheet is a planning aid; it does not carry out the actions it describes.

Get the most talked about stories directly in your inbox

Join the Decentralised News briefing for independent crypto, DeFi and AI analysis. No spam, unsubscribe anytime.