Embedded Files
Bits-to-Watts: Connecting Markets and Prices for Compute and Power
What economic incentives are required for data center flexibility?
What economic incentives are required for data center flexibility?
Many electricity systems around the world are observing unprecedented growth in load interconnection requests driven by AI data center development, raising the question of data center flexibility potential. Flexible data center operation could allow data centers to connect while mitigating impact on existing consumers.
We calculate the Value of Compute Load (VoCL), the short-run marginal benefit of electricity consumption for data centers.
We calculate the Value of Compute Load (VoCL), the short-run marginal benefit of electricity consumption for data centers.
Using best-in-class utilization benchmarks, and instance pricing from hyperscale cloud providers, we calculate ranges of VoCL for over 75 computing configurations across on-demand, spot, and forward-reserved reliability service categories.
View the full paper. For more information, contact farhad.billimoria@oxfordenergy.org and conleigh.byers@fticonsulting.com.
View the full paper. For more information, contact farhad.billimoria@oxfordenergy.org and conleigh.byers@fticonsulting.com.
There is a large range of VoCL across service agreements and compute resource specifications.
There is a large range of VoCL across service agreements and compute resource specifications.
Low VoCLs allow the possibility of price-responsive demand during extreme periods, in which demand would voluntarily shift in time or geography. High VoCLs can be higher than administratively-determined consumer Value of Lost Load (VoLL), motivating interest in ex-ante flexibility and priority service agreements to connect new data center load while protecting existing consumers.
Estimated range of willingness-to-pay for compute load by service class.
VoCL is at the core of efficient investment, operational, and contracting decisions — including non-firm interconnection and priority service agreements.
The ideal regulatory framework balances the economic opportunity of data centers with protection of existing consumers.
The ideal regulatory framework balances the economic opportunity of data centers with protection of existing consumers.
Non-firm access agreements in which demand is cleared subject to a security-constrained economic dispatch (SCED) is consistent with open-access principles to the transmission system. Where VoCL is relatively low, harnessing natural flexibility via bid-in demand is possible, and incentivizes the optimal mix of grid sue, spatio-temporal shifting, and self-supply. Where VoCL is relatively high, combined with the time value of money, data centers are incentivized to enter into ex-ante agreements for flexibility with the system operator and priority service agreements in which they are curtailed before mass-market consumers. While natural flexibility is incentivized by the price, ex-ante agreements require careful consideration of compliance and enforcement in real-time.
Price pressure on consumers can be alleviated while preserving optimal investment signals by requiring data centers to offer a hedge of electricity prices. This could be, e.g., a monthly average rate call option that prevents consumer electricity prices from rising higher than a cap over the month. Where the VoCL is high and the energy price is high, this results in a transfer of some data center surplus back to consumers. It also incentizes data centers to sign forward contracts for new generation at the correct locations in the grid to mitigage congestion and price spikes.
Below we describe the methodology and results in more detail.
Cloud computing pricing is relatively transparent, with publicly posted prices available from providers and pricing aggregators.
Cloud computing pricing is relatively transparent, with publicly posted prices available from providers and pricing aggregators.
Under Infrastructure-as-a-Service (IaaS), cloud providers supply raw compute infrastructure such as virtual machines, GPUs, storage, and networking resources.
Under Infrastructure-as-a-Service (IaaS), cloud providers supply raw compute infrastructure such as virtual machines, GPUs, storage, and networking resources.
The VoCL of a compute instance can be calculated by uniting cloud computing pricing and power draw for a compute instance.
The VoCL of a compute instance can be calculated by uniting cloud computing pricing and power draw for a compute instance.
MCEi denotes Marginal Compute Exposure of instance i, measured in $/hour. Marginal Compute Exposure reflects the incremental economic value associated with the compute instance service
MPDi denotes Marginal Power Draw of instance i, measured in MW. Marginal Power Draw reflects the incremental electrical load required to deliver that same unit of compute instance service, holding infrastructure and configuration fixed.
This framework considers VOCL from a bottom-up perspective focusing on the value associated with the underlying compute infrastructure. The focus is on marginality and short-run incentives in electricity markets, so we do not incorporate fixed costs nor a required profit margin. An analogy is a renewable generator where the projects may require an electricity price of, e.g., $50/MWh over the long-run to pay fixed and operating costs and meet hurdle returns, but in short-term spot markets it will bid at approximately $0/MWh in line with its near-zero short-run marginal cost.
Compute load may span a broad range of willingness-to-pay for electric service.
Compute load may span a broad range of willingness-to-pay for electric service.
Panel of box-plot distributions of the implied On-demand and Spot VoCL for training (trng) and inference (inf) median utilization for AWS EC2 GPU instances. Cases: on-demand with zero cumulative downtime (st = 0) (top row), and spot (bottom row). Monthly utilization of 10% is assumed.
This has important implications for short-run dispatch, scarcity pricing, demand participation, and the long-run coordination of generation, transmission, and large-load interconnection investment. Service-level agreements (SLAs), in which a portion of the consumers bill is repaid after a cumulative downtime threshold is reached, strongly influence VoCL.
VoCL can spike dramatically when threshold downtimes are reached in SLAs.
VoCL can spike dramatically when threshold downtimes are reached in SLAs.
Illustrative Cloud Service Availability SLAs and Service Credits (as at 7 May 2026) * Reflects On-Demand Capacity Reservations for Azure Virtual Machines; Azure maintains multiple SLA classes across products. ** EC2 also provides a region-level SLA with a more stringent Band 1 threshold of 0.01% downtime. *** GCP also provides more stringent downtime thresholds (e.g. 0.01%) for premium-tier services.
Sensitivity of implied on-demand VoCL (including SLA liabilities) for a range of canonical on-demand AWS cloud instances to cumulative downtime and monthly instance utilization for power draw based on a training median utilization. The left-hand side panel shows the VoCL (including SLA liabilities) against cumulative downtime in hours. The right-hand side panel shows the VoCL (including SLA liabilities) against customer utilization of the instances over the billing month. It assumes a fixed cumulative downtime of 3.7 hours (the first threshold). The service credit shares annotated in percentages.
Nevertheless, the short-run marginal incentive returns to the prevailing level after a downtime threshold is exceeded and before another downtime threshold is reached.
The difference between the on-demand and spot VoCL represent the value of flexibility to consumers. Competitive pressures and compute availability can push VoCL down over time.
The difference between the on-demand and spot VoCL represent the value of flexibility to consumers. Competitive pressures and compute availability can push VoCL down over time.
On-demand (excluding SLA liability) and spot VoCL timeseries for the p6-b300.48xlarge instance, for power draw based on training median utilization.
The graph aboves shows the VoCL of a frontier AI instance with NVIDIA’s Blackwell B300 GPU chip, launched in 2025. Relative to an on-demand VoCL of $13,600/MWh, the spot price has declined from around $6,700/MWh to $2,500/MWh between inception in November 2025 and May 2026.
Low price caps in the energy market can eliminate incentives for natural flexibility that could be enabled with full-strength price formation.
Low price caps in the energy market can eliminate incentives for natural flexibility that could be enabled with full-strength price formation.
Hypothetical aggregate demand bid curve. Assumes 80GW total demand, split between 90% on-demand and 10% interruptible, and a split in each service class split by GPU type as follows: 50% frontier AI, 40% mainstream AI and 10% early AI. Bids are set as per the median instance VoCL for each GPU type and service class.
Where VoCL is relatively low, bid-in demand with high price caps could incentivize natural flexibility.
Questions?
Questions?
For more information, contact farhad.billimoria@oxfordenergy.org and conleigh.byers@fticonsulting.com.
This analysis represents the authors’ work only and does not the represent the views of any organization, company, or institution.
Cover image by Cyril Fresillon/CC IN2P3/CNRS Images.
Page updated
Google Sites
Report abuse