inference/research
verifiedhigh · 0.90updated 2026-10-04

Cost to RUN, not to buy: US average industrial electricity price, 8.91 cents/kWh trailing-12-months to July 2026 (EIA Electric Power Monthly Table 5.3); 8.13 was the CY2024 figure

United States (national average across all 50 states and DC)

Fields

id
supply/us-electricity-industrial-tariff-2024
Kind
unknown
Region
United States (national average across all 50 states and DC)
Channels
contact-saleson-prem-cloud
Price (USD)
0.0891
Price basis
USD per kWh, INDUSTRIAL sector average retail price to ultimate customers, all US states and DC. REFRESHED 2026-10-04, and this is now the freshest figure the EIA publishes rather than a two-year-old annual average. The headline number 0.0891 (8.91 cents/kWh) is the ROLLING 12 MONTHS ENDING JULY 2026 industrial average, read from EIA Electric Power Monthly Table 5.3 'Average Price of Electricity to Ultimate Customers: Total by End-Use Sector', released 2026-09-24, series data through July 2026 (source pd-eia-epm-table-5-3). SUPERSEDED VALUE, KEPT VISIBLE SO NOBODY TREATS IT AS CURRENT: this record previously carried 0.0813 (8.13 c/kWh), the CALENDAR-YEAR 2024 figure. That number was correct as an annual average and is now ~2.7 years behind on a rising series. It is retained here because docs/08-cost-per-token.md still cites it inline and because the year-on-year change is itself the finding. DO NOT MIX THE TWO IN ONE COST MODEL. THE FULL SERIES, so drift is visible rather than asserted: industrial annual 2022 8.32, 2023 8.04, 2024 8.13, 2025 8.62 cents/kWh. Trailing-12-months ending July 2026 = 8.91. Calendar 2026 monthly: January 9.29, February 8.95, March 8.58, April 8.66, May 8.71, June 9.17, July 9.77. Year-to-date 2026 = 9.03. DIRECTION OF TRAVEL, which is why this record was refreshed rather than merely re-stamped: 8.13 -> 8.62 is +6.0% in one year, and 8.13 -> 8.91 on the trailing basis is +9.6%. Electricity is a RISING input to inference cost, not a flat one. CROSS-SECTOR, same source and same window: residential 17.99 and commercial 13.85 and all-sector total 14.12 cents/kWh on the trailing-12-months-to-July-2026 basis. WHAT STATUS THE EIA PUTS ON ITS OWN NEWEST NUMBERS, quoted from the table footnote: 'Values for 2024 and prior years are final. Values for 2026 and 2025 are preliminary estimates based on a cutoff model sample.' So 8.91 is PRELIMINARY and will be revised; the CY2024 8.13 is final. A refreshed preliminary figure is still the right stored figure, but it is provisional in a way the old one was not. NOT A QUOTE, NOT A DATA-CENTRE TARIFF - see notes for the scope limits that matter more than the headline number. PRICE CLASS: this is an AVERAGE RETAIL TARIFF COLLECTED FROM BILLS, i.e. an asking price paid by the average customer, NOT a sold or realised transaction price, NOT a negotiated contract rate and NOT a wholesale rate. A large industrial buyer pays something else entirely. OBSERVED: 2026-10-04. RE-VERIFY BY: 2026-11-04 for the preliminary monthly figures, because this series is revised MONTHLY and the next release was published on the page as October 23 2026; and 2027-01-04 for the CY2025 annual figure once the Electric Power Annual publishes data for 2025, due October 2026, which will make the annual series final. PERISHABLE-FIGURE STAMP (added 2026-10-04, uniform across data/supply/ so it is greppable). PRICE CLASS: PUBLISHED TARIFF AVERAGE, ASKING - an asking rate derived from billed retail revenue over billed volume, NOT a realised transaction price and NOT a negotiated contract rate. OBSERVED: 2026-10-04. RE-VERIFY BY: 2026-11-04 (monthly preliminary) / 2027-01-04 (annual final). UNIT: USD per kWh, industrial sector average retail price, US total. READING RULE: DERIVED, and the derivation is the point. The ORIGINAL 8.13 was computed as EPA Table 2.3 revenue divided by EPA Table 2.2 volume; the refreshed 8.91 is a figure the EIA computes and publishes directly in Table 5.3, so it needs no derivation. The old derivation has now been INDEPENDENTLY CONFIRMED rather than merely self-consistent: EPA Table 2.4 publishes 8.13 directly as the 2024 industrial average, reproducing the revenue/volume quotient to the cent (source pd-eia-epa-table-2-4). PROVIDER-TYPE SPREAD, because the blended 8.13 hid it: for 2024 the industrial average differs by what the customer actually buys - Full-Service Providers 7.74, Competitive Service Providers 9.84, Energy-Only 6.74, Delivery-Only 3.10 cents/kWh. A ~46% spread between 6.74 and 9.84 on the same year and the same sector, which the blended figure conceals. A colo tenant buying bundled energy plus delivery should read the Full-Service or Competitive row, not the national blend. No further perishable flag is published for this figure.
Availability
in-stock
Lead time source
structural-na
Lead time basis
lead_time_weeks is null because NO NUMBER APPLIES: this record prices a non-accelerator input (electricity, egress, PUE/uptime) and lists accelerators only as the load that consumes it. There is no order-to-delivery step, so a lead time is the wrong quantity rather than a missing one.
Export control regime
no EAR citation: no controlled item is transferred by this transaction
Export control note
MECHANISM: this record prices a non-accelerator input (electricity, egress, uptime/PUE) and lists accelerators only as the load that consumes it. No controlled item is transferred, so the 3A090 regime does not attach.
notes
THE MISSING COST INPUT FOR ANY COST-PER-TOKEN MODEL. Cloud per-GPU-hour is already priced and recorded elsewhere in this repo; the electricity underneath an on-prem or colo deployment is not, and it is the number that decides whether owning hardware beats renting it. Worked example against the measured 8xH200 node, using this record and the MLPerf trace together: the measured loaded plateau of 5806.93 W is 5.8069 kWh per node-hour, so ONE NODE-HOUR COSTS $0.4721 AT THE INDUSTRIAL RATE, $0.7404 AT THE COMMERCIAL RATE (12.75 c), and $0.7514 AT THE ALL-SECTOR AVERAGE (12.94 c). The idle floor of 2422.6 W still costs $0.197 per node-hour at the industrial rate. THAT IS THE WHOLE ARGUMENT IN THREE NUMBERS: 8xH200 hardware is a purchase order, but the electricity is a permanent annuity that runs whether or not the node is busy - roughly $345 per node-MONTH at the industrial rate if left idle 24x7 ($0.197 x 730 h), and about $828 per node-month under full load ($0.4721 x 730 h). SCOPE LIMITS, stated plainly because a wrong tariff is worse than no tariff: (1) These are AVERAGE RETAIL prices to ultimate customers across all sectors and states, including residential volume. A real hyperscale AI load is bought at industrial or wholesale rates in ONE state under a contract that often prices CAPACITY separately from ENERGY. (2) They exclude demand charges, ramp rates, power-factor penalties and transmission congestion charges, all of which matter enormously at 5-10 MW and none of which is visible here. (3) They exclude PUE, so the figure above is utility-side only - a colo-quoted price bundles PUE into a per-kW-month number instead. (4) They exclude taxes, transmission and distribution line items that make up a material share of a retail bill. COMPARISON THAT SHOULD MAKE ANY READER PAUSE: at 8.13 c/kWh, the ELECTRICITY for an 8xH200 node-hour is $0.4721, while the cheapest published peer-to-peer H100 rental in this repo is $1.69/GPU-hr and the RunPod Community H200 ask is $4.59/GPU-hr ($36.72/node-hour). Electricity is therefore roughly 1% of what renting the same hardware costs - the machine is the price, not the watt. That holds even at the 20 c/kWh that expensive-grid industrial buyers face. THE CONSEQUENCE FOR THIS REPO: joules per token decides SCHEDULING and BATCHING and PUE and cooling strategy, because those are the levers that change watts; it is NOT the lever that decides buy-versus-rent. Model-cost work should not treat energy price as the dominant term. EXPORT CONTROL (2026-10-04): NOT APPLICABLE - this record prices electricity (USD/kWh, EIA industrial sector average) and lists accelerators only as the load that consumes it. No controlled item is transferred, so the 3A090 hardware regime does not attach. Recorded null, not false. One genuinely relevant export-adjacent fact for anyone siting capacity on this figure, from the load side rather than the item side: NVIDIA's 91 FR 1684 licence-application condition (1) requires an exporter to certify US supply sufficiency 'taking into account normal lead times', i.e. US electricity and grid capacity is an input to whether licences are granted at all - a power-constrained US is a supply-constrained US. Sources: spl-bis-2026-01-advanced-computing-case-by-case. LEAD TIME: null - correct and not a gap. A token-priced serverless API is provisioned on request, so there is no lead time to publish; the binding constraint when a model is at capacity is provider-side queuing and any per-account concurrency limit, which this repo records in notes where a vendor publishes one (see the DeepSeek concurrency figures already here). Deliberately NOT recorded: any 'model capacity lead time' estimate, because vendors publish none and the SEO 'GPU lead time' numbers that circulate do not apply to a rented endpoint. Probe log in source spl-negative-no-oem-publishes-lead-time. STALENESS FLAG - RESOLVED 2026-10-04, and this is the audit result rather than a restatement of the problem. The previous version of this flag said the CY2024 figure was 'roughly TWO YEARS old ... well past the quarterly re-verification bar' and should not be used for a current power cost without a refresh. That refresh has now been done. Two findings, one of which corrects the original diagnosis and one of which makes the record sharper. (1) THE ORIGINAL DIAGNOSIS WAS PARTLY WRONG ABOUT WHY THE DATA WAS OLD. It was not neglect: on 2026-10-04 the EIA Electric Power Annual landing page states 'With Data for 2024, Release Date: October 16, 2025, Next Release: October 2026', so CY2024 was the publication FRONTIER of the annual series when this record was written. The prior wave flagged a stale value without checking whether a fresher one existed; a fresher one did, one table over, in the Electric Power Monthly series that runs ahead of the annual report. The lesson generalises: before calling a figure stale, check whether the publisher has a faster series carrying the same quantity. (2) THE FIGURE IS NOW FRESH BUT PRELIMINARY, which is a different failure mode from stale and worth naming. The Electric Power Monthly footnote says 2026 and 2025 values are 'preliminary estimates based on a cutoff model sample' while 2024 and prior are final. So this record has traded a final-but-dated number for a current-but-revisable one. That is the correct trade for a running-cost input, and it is why the stamp above carries a 2026-11-04 re-verify date, one month, not the usual quarter. RE-DERIVED WORKED EXAMPLE at the new rate, using the same measured MLPerf 8xH200 figures as before so the comparison is clean: the loaded plateau of 5806.93 W is 5.8069 kWh per node-hour, so ONE NODE-HOUR COSTS $0.5174 at 8.91 c/kWh, against $0.4721 at the old 8.13 c/kWh - the same node costs $0.0453 MORE per hour than this record used to say, about +9.6%, which is the tariff increase passed straight through. Idle floor of 2422.6 W still costs $0.2159 per node-hour (was $0.1970). Per node-MONTH at 730 h: about $377.70 loaded (was $344.64) and $157.57 idle (was $143.78). THE CONCLUSION OF THE ORIGINAL RECORD SURVIVES THE REFRESH, AND IS IF ANYTHING STRONGER: at $0.5174 per node-hour the electricity for an 8xH200 node remains roughly 1% of what renting the same hardware costs (the RunPod Community H200 ask is $4.59/GPU-hr = $36.72/node-hour, and the cheapest published peer-to-peer H100 rental in this repo is $1.69/GPU-hr). A 9.6% tariff increase does not move a number that was 1% of the alternative. Energy price still decides SCHEDULING, BATCHING, PUE and cooling - the levers that change watts - and still does NOT decide buy-versus-rent, where depreciation dominates. SCOPE LIMITS, restated because they matter more than the headline and they are unchanged by the refresh: (1) these are AVERAGE RETAIL prices to ultimate customers across all states, and a hyperscale load is bought at industrial or wholesale rates in ONE state under a contract that often prices CAPACITY separately from ENERGY; (2) they exclude demand charges, ramp rates, power-factor penalties and transmission congestion, all of which matter enormously at 5-10 MW; (3) they exclude PUE, so the figure is utility-side only, while a colo quotes PUE into a per-kW-month number instead; (4) they exclude taxes and T&D line items. ONE MORE CAVEAT THE EIA STATES ABOUT ITS OWN COLUMN, and it is specific to this repo's use: 'Utilities and energy service providers may classify commercial and industrial customers based on either NAICS codes or demands or usage falling within specified limits by rate schedule.' The industrial column is therefore a RATE-SCHEDULE CLASSIFICATION, not a census of industrial load, and it can absorb cryptocurrency-mining and data-centre consumption sitting behind an industrial tariff. Do not read 8.91 as what a data centre pays. UNVERIFIED, stated plainly: no data-centre-specific or PJM/ERCOT/Northern-Virginia industrial tariff was obtained for this pass. That is the number a siting decision actually needs and it is not in this record; a wholesale or utility-contract tariff is quote-only and would have to come from a tariff schedule or a utility account, neither of which is a citable public price list.

Notes

THE MISSING COST INPUT FOR ANY COST-PER-TOKEN MODEL. Cloud per-GPU-hour is already priced and recorded elsewhere in this repo; the electricity underneath an on-prem or colo deployment is not, and it is the number that decides whether owning hardware beats renting it. Worked example against the measured 8xH200 node, using this record and the MLPerf trace together: the measured loaded plateau of 5806.93 W is 5.8069 kWh per node-hour, so ONE NODE-HOUR COSTS $0.4721 AT THE INDUSTRIAL RATE, $0.7404 AT THE COMMERCIAL RATE (12.75 c), and $0.7514 AT THE ALL-SECTOR AVERAGE (12.94 c). The idle floor of 2422.6 W still costs $0.197 per node-hour at the industrial rate. THAT IS THE WHOLE ARGUMENT IN THREE NUMBERS: 8xH200 hardware is a purchase order, but the electricity is a permanent annuity that runs whether or not the node is busy - roughly $345 per node-MONTH at the industrial rate if left idle 24x7 ($0.197 x 730 h), and about $828 per node-month under full load ($0.4721 x 730 h). SCOPE LIMITS, stated plainly because a wrong tariff is worse than no tariff: (1) These are AVERAGE RETAIL prices to ultimate customers across all sectors and states, including residential volume. A real hyperscale AI load is bought at industrial or wholesale rates in ONE state under a contract that often prices CAPACITY separately from ENERGY. (2) They exclude demand charges, ramp rates, power-factor penalties and transmission congestion charges, all of which matter enormously at 5-10 MW and none of which is visible here. (3) They exclude PUE, so the figure above is utility-side only - a colo-quoted price bundles PUE into a per-kW-month number instead. (4) They exclude taxes, transmission and distribution line items that make up a material share of a retail bill. COMPARISON THAT SHOULD MAKE ANY READER PAUSE: at 8.13 c/kWh, the ELECTRICITY for an 8xH200 node-hour is $0.4721, while the cheapest published peer-to-peer H100 rental in this repo is $1.69/GPU-hr and the RunPod Community H200 ask is $4.59/GPU-hr ($36.72/node-hour). Electricity is therefore roughly 1% of what renting the same hardware costs - the machine is the price, not the watt. That holds even at the 20 c/kWh that expensive-grid industrial buyers face. THE CONSEQUENCE FOR THIS REPO: joules per token decides SCHEDULING and BATCHING and PUE and cooling strategy, because those are the levers that change watts; it is NOT the lever that decides buy-versus-rent. Model-cost work should not treat energy price as the dominant term.

EXPORT CONTROL (2026-10-04): NOT APPLICABLE - this record prices electricity (USD/kWh, EIA industrial sector average) and lists accelerators only as the load that consumes it. No controlled item is transferred, so the 3A090 hardware regime does not attach. Recorded null, not false. One genuinely relevant export-adjacent fact for anyone siting capacity on this figure, from the load side rather than the item side: NVIDIA's 91 FR 1684 licence-application condition (1) requires an exporter to certify US supply sufficiency 'taking into account normal lead times', i.e. US electricity and grid capacity is an input to whether licences are granted at all - a power-constrained US is a supply-constrained US. Sources: spl-bis-2026-01-advanced-computing-case-by-case.

LEAD TIME: null - correct and not a gap. A token-priced serverless API is provisioned on request, so there is no lead time to publish; the binding constraint when a model is at capacity is provider-side queuing and any per-account concurrency limit, which this repo records in notes where a vendor publishes one (see the DeepSeek concurrency figures already here). Deliberately NOT recorded: any 'model capacity lead time' estimate, because vendors publish none and the SEO 'GPU lead time' numbers that circulate do not apply to a rented endpoint. Probe log in source spl-negative-no-oem-publishes-lead-time.

STALENESS FLAG - RESOLVED 2026-10-04, and this is the audit result rather than a restatement of the problem. The previous version of this flag said the CY2024 figure was 'roughly TWO YEARS old ... well past the quarterly re-verification bar' and should not be used for a current power cost without a refresh. That refresh has now been done. Two findings, one of which corrects the original diagnosis and one of which makes the record sharper. (1) THE ORIGINAL DIAGNOSIS WAS PARTLY WRONG ABOUT WHY THE DATA WAS OLD. It was not neglect: on 2026-10-04 the EIA Electric Power Annual landing page states 'With Data for 2024, Release Date: October 16, 2025, Next Release: October 2026', so CY2024 was the publication FRONTIER of the annual series when this record was written. The prior wave flagged a stale value without checking whether a fresher one existed; a fresher one did, one table over, in the Electric Power Monthly series that runs ahead of the annual report. The lesson generalises: before calling a figure stale, check whether the publisher has a faster series carrying the same quantity. (2) THE FIGURE IS NOW FRESH BUT PRELIMINARY, which is a different failure mode from stale and worth naming. The Electric Power Monthly footnote says 2026 and 2025 values are 'preliminary estimates based on a cutoff model sample' while 2024 and prior are final. So this record has traded a final-but-dated number for a current-but-revisable one. That is the correct trade for a running-cost input, and it is why the stamp above carries a 2026-11-04 re-verify date, one month, not the usual quarter. RE-DERIVED WORKED EXAMPLE at the new rate, using the same measured MLPerf 8xH200 figures as before so the comparison is clean: the loaded plateau of 5806.93 W is 5.8069 kWh per node-hour, so ONE NODE-HOUR COSTS $0.5174 at 8.91 c/kWh, against $0.4721 at the old 8.13 c/kWh - the same node costs $0.0453 MORE per hour than this record used to say, about +9.6%, which is the tariff increase passed straight through. Idle floor of 2422.6 W still costs $0.2159 per node-hour (was $0.1970). Per node-MONTH at 730 h: about $377.70 loaded (was $344.64) and $157.57 idle (was $143.78). THE CONCLUSION OF THE ORIGINAL RECORD SURVIVES THE REFRESH, AND IS IF ANYTHING STRONGER: at $0.5174 per node-hour the electricity for an 8xH200 node remains roughly 1% of what renting the same hardware costs (the RunPod Community H200 ask is $4.59/GPU-hr = $36.72/node-hour, and the cheapest published peer-to-peer H100 rental in this repo is $1.69/GPU-hr). A 9.6% tariff increase does not move a number that was 1% of the alternative. Energy price still decides SCHEDULING, BATCHING, PUE and cooling - the levers that change watts - and still does NOT decide buy-versus-rent, where depreciation dominates. SCOPE LIMITS, restated because they matter more than the headline and they are unchanged by the refresh: (1) these are AVERAGE RETAIL prices to ultimate customers across all states, and a hyperscale load is bought at industrial or wholesale rates in ONE state under a contract that often prices CAPACITY separately from ENERGY; (2) they exclude demand charges, ramp rates, power-factor penalties and transmission congestion, all of which matter enormously at 5-10 MW; (3) they exclude PUE, so the figure is utility-side only, while a colo quotes PUE into a per-kW-month number instead; (4) they exclude taxes and T&D line items. ONE MORE CAVEAT THE EIA STATES ABOUT ITS OWN COLUMN, and it is specific to this repo's use: 'Utilities and energy service providers may classify commercial and industrial customers based on either NAICS codes or demands or usage falling within specified limits by rate schedule.' The industrial column is therefore a RATE-SCHEDULE CLASSIFICATION, not a census of industrial load, and it can absorb cryptocurrency-mining and data-centre consumption sitting behind an industrial tariff. Do not read 8.91 as what a data centre pays. UNVERIFIED, stated plainly: no data-centre-specific or PJM/ERCOT/Northern-Virginia industrial tariff was obtained for this pass. That is the number a siting decision actually needs and it is not in this record; a wholesale or utility-contract tariff is quote-only and would have to come from a tariff schedule or a utility account, neither of which is a citable public price list.

Cross-references

Accelerators (4)