On July 1, 2026, Virginia turned a hidden AI input into a taxable line item. Qualifying data center electricity consumption is now charged $0.011 per kilowatt-hour—1.1 cents—under a measure the state estimates can raise hundreds of millions of dollars while preserving its equipment-tax exemption.
That creates an obvious question for anyone buying cloud compute or AI tokens: does this end up on my bill?
The short answer is yes in economic terms, but probably not as a neat “Virginia power tax” beside your API usage. The legal payer, facility contract, workload location, provider margin, and pricing strategy all sit between the meter and your invoice.
The tax in one table
Item
Practical meaning
Rate
$0.011 for each taxable kWh
Effective date
July 1, 2026
Scheduled end
Before July 1, 2028 unless extended or replaced
Tax base
Qualifying data center electricity consumption
State rationale
Capture public value from exceptional power demand
Separate incentive
Virginia retains its major data center equipment exemption
Estimated revenue
Roughly $600 million annually in state estimates
This is a consumption tax, not a carbon price. A kilowatt-hour from nuclear, solar, gas, or a mixed grid faces the same nominal rate. It is also different from an ordinary utility tariff designed to fund substations and generation.
annual tax = average load in kW × operating hours × $0.011/kWh
Assuming a continuous average load:
Average facility load
Annual electricity
Approximate annual tax
10 MW
87.6 million kWh
$963,600
50 MW
438 million kWh
$4.82 million
100 MW
876 million kWh
$9.64 million
500 MW
4.38 billion kWh
$48.18 million
1 GW
8.76 billion kWh
$96.36 million
These are illustrations, not bills. Real load varies; legal definitions, meter boundaries, exemptions, and effective dates matter. Still, the table shows why a rate that sounds small becomes material at hyperscale.
Who pays first—and who pays eventually
The economic chain usually has several layers:
text
utility meter → facility/operator → colocation tenant or cloud provider
→ model company → API or subscription customer
Owner-operated campus
If a hyperscaler owns and operates the building, it receives the electricity service and bears the tax directly. It can absorb the expense, optimize the site, move incremental jobs elsewhere, or recover it through cloud pricing.
Colocation facility
In a colocation building, the operator may be the customer at the utility meter while tenants pay power under contracts. Those contracts often include electricity, capacity, utility riders, taxes, and a management markup. The legal incidence can sit with the operator while the economic incidence passes to tenants.
AI model provider renting cloud capacity
A model company may never see a Virginia tax invoice. It buys reserved accelerators or managed compute from a cloud provider whose price already reflects regional power, facility, and tax costs. Renegotiation dates determine how quickly new costs flow through.
API customer
The API customer buys tokens, not kilowatt-hours. A provider can spread one region’s higher expense across a national price, charge regional premiums, route workloads, or accept lower margin. You should not expect a precise public conversion.
What it means per million tokens
To isolate the tax, we need an energy assumption. That is harder than it sounds because vendors rarely publish task-level energy, and consumption changes with hardware, batching, utilization, model size, prompt length, output length, cooling, and the facility's power usage effectiveness.
Use a transparent range rather than false precision. Suppose one million tokens require between 0.1 and 10 kWh of facility electricity across very different workloads and architectures. The Virginia tax adds:
Illustrative energy per 1M tokens
Added Virginia tax
0.1 kWh
$0.0011
1 kWh
$0.011
10 kWh
$0.11
Even the top illustrative case is small beside many API prices. The point is not that power is irrelevant. The tax is only one component of power cost, and inference economics include expensive chips, servers, buildings, networking, engineering, financing, redundancy, and profit.
This is why token pricing rarely moves one-for-one with an electricity charge. Our AI token pricing explainer breaks down the full bill mechanics, while the agent monthly-cost guide converts them into a workflow budget.
The four most likely provider responses
1. Absorb it
Competitive providers may keep national token prices unchanged and accept slightly lower margin in Virginia. Falling inference costs can more than offset a new local tax.
2. Pass it through contractually
Colocation and enterprise cloud agreements may contain tax and utility pass-through clauses. Customers with Virginia-specific capacity are more likely to see a distinct adjustment than retail API customers.
3. Route marginal workloads elsewhere
Inference jobs that do not require a specific region can move to a cheaper site. Data residency, latency, capacity, and network transfer limit this response. Training jobs are easier to schedule geographically than latency-sensitive interactive queries.
4. Invest in efficiency
The cleanest response is doing more work per kWh: better batching, quantization, speculative decoding, efficient accelerators, cooling, and model routing. Our model-selection energy math shows that choosing a right-sized model can reduce energy much more than this tax changes it.
Why Virginia chose a tax instead of ending the incentive
Virginia became the world’s densest data center market partly through favorable tax treatment, fiber, land, government proximity, and utility access. Ending equipment exemptions abruptly could send future projects to competing states. Ignoring extraordinary electricity growth risks shifting infrastructure costs and political anger onto residents.
The consumption tax attempts a middle path: keep the capital-equipment incentive, then collect revenue in proportion to ongoing electricity use. It effectively says that a facility consuming more public-grid capacity should contribute more, regardless of how much equipment it buys.
Whether that is good policy depends on what revenue funds, whether utilities separately protect ordinary ratepayers, and whether the tax drives efficient design or merely relocates load.
Will other states copy it?
Virginia is a natural test because it has both a mature data center cluster and visible grid pressure. Other states will watch five outcomes:
Did projects relocate or continue?
Did operators successfully pass the cost to tenants?
Did revenue reduce household or infrastructure burdens?
Did power demand or design efficiency change?
Did litigation or administrative complexity undermine collection?
A copycat state may choose a lower rate, a tiered rate, a carbon-based charge, or a tax credit for dedicated clean generation. States with abundant power may reject the approach to preserve a recruitment advantage.
What buyers should ask in cloud contracts
Enterprise customers should stop treating data center taxes as someone else’s procurement detail. Ask:
Which taxes and utility riders are pass-through expenses?
Can the provider change price mid-term?
Is compute tied to a region, and can you move it without egress penalties?
Does reserved capacity include power escalation?
Are renewable or nuclear claims contractual or accounting attributes?
Can you see workload-level token and tool usage even if energy is not disclosed?
For retail subscriptions, the practical defense is simpler: compare total workflow value rather than obsess over invisible cents. Our AI subscription comparison uses cost per completed research, writing, spreadsheet, and meeting task.
The larger pricing lesson
Virginia’s tax exposes a fact the industry prefers to abstract: AI is a physical service. Every token depends on electricity, cooling, land, transmission, chips, and local permission. API pricing packages those inputs into a software-looking unit.
The direct tax per million tokens may be negligible. The indirect effect is not. If states demand infrastructure contributions, utilities assign upgrade costs to new loads, and communities delay sites, the marginal cost and timing of capacity change. Providers then make portfolio decisions about where to build and what workloads deserve scarce accelerators.
That is also why cheap and efficient models are related. A model that completes a task with fewer active parameters, fewer retries, and less context consumes less infrastructure and generally costs less to serve.
Verdict
Will every AI API add 1.1 cents per kWh to your invoice? No.
Does the industry ultimately bear a new operating cost? Yes.
Is electricity the dominant API cost? Usually not by itself; hardware economics and utilization remain decisive.
Could Virginia become a template? Yes, especially where communities believe data center incentives socialize grid costs.
Your AI bill already contained a power bill. Virginia simply made one part of it easier to see.
A procurement worksheet for Virginia exposure
Enterprise buyers can quantify exposure without pretending to know a provider's secret energy-per-token figure. List contracted regions, reserved capacity, colocation providers, renewal dates, tax pass-through clauses, and the share of workloads that can legally move. Ask the provider whether the new charge is included in current rates and whether a later adjustment requires notice.
Then model three cases: the provider absorbs the tax, passes through the full facility amount to Virginia-specific customers, or spreads it across a national service. The range is more honest than a single API-price prediction. Include data-egress, latency, residency, and migration engineering before declaring another region cheaper.
For internal budgeting, keep the tax separate from the much larger underlying electricity bill. That distinction lets teams see whether a change comes from policy, energy markets, infrastructure upgrades, or compute volume. It also prevents a highly visible tax from distracting from the bigger efficiency levers: right-sized models, bounded outputs, cache reuse, and fewer failed loops.
This article explains general pricing and policy mechanics, not tax, legal, or procurement advice. The Virginia measure and estimates are current through July 26, 2026 and may be amended, interpreted, or replaced before the scheduled 2028 endpoint.
