Open data on public expenditure — datasets, categories and a SPARQL endpoint for research and analysis

Tracking DOE supercomputer electricity bills in public records

Public spending data can reveal the cost of large-scale computing, but the meaning of a bill depends on where it was created. Records grouped under the label US Department of Energy National Laboratory Supercomputer Electricity Consumption Billing can include utility invoices, internal cost transfers and payments to contractors operating laboratory sites. The label may sound like a direct measure of power drawn by each supercomputer. It is not: a federal payment can cover an entire research campus without identifying which machine ran, or for how long.

Publicspending.net is an independent, self-funded initiative that gathers and standardises public payment records for open comparison. Its datasets, graphs and bulk downloads can be explored through a SPARQL endpoint and ontology-based tools. For this topic, the most useful approach is to trace each amount from the federal payment back to the contract, site and billing period while keeping financial allocation separate from physical measurement.

Billing is a chain, not a meter

Electricity billing for laboratory computing crosses several accounting boundaries. The federal site, operating contractor, utility and research programme may each record a different amount for the same electricity. A useful investigation begins by naming those layers rather than treating every row as machine use.

Record layer Main content Typical timing What it can establish
Utility invoice Site consumption, billed kWh, demand charges and service period Monthly or periodic The amount charged for a site
Contractor recharge Internal allocation to a laboratory, division or project Monthly or quarterly Which activity carries an allocated cost
Federal payment Cash paid by a DOE entity to a supplier or contractor After invoicing and approval Where public expenditure was recorded
Submeter telemetry Electrical readings for a machine, rack or circuit Minutes, hours or days The power used by particular equipment

The financial layers answer different questions and should not be added together without checking for duplication. A utility invoice establishes what was billed, a recharge allocates that cost internally, and a federal payment records cash transferred from the agency. Metering provides a different form of evidence because it measures equipment directly. When public data contain payment and contract records, they may support the first three layers while leaving machine-level consumption unobserved.

An Australian household shows the same distinction. A quarterly bill from Ausgrid or an energy retailer states total usage, tariffs and charges, but it rarely identifies which appliance caused the consumption. A federal laboratory invoice can therefore answer a cost question without answering a performance question.

What laboratory records can reveal

Many US Department of Energy national laboratories operate high-performance systems for climate modelling, materials science, nuclear research and other compute-intensive work. Names such as Oak Ridge, Lawrence Livermore, Argonne and Fermilab may appear as recipients, operators or cost centres in payment records. Those rows reveal where public money was booked, but machine-level allocation may occur inside a contractor’s charging system.

A useful record bundle can contain the agency, recipient, supplier, amount, currency, invoice or service date, posting date, payment date, contract identifier, accounting category and free-text description. Depending on the source, recharges may add an allocation rate, cost pool or service centre. These fields allow researchers to distinguish payments to utilities from broader site-service contracts and to connect expenditure with the laboratory responsible for it.

Descriptions often use broad labels such as electricity, utilities or site services. A chilled-water charge may not be separable from general infrastructure, contractor overhead may be bundled, and redactions can suppress supplier details. A payment may also represent an internal transfer rather than final expenditure at the utility. Record what each row proves, and label campus-billed, internally allocated and equipment-metered figures separately.

Dates, budgets and missing time

Chronology needs equal attention. A public-health warning about laboratory reporting delays applies to public accounts as well: an invoice date, a service date and a payment date describe different events. Treating them as interchangeable can create apparent changes in consumption that actually reflect reporting schedules.

US federal records generally align to a financial year beginning on 1 October, while a supplier invoice may cover a calendar month or a contract period straddling year-end. Data extracts may use publication, posting or payment dates as their principal time field. A consistent comparison must select one event, document the choice and test how results change under alternative assumptions.

Budget data introduces another layer. Congressional appropriations fund laboratory operations, contracts create obligations, and treasury records show payments. Contractor recharges then assign a share of those costs to programmes or projects. Cash paid indicates what left the agency, a recharge indicates where a laboratory assigned cost, and metered data indicates when equipment used electricity. None of those measures can substitute for the others without supporting documentation.

Comparing systems without false precision

Energy and demand must remain separate. kilowatt-hours measure energy, while kilowatts or kilovolt-amperes measure demand at a particular point. Demand charges can dominate a laboratory’s bill even when annual consumption resembles that of another site. Comparisons based on expenditure alone may therefore measure peak pricing and contract structure as much as computing activity.

Normalisation also requires operating context. Climate, cooling systems, power usage effectiveness, machine age, utilisation and operating schedules vary between laboratories and generations of hardware. A high bill can reflect a large research campus, an experimental facility or substantial cooling demand rather than inefficient computation. Even the timing of training runs and simulations can create sharp peaks that make monthly totals unstable.

A defensible comparison should adjust for inflation, identify contract variations and duplicate payments, and state which sites and cost categories are covered. Where only invoices exist, results should be described as billed electricity cost rather than measured supercomputer consumption. Confidence labels can distinguish directly metered figures from utility totals, internal allocations and estimated values, preventing a confident presentation from outrunning the underlying evidence.

Lessons from Australia’s electricity market

Australia’s electricity market offers useful contrasts. The National Electricity Market covers much of eastern and southern Australia and operates under the National Electricity Rules, with the Australian Energy Market Regulator overseeing major market participants. Western Australia and the Northern Territory operate outside that wholesale market. This fragmented structure means a comparison between Australian and US research facilities must account for different regulators, network arrangements and tariff structures.

Everyday household routines complicate comparisons further. Many Sydney households operate rooftop solar and face time-of-use tariffs, so daytime generation can offset evening consumption. Some Melbourne households use gas for heating and purchase electricity in fixed blocks, while others choose variable rates. NSW’s Energy Made Easy and Victoria’s Victorian Energy Compare help households benchmark available offers, although neither reveals appliance-level consumption from a bill.

Consumer protections sit within the National Energy Retail Law and National Energy Retail Rules. These establish obligations covering billing, estimates, marketing and disconnection, while the reference market price acts as a regulatory benchmark rather than every household’s chosen plan. State distribution networks and retail regulators also influence the charges visible on a household invoice. Laboratory-scale comparisons therefore face the same separation between regulated network costs, retailer charges and the underlying energy service.

Australia’s National Compute Infrastructure in Canberra and the Pawsey facility in Perth provide closer research analogues to a national laboratory. Their electricity costs may be distributed across Commonwealth funding agreements, university overheads, facility operators and utility accounts. Federal and state audit publications, together with freedom-of-information material, add transparency but not automatic comparability. Australian privacy requirements can also limit publication of operational meter data, making financial records an important public source without making them a complete measure of computing use.

From raw payments to defensible analysis

Publicspending.net provides datasets, standardised views, statistics, graphs and bulk downloads across its country and jurisdiction coverage. Its material includes Australia, Greece, the United Kingdom and the United States, with subnational coverage for Chicago, Massachusetts and Alaska. For Australian users, this wider collection offers a way to compare federal, state and local publication practices as well as electricity classifications.

A practical US workflow starts by filtering for DOE as the publishing agency, then identifying laboratory recipients and descriptions containing electricity, utilities or relevant service categories. A SPARQL query can group payments by laboratory, supplier and financial year, while ontology links can connect recipient organisations, contracts and expenditure records. A second pass should test whether electricity appears as a distinct charge, a combined utilities invoice or an internal recharge.

Results should retain the original amount, currency, dates, identifiers and source references alongside any calculated totals. Contract amendments, revised invoices and apparent duplicates need an explicit treatment, while inflation adjustments should be separated from the reported figures. Graphs can then show billed expenditure or consumption trends without implying that all rows represent the same physical service. Stable citations and reproducible queries make the analysis usable by journalists, researchers and public-interest groups.

Follow the evidence from invoice to insight

Start with the public records rather than a headline estimate. Download the relevant DOE payment extracts from publicspending.net, follow the ontology links to the recipient and contract, and inspect the underlying electricity or utilities descriptions. That route reveals what is documented, what is allocated and what remains unseen.

Run a SPARQL query for one laboratory and financial year, reconcile billed amounts against payments, then publish the reproducible result with its assumptions and limitations.