Data catalog

Every WattSaid number: what it is, what it covers, how often it changes, how far back WattSaid keeps it and how it is worked out, with its own page for the history, downloads and citation. The numbers are WattSaid’s own, so their sources are not named.

Onsite power

Generating capacity at US businesses' and factories' own sites

Unit: GW · Monthly · Since Aug 2024

Generating capacity in operation at US businesses' and factories' own sites, in GW.

Covers
The US, generators of 1 MW or more owned by commercial and industrial sites, with or without combined heat and power, monthly. Utilities, independent power producers and generators under 1 MW are not counted.
Method
The nameplate capacity of every such generator in operation in the month, added up and converted to GW.

Gas-fired generating capacity at US businesses' own sites

Unit: GW · Monthly · Since Aug 2024

Generating capacity in operation at US businesses' and factories' own sites that runs on natural gas, in GW.

Covers
The US, generators of 1 MW or more owned by commercial and industrial sites, with or without combined heat and power, monthly. Utilities, independent power producers and generators under 1 MW are not counted.
Method
The nameplate capacity of every such generator in operation in the month whose energy source is natural gas, added up and converted to GW.

Onsite generation US businesses have reported plans to add

Unit: GW · Monthly · Since Sep 2024

Generating capacity that US businesses and factories have reported plans to add at their own sites, in GW.

Covers
The US, planned generators of 1 MW or more in the commercial and industrial sectors, monthly. The newest month usually comes one to two months after it ends.
Method
The nameplate capacity of every planned commercial and industrial generator on the month's list, added up and converted to GW. Plans count whatever their stage, from early planning to construction, so not all of them will be built.

Gas-fired onsite generation US businesses have reported plans to add

Unit: GW · Monthly · Since Sep 2024

Gas-fired generating capacity that US businesses and factories have reported plans to add at their own sites, in GW.

Covers
The US, planned generators of 1 MW or more in the commercial and industrial sectors, monthly. The newest month usually comes one to two months after it ends.
Method
The nameplate capacity of every planned commercial and industrial generator on the month's list whose energy source is natural gas, added up and converted to GW. Plans count whatever their stage, so not all of them will be built.

California onsite generation incentive applications, last 4 weeks

Unit: applications · Weekly · Since 20 Oct 2024

How many applications California's onsite generation incentive received in the last 4 weeks.

Covers
California, every application on the program's weekly statewide list, by the date it was received. Applications cancelled later still count.
Method
At each weekly list date, the applications received in the 28 days up to that date are counted. Each new list recounts the past two years of weeks, so earlier values can change.

Storage applied for under California's onsite generation incentive, last 4 weeks

Unit: MWh · Weekly · Since 20 Oct 2024

How much energy storage was applied for under California's onsite generation incentive in the last 4 weeks, in MWh.

Covers
California, the storage capacity stated on every application on the program's weekly statewide list. It is capacity applied for, not installed.
Method
The storage capacity on the applications received in the 28 days up to each weekly list date, added up and converted to MWh. Each new list recounts the past two years of weeks, so earlier values can change.

California customer-sited systems connected in the month

Unit: systems · Monthly · Since Sep 2023

How many customer-sited power systems California's utilities approved for connection in the month.

Covers
California, systems on the customer's side of the meter at homes, businesses and public sites, by the month their connection was approved.
Method
A count of systems by approval month in the utilities' records of customer-sited systems. The last 36 months are counted again each week from the newest records, so recent months can change.

California customer-sited capacity connected at businesses and public sites

Unit: MW · Monthly · Since Sep 2023

The size of the customer-sited systems connected at California businesses and public sites in the month, in MW.

Covers
California, systems at every customer type other than homes, by the month their connection was approved. A system with no customer type given is treated as a home.
Method
The AC size of each non-residential system approved for connection in the month, added up and converted to MW. Recent months are counted again each week, so they can change.

California customer-sited systems connected with a battery

Unit: % · Monthly · Since Sep 2023

The share of customer-sited systems connected in California in the month that came with a battery, in percent.

Covers
California, all customer-sited systems approved for connection in the month, homes included.
Method
Systems with any storage capacity, divided by all systems approved for connection that month, in percent.

US data centers with an air permit record

Unit: data centers · Daily · Since 14 Oct 2026

How many US data centers have a record in a national register of facilities with air permits.

Covers
The US, facilities classed under the industry code for data processing, hosting and related services. Data centers classed under another code are not counted.
Method
A count of the matching facilities, taken each day.

Texas data center air permit applications with public notice pending

Unit: applications · Daily · Since 26 Oct 2026

How many air permit applications for Texas data centers are pending with a public notice.

Covers
Texas, applications for the full, case-by-case air permits that need public notice. Standard permits and permits by rule need no notice, so they are not counted, and concrete batch plants for building data centers are left out.
Method
A count of the applications on the day's list of pending applications whose listed names mention a data center. An application named otherwise, such as a power plant built for a data center, is missed.

Alberta data center power plant applications under review

Unit: applications · Daily · Since 31 Oct 2026

How many applications for power plants at Alberta data centers the provincial regulator is reviewing.

Covers
Alberta, proceedings on the regulator's list of applications under review. An application leaves the list once it is decided.
Method
Proceedings whose description or applicant mentions a data center or AI center, and whose description mentions a power plant, generation or a thermal plant, counted from a daily copy of the list. Cost claims and standby or emergency generators are left out, and a proceeding with several applications counts once. A developer whose filings mention neither a data center nor an AI center is missed.

Texas applications to run a large load behind an existing power plant's meter

Unit: applications · Daily · Since 1 Nov 2026

How many cases Texas has on record for a large power user to sit behind the meter of an existing power plant.

Covers
Texas, cases for a net metering arrangement under the 2025 state law that requires approval of each one. Every case on record counts, whatever its stage.
Method
Each case is counted once, from a daily copy of the case list. Compliance filings that quote a case are left out.

Arkansas data center air permit applications on record

Unit: applications · Daily · Since 2 Nov 2026

How many air permit applications for Arkansas data centers are on record.

Covers
Arkansas, every air permit application listed under the industry code for data processing and hosting, whatever its status. Data centers classed under another code are missed.
Method
A count of the listed applications, from a daily copy of the state's list of air permit applications.

Self-generation and cogeneration filings with the federal regulator, last 30 days

Unit: filings · Daily · Since 16 Jul 2026

How many filings to certify a self-generation or cogeneration plant were made with the federal regulator in the last 30 days.

Covers
The US, filings that certify a small power plant or a cogeneration plant as a qualifying facility. The search reads full text, so other documents that name the certification form also count.
Method
A count of documents filed in the 30 days up to and including each completed day whose text or description names the certification form. One plant can file more than once, for example to amend its certificate.

Company filings mentioning behind-the-meter power, last 7 days

Unit: filings · Daily · Since 16 Jul 2026

How many company filings mentioned behind-the-meter power in the last 7 days.

Covers
Company filings with the US securities regulator, every filing type, searched in full text.
Method
A count of filings containing the exact phrase behind-the-meter, filed in the 7 days up to and including each completed day. A passing mention counts as much as a filing about a project.

Company filings mentioning behind-the-meter power and data centers, last 7 days

Unit: filings · Daily · Since 16 Jul 2026

How many company filings mentioned both behind-the-meter power and data centers in the last 7 days.

Covers
Company filings with the US securities regulator, every filing type, searched in full text.
Method
A count of filings containing both the exact phrase behind-the-meter and the exact phrase data center, anywhere in the text, filed in the 7 days up to and including each completed day. The two phrases need not be about the same project.

Reviewed data center project identities; power availability not verified

Unit: projects · Daily · Since 29 Oct 2026 · In the API as btm-reviewed-projects

How many data center projects with onsite power have had their identity reviewed by WattSaid. It does not say whether their power is available.

Covers
Data center projects with onsite power in WattSaid's own research ledger. Cancelled projects are left out.
Method
A project counts only while a current review matches its exact research record and has checked its name, place, state, country and its link to onsite power against documents from a regulator or the company. A review expires, and a project whose record changes drops out until it is reviewed again. Capacity and operating status are not reviewed, so the count measures reviewed identities, not sites with power they can deliver.

US customer-sited generating capacity

Unit: GW · Monthly · Since Jan 2025

Generating capacity that US homes and businesses have installed on their own side of the meter.

Covers
The US, customer-sited capacity, net metered or not, monthly.
Method
The sum of net-metered and other customer-sited capacity for the month.

New York behind-the-meter solar, peak output against installed capacity

Unit: % · Daily · Since 1 Aug 2026

How much of New York's behind-the-meter solar was producing at the day's peak hour.

Covers
New York State, behind-the-meter solar, daily.
Method
The peak hour's estimated output divided by the installed capacity on the same day, in percent.

Onsite power filing activity against its usual

Unit: index · Daily · Since 5 Aug 2026 · In the API as onsite-activity-index

How busy onsite power filings are this week against their previous four weeks. 100 is usual.

Covers
US self-generation and cogeneration filings, company filings mentioning behind-the-meter power, and Texas data center air permit notices.
Method
Each input's average over the last 7 days is divided by its median over the 28 days before; the index is 100 times the average of those ratios, with equal weight for each input available that day. At least two inputs a day.

Data centers

Gas power built for data centers, operating

Unit: GW · Quarterly · Since Q3 2026

Generating capacity of gas-fired power plants built to supply US data centers that are running, in GW.

Covers
US gas-fired plants tied to data centers, decided project by project (unit by unit where only some of a plant's units are for data centers): classified in the plant records as serving data centers, or tied to them by a company statement or a regulator's approval.
Method
The capacity of each counted project's operating units, added up, in GW to one decimal. A project's units are counted under their own status, so one project can add to more than one of these numbers. Projects whose names suggest data center use are checked by hand before they count, so a new one can be added late. A new value is worked out each time the plant records are updated.

Gas power built for data centers, under construction

Unit: GW · Quarterly · Since Q3 2026

Generating capacity of gas-fired power plants for US data centers that is under construction, in GW.

Covers
US gas-fired plants tied to data centers, decided project by project (unit by unit where only some of a plant's units are for data centers): classified in the plant records as serving data centers, or tied to them by a company statement or a regulator's approval.
Method
The capacity of each counted project's units under construction, added up, in GW to one decimal. A project's units are counted under their own status, so one project can add to more than one of these numbers. Projects whose names suggest data center use are checked by hand before they count, so a new one can be added late. A new value is worked out each time the plant records are updated.

Gas power proposed for data centers

Unit: GW · Quarterly · Since Q3 2026

Generating capacity of gas-fired power plants proposed for US data centers and not yet under construction, in GW.

Covers
US gas-fired plants tied to data centers, decided project by project (unit by unit where only some of a plant's units are for data centers): classified in the plant records as serving data centers, or tied to them by a company statement or a regulator's approval. Proposed means announced or in pre-construction; cancelled and shelved units are not counted.
Method
The capacity of each counted project's announced and pre-construction units, added up, in GW to one decimal. A project's units are counted under their own status, so one project can add to more than one of these numbers. Projects whose names suggest data center use are checked by hand before they count, so a new one can be added late. A new value is worked out each time the plant records are updated.

Gas power projects for data centers, operating or in development

Unit: projects · Quarterly · Since Q3 2026

The number of gas-fired power projects for US data centers that are operating, under construction or proposed.

Covers
US gas-fired plants tied to data centers, decided project by project (unit by unit where only some of a plant's units are for data centers): classified in the plant records as serving data centers, or tied to them by a company statement or a regulator's approval. Projects whose units are all cancelled, shelved, mothballed or retired are not counted.
Method
A count of the projects with at least one unit operating, under construction, in pre-construction or announced; a project counts once whatever its number of units. Projects whose names suggest data center use are checked by hand before they count, so a new one can be added late. A new value is worked out each time the plant records are updated.

Output of running reactors contracted by data center companies

Unit: GW · Daily · Since 15 Oct 2026

The output of nuclear reactors already running that data center companies have contracted to buy, in GW.

Covers
US deals with a named data center buyer, a named plant or developer and a stated size. Non-binding frameworks, a buyer's own targets and plants with no data center buyer signed are left out.
Method
The megawatts of every deal of this kind, added up, in GW to two decimals, from a list of deals WattSaid keeps by hand from public announcements. A deal counts from its announcement, whatever year power is due to start, so this is contracted capacity, not power delivered. Where one agreement also covers an uprate, that part is counted with restarts and uprates. The number changes only when a deal is added or revised.

Nuclear restarts and uprates contracted by data center companies

Unit: GW · Daily · Since 15 Oct 2026

Nuclear capacity that data center companies have contracted from closed reactors being brought back and from added output at running reactors, in GW.

Covers
US deals with a named data center buyer, a named plant or developer and a stated size. Non-binding frameworks, a buyer's own targets and plants with no data center buyer signed are left out.
Method
The megawatts of every restart and uprate deal, added up, in GW to two decimals, from a list of deals WattSaid keeps by hand from public announcements. A deal counts from its announcement, whatever year power is due to start, so this is contracted capacity, not power delivered. Where one agreement covers both running output and an uprate, only the uprate is counted here. The number changes only when a deal is added or revised.

New reactors contracted by data center companies

Unit: GW · Daily · Since 15 Oct 2026

The capacity of nuclear reactors not yet built that data center companies have contracted for, in GW.

Covers
US deals with a named data center buyer, a named plant or developer and a stated size, including agreements to fund or develop reactors with an option to buy their power. Non-binding frameworks, a buyer's own targets and plants with no data center buyer signed are left out.
Method
The megawatts of every deal for new reactors, added up, in GW to two decimals, from a list of deals WattSaid keeps by hand from public announcements. A deal counts from its announcement, whatever year power is due to start, so this is contracted capacity, not power delivered. The number changes only when a deal is added or revised.

US data center pipeline against its usual

Unit: index · Daily · Since 5 Nov 2026 · In the API as dc-pipeline-index

How fast the US data center pipeline is moving this week against its previous four weeks. 100 is usual.

Covers
US data centers planned and under construction, data centers with an air permit record, and reviewed data center projects with onsite power.
Method
Each input's average over the last 7 days is divided by its median over the 28 days before; the index is 100 times the average of those ratios, with equal weight for each input available that day. At least two inputs a day.

Power added at the largest AI data centers, last 30 days

Unit: GW · Daily · Since 14 Nov 2026

Power capacity added at the largest AI data centers worldwide over the last 30 days.

Covers
The largest AI data centers worldwide, built and under construction.
Method
Their total power on the day minus their total power 30 days before.

US data center construction spending, growth on a year before

Unit: % · Monthly · Since Oct 2017

How much more was spent building US data centers than in the same month a year before.

Covers
US data center construction, private, at an annual rate, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

Physical AI

Open jobs at 43 robotics companies

Unit: jobs · Daily · Since 21 Oct 2026 · In the API as jobs-robots

How many jobs 43 robotics companies have open on the day.

Covers
Open postings at 43 companies that build robots, the AI that runs them or robot software, from humanoids to warehouse, farm, construction and delivery robots. Defence and drone makers are not in the set. They are not part of the 30 AI infrastructure companies. The set was widened from 10 companies on 17 November 2026; earlier days count those 10.
Method
The postings each company lists on the job board it publishes, counted once a day and added up. A day is given only when every one of the 43 boards was read. Postings are counted as listed, every role and place, so one role posted for several places can count more than once.

Open jobs at 18 self-driving companies

Unit: jobs · Daily · Since 21 Oct 2026 · In the API as jobs-self-driving

How many jobs 18 self-driving companies have open on the day.

Covers
Open postings at 18 companies developing self-driving cars, robotaxis, trucks, yard and mine haulage, pod transit and the software that drives them. They are not part of the 30 AI infrastructure companies. The set was widened from 8 companies on 17 November 2026; earlier days count those 8.
Method
The postings each company lists on the job board it publishes, counted once a day and added up. A day is given only when every one of the 18 boards was read. Postings are counted as listed, every role and place, so one role posted for several places can count more than once.

Crashes reported to the federal safety regulator by companies running self-driving cars, month

Unit: crashes · Monthly · Since Jun 2025

The number of crashes involving self-driving cars that their operators reported to the federal safety regulator, by the month of the crash.

Covers
The US, every crash involving a vehicle driven by an automated driving system that its operator must report, whoever caused it, from June 2025, when the current reporting rules took effect.
Method
Each report is counted once, under the month of the crash, however often it is amended. Reports are due by the middle of the month after a crash, so a month is shown only once reports filed in a later month have appeared, and late reports can still add to it. Reports are filed whoever was at fault, so a count is not a measure of fault.

Driverless robotaxi trips in California, all companies, month

Unit: trips · Monthly · Since Apr 2025

The number of passenger trips robotaxis with no safety driver made in California in a month, all companies together.

Covers
California, passenger trips in vehicles with no safety driver, as the companies report them to the state. Trips with a safety driver are not counted, and no company is shown on its own.
Method
The trips each company reports for the month, added up. A month is shown only when every company's driverless report for it has been read; a company that reported before and is missing holds the month back, unless it stays absent for three later months. Reports come in by quarter, so a month appears only after its quarter's reports are in.

Driverless robotaxi miles with passengers in California, all companies, quarter

Unit: million miles · Quarterly · Since Q1 2025

How many miles robotaxis with no safety driver drove in California with passengers on board in a calendar quarter, all companies together, in millions.

Covers
California, miles driven with a passenger on board in vehicles with no safety driver, as the companies report them to the state. Empty miles between trips and miles with a safety driver are not counted, and no company is shown on its own.
Method
Each company's monthly miles are added up by calendar quarter, or its quarter total is used where it reports only that. A quarter is shown only when every company with driverless service has given its miles for the whole quarter, in million miles to two decimals.

Companies authorized to run driverless vehicles in Texas

Unit: companies · Daily · Since 31 Oct 2026

The number of companies authorized to run vehicles with no human driver in Texas.

Covers
Texas, companies whose authorization is active in the state's public register. Companies with any other status, restricted included, are not counted.
Method
A count of the active authorizations in the register, read in full once a day.

Driverless vehicles registered in Texas by authorized companies

Unit: vehicles · Daily · Since 31 Oct 2026

The number of vehicles that companies authorized to run without a human driver have registered in Texas.

Covers
Texas, vehicles registered by companies whose authorization is active in the state's public register. Vehicles of companies with a restricted or other status are not counted.
Method
Each authorized company's registered vehicles, added up from the register, read in full once a day. This counts vehicles registered, not vehicles on the road that day.

Licence changes on open robotics models, last 30 days

Unit: changes · Daily · Since 23 Oct 2026 · In the API as robot-licence-changes

How many licence changes were seen on open robotics models in the last 30 days.

Covers
Open models tagged for robotics on a public model hub: the licence each declares, for every such model, and the contents of the model card and licence files for the models in WattSaid's Physical AI catalog and the 100 most downloaded robotics models.
Method
Each day's record of every model is compared with the day before's. A change counts when the declared licence changes, or when a tracked licence file is added, removed or edited, even if the licence's name stays the same; a change to access rules alone does not count. The total is the changes seen in the 30 days to the day, the day included, so a model that changes twice counts twice; the earliest values cover fewer than 30 days.

Open robotics models fine-tuned from SmolVLA

Unit: models · Daily · Since 22 Oct 2026

The number of open robotics models that declare SmolVLA as the model they were fine-tuned from.

Covers
Open models tagged for robotics on a public model hub that name the published SmolVLA base checkpoint as their base, as a full fine-tune or an adapter. Fine-tunes not tagged for robotics, or that do not name their base, are not counted.
Method
A count of the models listed on the day that declare it; a model declared both ways counts once. A model that is taken down drops out, so the count can fall.

Open robotics models fine-tuned from π0.5

Unit: models · Daily · Since 22 Oct 2026

The number of open robotics models that declare π0.5 as the model they were fine-tuned from.

Covers
Open models tagged for robotics on a public model hub that name the published π0.5 base checkpoint as their base, as a full fine-tune or an adapter. Fine-tunes not tagged for robotics, that do not name their base, or that start from another copy of π0.5, are not counted.
Method
A count of the models listed on the day that declare it; a model declared both ways counts once. A model that is taken down drops out, so the count can fall.

Open robotics models fine-tuned from GR00T N1.5

Unit: models · Daily · Since 22 Oct 2026

The number of open robotics models that declare GR00T N1.5 as the model they were fine-tuned from.

Covers
Open models tagged for robotics on a public model hub that name the published GR00T N1.5 checkpoint as their base, as a full fine-tune or an adapter. Fine-tunes not tagged for robotics, or that do not name their base, are not counted.
Method
A count of the models listed on the day that declare it; a model declared both ways counts once. A model that is taken down drops out, so the count can fall.

Open robotics models fine-tuned from OpenVLA

Unit: models · Daily · Since 22 Oct 2026

The number of open robotics models that declare OpenVLA as the model they were fine-tuned from.

Covers
Open models tagged for robotics on a public model hub that name the published OpenVLA checkpoint as their base, as a full fine-tune or an adapter. Fine-tunes not tagged for robotics, or that do not name their base, are not counted.
Method
A count of the models listed on the day that declare it; a model declared both ways counts once. A model that is taken down drops out, so the count can fall.

Problem reports to the federal device regulator about robotic surgery systems, month

Unit: reports · Monthly · Since Jan 2025

The number of problem reports about robotic surgery systems that the federal medical device regulator received, by the month received.

Covers
The US, reports on computer-controlled surgical instruments (robotic surgery systems) from manufacturers, hospitals and the public: malfunctions, injuries and deaths, from January 2025.
Method
Each report is counted once under the month the regulator received it. The regulator's data is updated weekly, so a month is shown once the data reaches its last days. Most reports are of a malfunction, manufacturers may send some in batches, and a report does not show that the system caused the problem, so a count is not a measure of safety.

Open robotics tool use against its usual

Unit: index · Daily · Since 29 Jul 2026 · In the API as robotics-adoption-index

How much open robotics models, datasets and tools are used this week against their previous four weeks. 100 is usual.

Covers
Downloads of open robotics models, datasets and packages, new stars on open robotics projects, container image pulls and release downloads, worldwide.
Method
Each input's average over the last 7 days is divided by its median over the 28 days before; the index is 100 times the average of those ratios, with equal weight for each input available that day. At least two inputs a day.

US imports of industrial robots, growth on a year before

Unit: % · Monthly · Since Aug 2026

How much more the US imported in industrial robots than in the same month a year before.

Covers
The US, industrial robots, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

China and the US

Electricity generated, China, large plants

Unit: TWh · Yearly · Since 2001

Electricity generated in China over a year, in TWh.

Covers
China, plants of enterprises above a set size only, so smaller plants are left out. Calendar years since 2001, plus the latest 12 months both China and the US have published. The US figure covers every sector.
Method
The monthly figures are running totals from January, so the latest 12 months are this year so far, plus the whole of last year, less last year to the same month. Converted to TWh. The change compares the latest 12 months with the same 12 months a year before.

Electricity generated, US

Unit: TWh · Yearly · Since 2001

Electricity generated in the US over a year, in TWh.

Covers
The US, every sector. Calendar years since 2001, plus the latest 12 months both the US and China have published. China's figure counts only plants of enterprises above a set size.
Method
Monthly generation added up over each calendar year, or over the latest 12 months when all 12 are published, and converted to TWh. The change compares the latest 12 months with the same 12 months a year before.

Electricity used, China

Unit: TWh · Yearly · Since 2009

Electricity used in China in a calendar year, in TWh.

Covers
China, all electricity use across the economy. Calendar years since 2020, ending at the last year both countries have published. The US figure counts only electricity sold to end users.
Method
The year's total, converted to TWh. Calendar years only, with no 12-month figure. The change is on the year before.

Electricity sold to end users, US

Unit: TWh · Yearly · Since 2009

Electricity sold to end users in the US in a calendar year, in TWh.

Covers
The US, retail sales to every type of customer. Calendar years since 2009, ending at the last year both countries have published. Electricity that is not sold, such as power a site makes for itself, is not counted; China's figure counts all use.
Method
The year's total retail sales, converted to TWh. Calendar years only, with no 12-month figure. The change is on the year before.

Share from fuel-burning plants, China

Unit: % · Yearly · Since 2001

The share of China's electricity that came from fuel-burning (thermal) plants, in percent.

Covers
China, plants of enterprises above a set size. Calendar years since 2001, plus the latest 12 months both countries have published. China counts every fuel-burning plant, not only fossil-fuel ones; the US figure counts fossil fuels only.
Method
Thermal generation divided by all generation over the same 12 months, in percent, with the 12 months formed from running totals as for electricity generated. The change is in percentage points against the same 12 months a year before.

Share from fossil-fuel plants, US

Unit: % · Yearly · Since 2001

The share of US electricity that came from plants burning fossil fuels, in percent.

Covers
The US, every sector. Calendar years since 2001, plus the latest 12 months both countries have published. China's matching figure counts every fuel-burning plant, not only fossil-fuel ones.
Method
Generation from fossil fuels divided by all generation over the same 12 months, in percent. The change is in percentage points against the same 12 months a year before.

Share from wind and solar, China, large plants

Unit: % · Yearly · Since 2010

The share of China's electricity that came from wind and solar, in percent.

Covers
China, plants of enterprises above a set size only. Calendar years since 2001, plus the latest 12 months both countries have published. The US figure covers every sector.
Method
Wind and solar generation added together and divided by all generation over the same 12 months, in percent, with the 12 months formed from running totals as for electricity generated. The change is in percentage points against the same 12 months a year before.

Share from wind and solar, US

Unit: % · Yearly · Since 2001

The share of US electricity that came from wind and solar, in percent.

Covers
The US, every sector. Calendar years since 2001, plus the latest 12 months both countries have published. China's figure covers only plants of enterprises above a set size.
Method
Wind and solar generation added together and divided by all generation over the same 12 months, in percent. The change is in percentage points against the same 12 months a year before.

Coal mined, China

Unit: million tonnes · Yearly · Since 2001

Coal mined in China in a calendar year, in million tonnes.

Covers
China, raw coal from enterprises above a set size. Calendar years since 2001, ending at the last year both countries have published.
Method
The year's total output in million tonnes. Tonnes are weight, not energy content, so a tonne of coal need not hold the same energy in both countries. The change is on the year before.

Coal mined, US

Unit: million tonnes · Yearly · Since 2001

Coal mined in the US in a calendar year, in million tonnes.

Covers
The US, every coal rank and mine type. Calendar years since 2001, ending at the last year both countries have published.
Method
The year's total production, converted from short tons to metric tonnes. Tonnes are weight, not energy content. The change is on the year before.

Crude oil produced, China

Unit: million barrels a day · Yearly · Since 2001

Crude oil produced in China, as a daily average in million barrels a day.

Covers
China, enterprises above a set size. Calendar years since 2001, plus the latest 12 months both countries have published.
Method
Output in tonnes over the year or the latest 12 months (this year so far, plus last year, less last year to the same month), converted at 7.33 barrels a tonne and divided by 365 days. The 7.33 is an average: the true number of barrels in a tonne depends on the oil. The change compares the latest 12 months with the same 12 months a year before.

Crude oil produced, US

Unit: million barrels a day · Yearly · Since 2001

Crude oil produced in the US, as a daily average in million barrels a day.

Covers
The US, the whole country. Calendar years since 2001, plus the latest 12 months both countries have published.
Method
Monthly production in barrels, added up over the year or the latest 12 months when all 12 are published, and divided by 365 days. The change compares the latest 12 months with the same 12 months a year before.

Natural gas produced, China

Unit: bcm · Yearly · Since 2001

Natural gas produced in China in a year, in billion cubic meters.

Covers
China, enterprises above a set size. Calendar years since 2001, plus the latest 12 months both countries have published.
Method
Output over the year or the latest 12 months (this year so far, plus last year, less last year to the same month), in billion cubic meters. The change compares the latest 12 months with the same 12 months a year before.

Natural gas produced, US

Unit: bcm · Yearly · Since 2001

Natural gas produced in the US in a year, in billion cubic meters.

Covers
The US, dry gas, the whole country. Calendar years since 2001, plus the latest 12 months both countries have published.
Method
Monthly dry gas production, added up over the year or the latest 12 months when all 12 are published, and converted from cubic feet to billion cubic meters. The change compares the latest 12 months with the same 12 months a year before.

Solar capacity, China

Unit: GW · Monthly · Since Apr 2021

Solar generating capacity installed in China, in GW.

Covers
China, the installed total for the whole country at each month, with no lower size limit stated. The US figure counts only plants of 1 MW or more.
Method
The installed total for each month, converted to GW. When a month is reported more than once, the latest report is used.

Solar capacity, US plants of 1 MW or more

Unit: GW · Monthly · Since Jul 2026

Solar generating capacity in operation in the US, in GW.

Covers
The US, solar generators of 1 MW or more in operation, monthly, so most rooftop systems are not counted.
Method
The nameplate capacity of every solar generator in operation in the month, added up and converted to GW. A point is added each time a new month is published.

Wind capacity, China

Unit: GW · Monthly · Since Apr 2021

Wind generating capacity installed in China, in GW.

Covers
China, the installed total for the whole country at each month. The US figure counts only generators of 1 MW or more.
Method
The installed total for each month, converted to GW. When a month is reported more than once, the latest report is used.

Wind capacity, US

Unit: GW · Monthly · Since Jul 2026

Wind generating capacity in operation in the US, in GW.

Covers
The US, wind generators of 1 MW or more in operation, monthly.
Method
The nameplate capacity of every wind generator in operation in the month, added up and converted to GW. A point is added each time a new month is published.

New-type storage added this year, China

Unit: GW · Monthly · Since Jun 2023

New-type energy storage added in China since January, in GW of power.

Covers
China, nationwide, since 2023. New-type storage means batteries and other storage except pumped hydro; the US figure counts batteries only. China's national total is published two to four times a year (the end of a quarter, usually June, September and December), so the line has those points only, and starts again each January.
Method
WattSaid's subtraction: China's new-type storage built and running at the end of the period, minus the same total at the end of the year before, in GW of power, not GWh of energy. Since it is a total for the year so far, no change is shown.

Batteries added this year, US

Unit: GW · Monthly · Since Jul 2026

Battery capacity that started operating in the US this calendar year, in GW of power.

Covers
The US, batteries of 1 MW or more in operation in the newest month, monthly. China's matching figure counts all new-type storage, not only batteries.
Method
The nameplate power of every battery in operation in the month whose first month of operation falls in the same year, added up and converted to GW, not GWh of energy. Since it is a total for the year so far, no change is shown.

Electricity growth at computing centres above the whole economy, China

Unit: points · Monthly · Since Aug 2026

How much faster electricity use at China's computing centres grew than the whole economy's, on the same month a year before, in percentage points.

Covers
China, computing centres nationwide and every sector of the economy, monthly, from August 2026. There is no US counterpart.
Method
Computing centres' growth on the same month a year before, less the whole economy's growth for that month, both from official monthly electricity-use statistics, in percentage points to one decimal. Above zero, computing centres grew faster.

Compute in known AI supercomputers, China

Unit: thousand H100 equivalents · Monthly · Since Oct 2026

The computing power of China's known AI supercomputers, in thousands of H100 equivalents.

Covers
Known AI supercomputers listed for China, monthly. Any that have not been reported are missing.
Method
Each cluster's computing power is given as the number of H100 chips that would match it, and the country's clusters are added up, in thousands. A cluster that has been expanded and listed again counts once, at its newest size. A month holds its latest reading.

Compute in known AI supercomputers, US

Unit: thousand H100 equivalents · Monthly · Since Oct 2026

The computing power of the US's known AI supercomputers, in thousands of H100 equivalents.

Covers
Known AI supercomputers listed for the US, monthly. Any that have not been reported are missing.
Method
Each cluster's computing power is given as the number of H100 chips that would match it, and the country's clusters are added up, in thousands. A cluster that has been expanded and listed again counts once, at its newest size. A month holds its latest reading.

Generative AI services filed with the national regulator, China, total

Unit: services · Monthly · Since Jun 2025

How many generative AI services have been filed with China's national regulator, in total.

Covers
China, services that completed filing, and since August 2026 on-device models on phones. Services that were registered rather than filed are not counted. A point comes with each release of the lists, about every two months.
Method
Each release adds rows to lists numbered on from the first release, so a list's highest number so far is how many it holds; the lists' totals are added up. A release whose rows cannot be placed in a list adds no point of its own.

Flagship AI model API price, input, China

Unit: $ per million tokens · Daily · Since 31 Oct 2026

The average list price of input to five Chinese flagship AI models through their APIs, in US dollars per million tokens.

Covers
Five Chinese vendors' most capable general models, named in the tile's note, daily; three models before 1 November 2026. When a vendor releases a new flagship it replaces the old one, which can move the line.
Method
A simple average of the five models' published list prices, each converted from yuan to dollars at the exchange rate of that day or the latest day before it. One rate is fixed for each model so the line stays like for like: input not served from a cache, the daytime rate where the price changes with the time of day, the real-time rate rather than batch, and one region and request-size tier. If any model's price is missing, no new point is added that day.

Flagship AI model API price, input, US

Unit: $ per million tokens · Daily · Since 31 Oct 2026

The average list price of input to five US flagship AI models through their APIs, in US dollars per million tokens.

Covers
Five US vendors' most capable general models, named in the tile's note, daily; three models before 1 November 2026. When a vendor releases a new flagship it replaces the old one, which can move the line.
Method
A simple average of the five models' standard input list prices, in dollars. If any model's price is missing, no new point is added that day.

Flagship AI model API price, output, China

Unit: $ per million tokens · Daily · Since 31 Oct 2026

The average list price of output from five Chinese flagship AI models through their APIs, in US dollars per million tokens.

Covers
Five Chinese vendors' most capable general models, named in the tile's note, daily; three models before 1 November 2026. When a vendor releases a new flagship it replaces the old one, which can move the line.
Method
A simple average of the five models' published list prices, each converted from yuan to dollars at the exchange rate of that day or the latest day before it. One rate is fixed for each model so the line stays like for like: the daytime rate where the price changes with the time of day, the real-time rate rather than batch, and one region and request-size tier. If any model's price is missing, no new point is added that day.

Flagship AI model API price, output, US

Unit: $ per million tokens · Daily · Since 31 Oct 2026

The average list price of output from five US flagship AI models through their APIs, in US dollars per million tokens.

Covers
Five US vendors' most capable general models, named in the tile's note, daily; three models before 1 November 2026. When a vendor releases a new flagship it replaces the old one, which can move the line.
Method
A simple average of the five models' standard output list prices, in dollars. If any model's price is missing, no new point is added that day.

Industrial robots made, China, larger manufacturers

Unit: thousand robots · Yearly · Since 2016

Industrial robots made in China over a year, in thousands.

Covers
China, robots made by manufacturers above a set size, some of them sold abroad. Calendar years since 2016, plus the latest 12 months both countries have published. The US publishes no count of robots made, so its matching figure is robots imported. Neither counts robots installed.
Method
The monthly figures are running totals from January, so the latest 12 months are this year so far, plus last year, less last year to the same month. Divided by 1,000. The change compares the latest 12 months with the same 12 months a year before.

Industrial robots imported, US

Unit: thousand robots · Yearly · Since 2016

Industrial robots imported into the US over a year, in thousands.

Covers
The US, industrial robots entering the country, counted in units. Calendar years since 2016, plus the latest 12 months both countries have published. It is not a count of robots made or installed; China's matching figure is robots made.
Method
Monthly imports in number of robots under the trade codes for industrial robots: one code, and since 2022 a second one, after a revision of the international product codes moved robots that lift and handle there. Added up over the year or the latest 12 months when all 12 are published, and divided by 1,000. The latest 13 months are taken again on every run, as recent months are revised. The change compares the latest 12 months with the same 12 months a year before.

Power

Lower 48 power from natural gas

Unit: % · Hourly · Since Jan 2019

The share of electricity generated in the Lower 48 states that came from natural gas, hour by hour, in percent.

Covers
The Lower 48 states together, all generation reported for each hour.
Method
Generation from natural gas divided by generation from all fuels in the same hour, in percent, to one decimal. An hour is given once every fuel that reports in at least half of the recent hours is in, so the newest hours arrive about a day late.

Lower 48 power from wind and solar

Unit: % · Hourly · Since Jan 2019

The share of electricity generated in the Lower 48 states that came from wind and solar, hour by hour, in percent.

Covers
The Lower 48 states together, all generation reported for each hour.
Method
Generation from wind and solar plants, those paired with a battery included, divided by generation from all fuels in the same hour, in percent, to one decimal. An hour is given once every fuel that reports in at least half of the recent hours is in, so the newest hours arrive about a day late.

Texas household power, median 12-month fixed offer

Unit: ¢ per kWh · Daily · Since 13 Oct 2026

The middle price of the 12-month fixed-rate electricity plans offered to Texas households on the day, for a home using 1,000 kWh a month.

Covers
Texas, retail electricity plans on offer that day: 12-month fixed-rate plans only; prepaid and time-of-use plans are left out.
Method
The median of each qualifying offer's price per kWh at 1,000 kWh a month, in cents to two decimals. Every offer counts once, however many customers take it, and only English-language listings are used.

Projects waiting to connect to PJM's grid

Unit: projects · Weekly · Since 13 Oct 2026

The number of power projects waiting to connect to PJM's grid.

Covers
PJM's public interconnection queue: projects whose status is active. Projects with any other status are not counted.
Method
A count of the active projects in a copy of the queue taken each Monday, one value a week.

Capacity waiting to connect to PJM's grid

Unit: GW · Weekly · Since 13 Oct 2026

How much generating capacity the projects waiting to connect to PJM's grid would add, in GW.

Covers
PJM's public interconnection queue: projects whose status is active. Projects with any other status are not counted.
Method
The megawatts of energy each active project asks to deliver, added up and shown in GW to one decimal, from a copy of the queue taken each Monday. This is capacity requested, not capacity that will be built.

Gas-fired capacity waiting to connect to PJM's grid

Unit: GW · Weekly · Since 13 Oct 2026

How much gas-fired generating capacity the projects waiting to connect to PJM's grid would add, in GW.

Covers
PJM's public interconnection queue: active projects whose fuel includes natural gas, hybrids with another fuel included.
Method
The megawatts of energy each of those projects asks to deliver, added up and shown in GW to one decimal, from a copy of the queue taken each Monday. A project with natural gas among its fuels counts in full. This is capacity requested, not capacity that will be built.

Average power level of US nuclear reactors

Unit: % of full power · Daily · Since 9 Jul 2026

The average power level of US nuclear reactors on the day, as a percent of full power.

Covers
Every US power reactor in the daily reactor status reports, each counted equally whatever its size. A reactor that is shut down, for refueling or any other reason, counts as 0.
Method
The simple average of each reactor's reported percent of full power for the day, to one decimal. Because every reactor weighs the same, this is not the share of the fleet's capacity that is running.

Lower 48 power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in Lower 48 in the latest hours against the normal for the same hour, in percent.

Covers
All 48 contiguous states, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

PJM power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in PJM in the latest hours against the normal for the same hour, in percent.

Covers
13 eastern states and DC, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

ERCOT power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in ERCOT in the latest hours against the normal for the same hour, in percent.

Covers
Most of Texas, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

MISO power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in MISO in the latest hours against the normal for the same hour, in percent.

Covers
15 Midwest and Southern states, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

SPP power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in SPP in the latest hours against the normal for the same hour, in percent.

Covers
14 central states, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

CAISO power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in CAISO in the latest hours against the normal for the same hour, in percent.

Covers
Most of California, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

NYISO power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in NYISO in the latest hours against the normal for the same hour, in percent.

Covers
New York State, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

ISO-NE power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in ISO-NE in the latest hours against the normal for the same hour, in percent.

Covers
New England, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

Northwest power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in Northwest in the latest hours against the normal for the same hour, in percent.

Covers
Northwest, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

Southwest power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in Southwest in the latest hours against the normal for the same hour, in percent.

Covers
Southwest, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

Southeast power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in Southeast in the latest hours against the normal for the same hour, in percent.

Covers
Southeast, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

Florida power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in Florida in the latest hours against the normal for the same hour, in percent.

Covers
Florida, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

Carolinas power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in Carolinas in the latest hours against the normal for the same hour, in percent.

Covers
Carolinas, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

Tennessee power demand against normal

Unit: % · Hourly · Since Jan 2022

Power drawn in Tennessee in the latest hours against the normal for the same hour, in percent.

Covers
Tennessee, every hour.
Method
The normal is the median demand at the same hour, on the same kind of day (weekday or weekend), within 14 days either side of the date, in each of the three previous years. The indicator is actual demand over the normal, minus one, in percent.

Household power price in Virginia against the US average

Unit: % · Monthly · Since Jan 2024

How much more (or less) households in Virginia pay for power than the US average, in percent.

Covers
Virginia and the US, residential customers, monthly.
Method
The average residential price in the state divided by the US average for the same month, minus one, in percent.

Electricity sold to businesses in Virginia, growth on a year before

Unit: % · Monthly · Since Jan 2025

How much more electricity businesses in Virginia bought than in the same month a year before.

Covers
Virginia, commercial customers, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

Household power price in Texas against the US average

Unit: % · Monthly · Since Jan 2024

How much more (or less) households in Texas pay for power than the US average, in percent.

Covers
Texas and the US, residential customers, monthly.
Method
The average residential price in the state divided by the US average for the same month, minus one, in percent.

Electricity sold to businesses in Texas, growth on a year before

Unit: % · Monthly · Since Jan 2025

How much more electricity businesses in Texas bought than in the same month a year before.

Covers
Texas, commercial customers, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

Household power price in Ohio against the US average

Unit: % · Monthly · Since Jan 2024

How much more (or less) households in Ohio pay for power than the US average, in percent.

Covers
Ohio and the US, residential customers, monthly.
Method
The average residential price in the state divided by the US average for the same month, minus one, in percent.

Electricity sold to businesses in Ohio, growth on a year before

Unit: % · Monthly · Since Jan 2025

How much more electricity businesses in Ohio bought than in the same month a year before.

Covers
Ohio, commercial customers, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

Household power price in Georgia against the US average

Unit: % · Monthly · Since Jan 2024

How much more (or less) households in Georgia pay for power than the US average, in percent.

Covers
Georgia and the US, residential customers, monthly.
Method
The average residential price in the state divided by the US average for the same month, minus one, in percent.

Electricity sold to businesses in Georgia, growth on a year before

Unit: % · Monthly · Since Jan 2025

How much more electricity businesses in Georgia bought than in the same month a year before.

Covers
Georgia, commercial customers, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

Household power price in Pennsylvania against the US average

Unit: % · Monthly · Since Jan 2024

How much more (or less) households in Pennsylvania pay for power than the US average, in percent.

Covers
Pennsylvania and the US, residential customers, monthly.
Method
The average residential price in the state divided by the US average for the same month, minus one, in percent.

Electricity sold to businesses in Pennsylvania, growth on a year before

Unit: % · Monthly · Since Jan 2025

How much more electricity businesses in Pennsylvania bought than in the same month a year before.

Covers
Pennsylvania, commercial customers, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

Household power price in Arizona against the US average

Unit: % · Monthly · Since Jan 2024

How much more (or less) households in Arizona pay for power than the US average, in percent.

Covers
Arizona and the US, residential customers, monthly.
Method
The average residential price in the state divided by the US average for the same month, minus one, in percent.

Electricity sold to businesses in Arizona, growth on a year before

Unit: % · Monthly · Since Jan 2025

How much more electricity businesses in Arizona bought than in the same month a year before.

Covers
Arizona, commercial customers, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

Gas and oil

Gas scheduled to LNG export terminals, on 12 pipelines

Unit: million MMBtu a day · Daily · Since 15 Jul 2026

How much gas the pipelines have scheduled to the points that feed US LNG export terminals, in million MMBtu a day.

Covers
Eight LNG export terminals on the Gulf Coast and in Georgia, as fed by 12 interstate pipelines. Gas that reaches them on other pipelines, and terminals elsewhere, are not counted. Scheduled, not metered.
Method
The pipeline operators' scheduled quantities at the points that feed the terminals, added up for each gas day (9 a.m. to 9 a.m. Central) at the latest scheduling cycle. A point that hands gas to another of the 12 pipelines on its way to a terminal is left out, so the same gas is not counted twice. A day is given only when all 12 pipelines' figures for it are in, and the newest days can change as later cycles come in.

Gas scheduled for delivery on 12 US pipelines

Unit: million MMBtu a day · Daily · Since 15 Jul 2026

How much gas is scheduled for delivery off 12 US interstate pipelines, in million MMBtu a day.

Covers
Every delivery point on 12 interstate pipelines that serve Virginia, Texas and the LNG export terminals on the Gulf Coast and in Georgia. Scheduled, not metered.
Method
The pipeline operators' scheduled quantities at every delivery point, added up for each gas day (9 a.m. to 9 a.m. Central) at the latest scheduling cycle. A delivery that hands gas to another of the 12 is left out, so gas passed between them counts once, where it leaves the group; deliveries into other pipelines and into storage count. A day is given only when all 12 pipelines' figures for it are in, and the newest days can change as later cycles come in.

US gas in storage against the 5-year average

Unit: % · Weekly · Since 2 Oct 2026

How much more (or less) gas the US has in storage than the five-year average for the same week, in percent.

Covers
The Lower 48 states together, weekly.
Method
The week's storage level divided by the five-year average for the same week, as given with the weekly storage figures, minus one, in percent, to one decimal. One value a week, dated by the week the level is for.

Compute

Training GPUs, cloud spot price

Unit: $ per GPU-hour · Daily · Since 15 Jul 2026

The average spot price of one hour of a training GPU in the cloud, in dollars.

Covers
Four server types with eight training GPUs each, Linux, in every zone of five US and Canadian regions of one large cloud provider.
Method
For each server type, a zone whose price changed that day takes the average of that day's prices, and a zone that posted no change keeps its last price for up to 14 days. The zones are averaged, the result is divided by the server's eight GPUs, and the four types are averaged with equal weight. A day without a price for all four types is left out.

Inference GPUs, cloud spot price

Unit: $ per GPU-hour · Daily · Since 15 Jul 2026

The average spot price of one hour of an inference GPU in the cloud, in dollars.

Covers
Four server types with one inference GPU each, Linux, in every zone of five US and Canadian regions of one large cloud provider.
Method
For each server type, a zone whose price changed that day takes the average of that day's prices, and a zone that posted no change keeps its last price for up to 14 days. The zones are averaged and the four types are averaged with equal weight. A day without a price for all four types is left out.

Capital spending of Microsoft, Alphabet, Amazon, Meta and Oracle, calendar quarter

Unit: $ billion · Quarterly · Since Q3 2017 · In the API as capex-five

What Microsoft, Alphabet, Amazon, Meta and Oracle together spent on property and equipment in a calendar quarter, in billions of dollars.

Covers
The five companies' whole spending on property and equipment, worldwide and not only on AI or data centers, as each reports it in its cash flow statement.
Method
Each company's quarter is its three-month figure where one is reported, and otherwise its year to date less its year to date a quarter before; a revised figure replaces the earlier one. A fiscal quarter is placed in the calendar quarter of its last month (Oracle's quarter to August counts as the quarter to September), and a quarter is given only when all five have reported it, to $0.1 billion.

Capital spending of Microsoft, Alphabet, Amazon, Meta and Oracle, growth on a year before

Unit: % · Quarterly · Since Q3 2018 · In the API as capex-five-growth

How much more the five largest cloud and AI spenders invested than in the same quarter a year before.

Covers
Microsoft, Alphabet, Amazon, Meta and Oracle, calendar quarters.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

US imports of computers and parts, growth on a year before

Unit: % · Monthly · Since Aug 2026

How much more the US imported in computers and computer parts than in the same month a year before.

Covers
The US, computers and parts, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

Prices of data processing and hosting, change on a year before

Unit: % · Monthly · Since Jan 2025

How much the prices US providers charge for data processing and hosting changed on a year before.

Covers
The US, data processing and hosting services, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

Open AI model use against its usual

Unit: index · Daily · Since 2 Nov 2026 · In the API as ai-adoption-index

How much open AI models and datasets are used this week against their previous four weeks. 100 is usual.

Covers
Downloads of open AI models and datasets, and tokens sent to AI models through a large model router, worldwide.
Method
Each input's average over the last 7 days is divided by its median over the 28 days before; the index is 100 times the average of those ratios, with equal weight for each input available that day. At least two inputs a day.

Hiring

Open jobs at 30 AI infrastructure companies

Unit: jobs · Daily · Since 13 Oct 2026 · In the API as jobs-all

How many jobs 30 AI infrastructure companies have open on the day.

Covers
Open postings at 30 companies that build, power and use AI compute: 13 compute and data center companies, 6 AI labs, 6 AI chip makers and 5 new energy companies. Robot and self-driving companies are counted separately.
Method
The postings each company lists on the job board it publishes, counted once a day and added up. A day is given only when all 30 boards were read, so a board that could not be read never looks like jobs closing. Postings are counted as listed, every role and place, so one role posted for several places can count more than once.

Open jobs at compute and data center companies

Unit: jobs · Daily · Since 13 Oct 2026 · In the API as jobs-compute

How many jobs 13 compute and data center companies have open on the day.

Covers
Open postings at 13 companies that rent out AI compute, run AI models for customers or build data centers.
Method
The postings each company lists on the job board it publishes, counted once a day and added up. A day is given only when every one of the 13 boards was read. Postings are counted as listed, every role and place, so one role posted for several places can count more than once.

Open jobs at AI labs

Unit: jobs · Daily · Since 13 Oct 2026 · In the API as jobs-labs

How many jobs 6 AI labs have open on the day.

Covers
Open postings at 6 companies that build AI models or the data and products around them.
Method
The postings each company lists on the job board it publishes, counted once a day and added up. A day is given only when every one of the 6 boards was read. Postings are counted as listed, every role and place, so one role posted for several places can count more than once.

Open jobs at AI chip makers

Unit: jobs · Daily · Since 13 Oct 2026 · In the API as jobs-chips

How many jobs 6 AI chip makers have open on the day.

Covers
Open postings at 6 companies that design chips for AI. The largest established chip makers are not in the set.
Method
The postings each company lists on the job board it publishes, counted once a day and added up. A day is given only when every one of the 6 boards was read. Postings are counted as listed, every role and place, so one role posted for several places can count more than once.

Open jobs at new energy companies

Unit: jobs · Daily · Since 13 Oct 2026 · In the API as jobs-energy

How many jobs 5 new energy companies have open on the day.

Covers
Open postings at 5 companies developing advanced nuclear power, fusion and batteries.
Method
The postings each company lists on the job board it publishes, counted once a day and added up. A day is given only when every one of the 5 boards was read. Postings are counted as listed, every role and place, so one role posted for several places can count more than once.

Hiring at AI infrastructure companies against its usual

Unit: index · Daily · Since 2 Nov 2026 · In the API as hiring-momentum

How many jobs 30 AI infrastructure companies have open this week against their previous four weeks. 100 is usual.

Covers
Open jobs at 30 companies that build, power and use AI compute.
Method
Each input's average over the last 7 days is divided by its median over the 28 days before; the index is 100 times the average of those ratios, with equal weight for each input available that day.

US jobs in data hosting, growth on a year before

Unit: % · Monthly · Since Jan 2025

How many more people work in US data hosting and computing infrastructure than a year before.

Covers
The US, data processing, hosting and related services, monthly.
Method
The value for each period divided by the value for the same period a year before, minus one, in percent.

Buzz

AI: attention against its usual

Unit: index · Daily · Since 19 Oct 2026 · In the API as attention-ai

Attention to AI this week against its own previous four weeks. 100 is usual; 150 is half as much again.

Covers
New code repositories, new research papers and reference page views about AI, worldwide, in English.
Method
Each input's average over the last 7 days is divided by its median over the 28 days before; the index is 100 times the average of those ratios, with equal weight for each input available that day (at least two). Completed UTC days only.

AI agents: attention against its usual

Unit: index · Daily · Since 19 Oct 2026 · In the API as attention-ai-agent

Attention to AI agents this week against its own previous four weeks. 100 is usual; 150 is half as much again.

Covers
New code repositories, new research papers and reference page views about AI agents, worldwide, in English.
Method
Each input's average over the last 7 days is divided by its median over the 28 days before; the index is 100 times the average of those ratios, with equal weight for each input available that day (at least two). Completed UTC days only.

GPUs: attention against its usual

Unit: index · Daily · Since 19 Oct 2026 · In the API as attention-gpu

Attention to GPUs this week against its own previous four weeks. 100 is usual; 150 is half as much again.

Covers
New code repositories, new research papers and reference page views about GPUs, worldwide, in English.
Method
Each input's average over the last 7 days is divided by its median over the 28 days before; the index is 100 times the average of those ratios, with equal weight for each input available that day (at least two). Completed UTC days only.

Data centers: attention against its usual

Unit: index · Daily · Since 19 Oct 2026 · In the API as attention-data-center

Attention to data centers this week against its own previous four weeks. 100 is usual; 150 is half as much again.

Covers
New code repositories, new research papers and reference page views about data centers, worldwide, in English.
Method
Each input's average over the last 7 days is divided by its median over the 28 days before; the index is 100 times the average of those ratios, with equal weight for each input available that day (at least two). Completed UTC days only.

Robots: attention against its usual

Unit: index · Daily · Since 5 Nov 2026 · In the API as attention-robotics

Attention to robots this week against its own previous four weeks. 100 is usual; 150 is half as much again.

Covers
New code repositories, new research papers and reference page views about robots, worldwide, in English.
Method
Each input's average over the last 7 days is divided by its median over the 28 days before; the index is 100 times the average of those ratios, with equal weight for each input available that day (at least two). Completed UTC days only.