Skip to content

Charts

19 charts Newest first
Robotics

Black Rice Was Harder for This Sushi Robot to Shape

Black Rice Was Harder for This Sushi Robot to Shape

Rice portions judged acceptable or better · One robot hand · 20 fresh attempts with each kind of rice

Each circle is one attempt. Filled circles met the panel’s threshold for shape.

Source: Hu, Kogashi and Koike-Akino · MERL workshop report, Table 2 ·
Sources, data and detail

What the figures show

A sushi chef can switch ingredients and adjust by feel. What happens when a robot hand that has practiced shaping rice meets a different variety? Researchers gave TacSushi one job with two kinds of rice. It had to squeeze a portion into an oval that could hold its shape. The hand had cameras and touch sensors at its fingertips for both tests.

With the rice it had practiced on, 15 of its 20 attempts passed. With black rice, an ingredient the team reserved for testing, seven passed. Both counts come from the full model in the researchers’ November 13 workshop report. The broader experiment also included wrapping seaweed and placing toppings.

What counted as a good rice ball? Five people and three AI models scored a photo of each finished portion. An average of at least three on a five-point scale counted as acceptable. This captures what a diner could see on the plate. Taste and texture would require a different test. With just 20 attempts per ingredient, each portion moves the score by five percentage points.

Data and method

20 fresh rice-shaping attempts per ingredient with the full TacSushi model.

Rice portions judged acceptable or better
RiceSuccessful attempts
Familiar rice15 / 20
Black rice7 / 20

Five people and three AI models rated the final image from each attempt. The chart counts portions that averaged at least 3 out of 5.

Sources

Compute

Who Gets There Before the Ambulance?

Who Gets There Before the Ambulance?

Missions where a trained volunteer arrived before the ambulance · Freiburg region, Germany · Suspected cardiac arrests

The 2.5 km radius and travel-time alerts performed similarly in the trial.

Source: Ganter et al. · ALARM randomized trial ·
Sources, data and detail

What the figures show

When a dispatcher in the Freiburg region of Germany suspects a cardiac arrest, a smartphone alert can send trained volunteers to the scene while the ambulance is on its way. A local trial asked which volunteers should get the call. The answer depended partly on how wide a circle the system drew around the patient.

On days when alerts went to volunteers within 600 metres, someone reached the scene before the ambulance on 31% of missions. A radius of 2.5 kilometres brought that figure to 57%. Selecting volunteers by estimated travel time produced a similar result at 58%. The study could distinguish both approaches from the smaller radius. It could not reliably separate the two higher rates.

The researchers randomly assigned a strategy each day and tracked arrivals by phone GPS. The trial covered 823 alerts for suspected cardiac arrest from January through July 2024. It measured who arrived first. Whether any strategy improved survival remains an open question. The travel-time system also spared some volunteers a trip. It stood down 31% of those who accepted an alert because the ambulance was expected to arrive first.

Data and method

Suspected cardiac-arrest missions assigned to each alert strategy in the Freiburg region.

Missions where a trained volunteer arrived before the ambulance
Alert strategyShare of missions
Within 600 m31%
Within 2.5 km57%
By estimated travel time58%

823 app activations across all three strategies, January–July 2024. GPS recorded volunteer arrivals. The study measured arrival order.

Sources

AI

Where AI Users Say They Save Time at Work

Where AI Users Say They Save Time at Work

Frequent workplace AI users who say they spend less time on the task · U.S. employees using AI at least a few times a week · August 2026

People could report savings on several tasks. The bars count people, not hours.

Source: Gallup · U.S. employee survey, August 2026 ·
Sources, data and detail

What the figures show

When Gallup asked regular AI users where they spent less time at work, finding information topped the list. In Gallup’s survey, 52% said they spent less time searching for or gathering information since they started using AI. Writing or drafting came next at 47%, followed by analyzing data at 38%. For manual or administrative work, the figure was 29%.

Meetings tell a different story. Among those same frequent users, 83% said they spent about as much time in meetings as before. That helps explain what a faster first draft may feel like in practice. A document can take less of someone’s morning while their calendar stays just as full.

Gallup surveyed 2,648 U.S. employees from August 5 to 19, 2026, and published the findings on November 16. The four bars show answers from the subset who used AI at least a few times a week. Someone who saved five minutes and someone who saved an hour both count once. The survey therefore tells us where people noticed a change and leaves the size of that change open.

Data and method

U.S. employees who use AI at work at least a few times a week.

Frequent workplace AI users who say they spend less time on the task
TaskShare of frequent users
Finding information52%
Writing or drafting47%
Analyzing data38%
Manual or administrative work29%

Gallup surveyed 2,648 U.S. employees overall, August 5–19, 2026. Bars show self-reported changes among frequent AI users. Each person could report changes on several tasks.

Sources

Longevity

Global Lifespan Recovered More Than Healthy Life Expectancy

Global Lifespan Recovered More Than Healthy Life Expectancy

WHO modeled estimates · World, both sexes · 2000–2023

Expected years at birth. Shaded bands show the original uncertainty ranges.

━ Life expectancy┄ Healthy life expectancy
04080200020102023

Healthy life expectancy weights years by health and disability severity, rather than counting only illness-free years.

YearLifeHealthy life
201973.463.2
202171.561.4
202373.362.8

By 2023, the point estimates were 0.1 and 0.4 years below 2019. This comparison does not establish a statistically significant difference in recovery.

Source: WHO, Global Health Estimates 2000–2023 · Chart: VastkindSource revision: October 1, 2026 · Accessed November 7, 2026

Longer life and better health do not necessarily recover together. In WHO’s estimates, global life expectancy was almost back to its 2019 level by 2023. Healthy life expectancy remained further below its earlier level. Both measures had risen over the longer period shown.

The distinction is about population estimates, not the age at which one person becomes ill. A year lived with impaired health contributes less than a full healthy-equivalent year to the second measure.

More Detail, Sources and Data

Two Measures, Not Two Stages of a Life

Period life expectancy applies the mortality conditions of a given year across a hypothetical lifetime. It is not a prediction of the actual lifespan of babies born that year. Healthy life expectancy, or HALE, adjusts years lived for the severity of health loss. The gap between the measures is not a block of healthy years followed by a block of sick years.

Read the Estimates With Their Uncertainty

The lines use the same zero-to-80-year scale and show every annual point from 2000 through 2023. The bands preserve WHO’s published lower and upper bounds for each year. For 2023, life expectancy is 73.3 years [70.7–75.8] and healthy life expectancy is 62.8 years [57.9–67.4]. These are marginal uncertainty ranges, not uncertainty intervals for a change between years.

Vastkind compares the published point estimates. It does not reconstruct WHO’s population model or test whether the two recovery paths differ statistically. Overlapping ranges do not prove equality, and the point differences do not prove a significant difference. The chart does not identify the cause of the changes or measure the effect of any treatment.

Global Expected Years at Birth, Both Sexes
YearLife Expectancy [Range]Healthy Life Expectancy [Range]
200067.1 [65.4–68.6]57.9 [54.0–61.4]
200167.4 [65.8–69.0]58.2 [54.3–61.7]
200267.7 [66.1–69.3]58.4 [54.6–62.0]
200368.0 [66.4–69.6]58.7 [54.9–62.3]
200468.4 [66.8–70.0]59.0 [55.1–62.6]
200568.9 [67.3–70.4]59.4 [55.5–63.0]
200669.4 [67.9–70.9]59.9 [56.0–63.4]
200769.9 [68.3–71.4]60.2 [56.3–63.8]
200870.2 [68.6–71.7]60.5 [56.6–64.1]
200970.6 [69.1–72.1]60.9 [56.9–64.4]
201071.0 [69.5–72.4]61.1 [57.2–64.7]
201171.3 [69.8–72.7]61.4 [57.5–65.0]
201271.7 [70.3–73.1]61.7 [57.8–65.3]
201372.1 [70.7–73.4]62.0 [58.1–65.6]
201472.3 [71.0–73.6]62.2 [58.3–65.7]
201572.5 [71.1–73.8]62.3 [58.4–65.9]
201672.7 [71.3–74.0]62.5 [58.6–66.1]
201772.9 [71.6–74.2]62.7 [58.8–66.3]
201873.2 [71.9–74.4]63.0 [59.1–66.5]
201973.4 [72.1–74.5]63.2 [59.3–66.6]
202072.7 [71.7–73.7]62.4 [58.7–65.8]
202171.5 [70.2–72.7]61.4 [57.5–64.9]
202272.7 [70.5–74.8]62.3 [57.7–66.5]
202373.3 [70.7–75.8]62.8 [57.9–67.4]

Original Data and Reuse

WHO, Global Health Estimates 2000–2023; Global Health Observatory indicators WHOSIS_000001 (life expectancy at birth) and WHOSIS_000002 (healthy life expectancy at birth), GLOBAL, both sexes. Both series carry the October 1, 2026 revision stamp; accessed November 7, 2026. We acknowledge the countries supplying the underlying national data, as described in WHO’s methods linked from the source page. These are global estimates, not an unweighted average of country values.

All 24 annual values and their original bounds are displayed for each indicator. Differences of 0.1 and 0.4 years are simple subtraction of the published 2019 and 2023 point estimates. The data period ends in 2023, not in the year of this chart’s publication.

Copyright remains with WHO. This editorial public-health display follows the WHO dataset terms; no unrestricted open-data license or WHO endorsement is claimed. No WHO emblem is used.

Energy

Batteries Supplied 15% of California’s Evening Grid Demand

Batteries Supplied 15% of California’s Evening Grid Demand

2025 reported averages · Report-defined western regions

Share of each region’s electricity demand during hours ending 18–22 (five hourly intervals).

California15%
Desert Southwest11%
Intermountain West3%
Pacific Northwest1%
0%5%10%15%20%

Operator-reported battery schedules, not a guarantee of supply on every evening. Percentages use different regional demand totals.

Source: California ISO, Department of Market Monitoring · Chart: Vastkind2025 data · Report dated October 28, 2026; posted October 30

Batteries already supply a meaningful part of evening electricity demand in parts of the western grid. The California group has the largest reported share in this comparison, followed by the Desert Southwest. These are shares of the electricity each region needs, not rankings of battery size or total electricity delivered.

That distinction matters: installing a battery and using it to meet demand are different measures. This comparison captures batteries’ reported role during the evening window, when California ISO says net demand is highest. It does not show where their charging electricity came from.

More Detail, Sources and Data

What the Percentages Measure

California ISO’s Department of Market Monitoring reports these four averages together on pages 3 and 16 of its 2025 Special Report on Battery Storage. Each value expresses reported battery supply as a share of the load, or electricity demand, of that report-defined region during hours ending 18 through 22 in 2025. The source discusses the figures in its five-minute-market battery-schedules section. These are published operator aggregates, not a new Vastkind calculation from raw meter readings.

The window contains five hourly intervals, not four hours from 18:00 to 22:00. The share passage does not specify a clock timezone or its precise averaging weights, so we retain the source’s hour-ending labels rather than translate them into local-clock times.

Reported Average Battery Share of Regional Demand, 2025
Report RegionShare
California15%
Desert Southwest11%
Intermountain West3%
Pacific Northwest1%

Comparable Shares, Different Regional Totals

All bars start at zero and use the same 0–20% scale. Vastkind transcribed the published whole-percentage values, checked both source passages and drew this comparison. No additional precision is inferred. The shares cannot be summed into a western-grid total or averaged without the underlying regional demand totals.

A larger share does not necessarily mean more electricity delivered in absolute terms. The comparison does not establish lower bills, avoided emissions, policy effects or reliable supply on every evening. The report discusses hybrid resources separately; the share passage does not establish complete coverage of every hybrid battery.

The Report’s Four Regional Groups

These are the report’s balancing-area groups on page 9, not a new Vastkind grouping of entire states. A balancing area is a part of the power system managed to keep electricity supply and demand in balance.

  • California: California ISO; Balancing Authority of Northern California; Los Angeles Department of Water and Power; Turlock Irrigation District.
  • Desert Southwest: Arizona Public Service; El Paso Electric; NV Energy; Public Service Company of New Mexico; Salt River Project; Tucson Electric; WAPA-Desert Southwest.
  • Intermountain West: Avista Corporation; Idaho Power Company; NorthWestern Energy; PacifiCorp East.
  • Pacific Northwest: Avangrid Power; Bonneville Power Administration; PacifiCorp West; Portland General Electric; Powerex; Puget Sound Energy; Seattle City Light; Tacoma Power.

Source, Dates and Reuse

California ISO, Department of Market Monitoring, 2025 Special Report on Battery Storage, pages 3 and 16 (values), page 9 (regions). The document is dated October 28, 2026; its official posting notice is dated October 30. The measurements describe 2025, not today’s grid.

Original factual comparison by Vastkind. Source: California Independent System Operator; copyright and other proprietary rights remain with their respective owners. Reuse follows CAISO’s public-information terms; no Creative Commons or public-domain status is claimed. No CAISO artwork or logo is copied.

Compute

Servers Take 56% of This AI Data Center’s Upfront Budget

Servers Take 56% of This AI Data Center’s Upfront Budget

Hypothetical U.S. site · 1 GW of IT capacity · May 2026 model

Share of an estimated $37.9 billion initial investment. GB200 NVL72 hardware, not an actual project’s bill.

Servers55.9%
Facility & Cooling30.2%
Networking & Other IT13.0%
Land0.5%
Utility Works0.4%
0%25%50%75%100%

Initial purchases and construction only, not annual operating costs. Hardware and construction inputs include 2025 estimates.

Source: Epoch AI cost model · Calculation and chart: VastkindPublished May 14, 2026 · Upfront data updated May 19, 2026

The computers cost more than everything else put together. In Epoch AI’s hypothetical U.S. AI data center, servers account for about $21.2 billion of the $37.9 billion initial investment. The facility, cooling, other IT equipment, land and utility works make up the rest.

This explains why a data center’s headline construction budget may describe only part of the investment. A building and its cooling systems are not the same purchase as a fully equipped AI computing site.

This is a model, not an average of real projects or a current price quote. It uses NVIDIA GB200 NVL72 systems and one gigawatt of rated IT capacity. That describes the computing equipment’s power rating, not the site’s electricity use over a year.

Updated explainer, November 3: For the separate question of energy needed to run a model, our Extropic analysis explains why a projected chip-energy saving is not yet a whole-task result.

More Detail, Sources and Data

What the Budget Includes

“Facility & Cooling” is Epoch’s facility category, including its liquid-cooling cost adjustment. “Networking & Other IT” is the source’s network-infrastructure category: a proxy for the cluster’s non-server investment, including networking, storage, service, software and installation. Land and utility works are separate inputs. A small budget share does not establish that land or a grid connection is easy to obtain.

Modeled Initial Investment
ComponentUSD MillionsShare
Servers21,18855.9%
Facility & Cooling11,43330.2%
Networking & Other IT4,92513.0%
Land1720.5%
Utility Works1640.4%
Reported Total37,883100%

How We Calculated the Shares

Vastkind divided each component in Epoch’s upfront-investment CSV by its reported total of $37,883 million, then multiplied by 100. Bars start at zero and use unrounded shares; labels show one decimal place. The title rounds the server share, 55.93%, to 56%.

The individually rounded components add to $37,882 million, $1 million below the rounded total. The underlying model workbook totals approximately $37,882.5 million. This rounding difference does not change the displayed shares.

A Model of Purchases, Not Annual Spending

Epoch published the analysis on May 14, 2026; its upfront chart data were updated May 19. Hardware and construction assumptions include 2025 estimates. The model uses GB200 NVL72 systems and a hypothetical U.S. hyperscale site with 1 GW of rated IT capacity, powered entirely from the grid. It is not a November 2026 supplier quote, a measured industry average or a forecast for every future AI facility.

These are initial capital costs. Electricity bills and other operating expenses belong to a different comparison. Epoch’s separate annualized-cost analysis accounts for recurring costs and different asset lifetimes; its percentages must not be mixed with these upfront shares.

For the separate question of what a large chip count can enable, see our explanation of Colossus 2’s expansion plans. This chart does not estimate that project’s budget.

Data and model: Amelia Michael and Ben Cottier, Epoch AI, AI Data Center Cost Breakdown, reused with attribution under CC BY 4.0. Vastkind calculated the shares, clarified category labels and created this chart; no source graphic was copied.

Energy

Seven in Ten Grid-Connection Requests Were Withdrawn

Seven in Ten Grid-Connection Requests Were Withdrawn

U.S. power-plant and storage grid requests · Submitted 2000–2020 · Status at end of 2025

Share of 21,927 requests in the report’s outcome sample. Counts, not generating capacity.

Withdrawn70.8%

15,515 requests

Operating19.1%

4,192 requests

Still active8.8%

1,921 requests

Suspended1.4%

299 requests

0%25%50%75%100%

Suspended means on hold. Withdrawn does not mean rejected by a grid operator; the data do not assign a reason for each withdrawal.

Source: Berkeley Lab and GridTracker, Queued Up 2026 · Original chart: Vastkind · CC BY 4.0 dataRequests submitted 2000–2020 · Status at end of 2025 · 2026 edition

Most requests in this older cohort did not become operating power resources. By the end of 2025, 15,515 had been withdrawn, while 4,192 were operating. Another 2,220 were still active or suspended, even though every request had entered a queue at least five years earlier.

A grid queue records developers’ plans to connect equipment, not a stock of completed plants. These outcomes show why a large queue cannot be read as a promise of future electricity supply. They do not show how many withdrawals were caused by grid delays, financing, permitting or a change of plan.

MISO’s current proposal concerns a different side of the grid: connecting large electricity users. This chart instead follows generation and storage requests, using outcomes recorded before that October 2026 dispute.

More Detail, Sources and Data

Requests Are Not the Same as Plants or Megawatts

The denominator is the 21,927 requests in the report’s published outcome analysis, submitted from 2000 through 2020 inclusive. The report selects regions with comprehensive status information; this is not a census of every U.S. connection application. Load requests, distribution-connected systems and equipment behind a customer’s meter are outside this dataset’s scope.

Each request has equal weight. A large plant counts the same as a small one. The operating share is 19.1% by request count, not the report’s separate 13% capacity-weighted result. Active and suspended requests are not treated as permanent failures; this is a status snapshot, not a prediction for a new application.

Status at End of 2025 for Requests Submitted in 2000–2020
StatusRequestsShare
Withdrawn15,51570.8%
Operating4,19219.1%
Still active1,9218.8%
Suspended2991.4%
Total21,927100%

Source, Calculation and a Conflicting Footnote

Vastkind summed the 21 annual rows in the original workbook, sheet “23. Completion Rate Trend,” cells A28:E48. Each share equals its status count divided by 21,927, multiplied by 100. Bars use unrounded shares; labels use one decimal place, so displayed shares add to 100.1% after rounding.

The workbook tables and the figure on page 35 of the 2026 report support 21,927 requests. Footnote 5 on that page instead says 22,061. The source does not explain the difference. This chart uses the reproducible table total, not the conflicting footnote.

The publication catalogue and workbook are dated May 2026; the report cover is dated June 2026. The outcome snapshot ends in 2025. Data attribution: Lawrence Berkeley National Laboratory and GridTracker/Interconnection.fyi, CC BY 4.0. Aggregation and visualization by Vastkind; no source illustration was copied.

Biotech

Enzyme-Loaded Films Lost 12% to 100% of Their Mass on Average

Enzyme-Loaded Films Lost 12% to 100% of Their Mass on Average

Mass lost as % of initial film mass · 150 days · 757 m depth · 4.4°C · Misaki, Japan, 2023

Each dot is one film sample. Vertical marks show the mean of four samples per group. Added enzyme: 3% of film weight.

PBS

Polybutylene succinate

No added enzyme1.2% mean
With enzyme97.3% mean

PBSA

Polybutylene succinate-co-adipate

No added enzyme3.4% mean
With enzyme100.0% mean

PCL

Polycaprolactone

No added enzyme2.5% mean
With enzyme100.0% mean

PBAT

Polybutylene adipate-co-terephthalate

No added enzyme1.1% mean
With enzyme12.1% mean

Thin films, not ordinary discarded objects. Mass loss does not establish the fate of all the carbon or rule out small fragments.

Source: Huang et al., Nature, Figure 3a source data · Original chart: Vastkind · CC BY 4.0Published October 21, 2026 · Measurements June–November 2023

The added enzyme made a large difference for three of the four plastics: their films lost 97–100% of their mass, compared with 1–3% without the enzyme. PBAT responded much less, losing about 12% with enzyme versus 1% without it. The same treatment did not make every material disappear at the same rate.

These were films about 100–300 micrometers thick, exposed in the same cold deep-sea setting. The result is a materials comparison, not evidence that existing ocean plastic can be cleaned up this way. Adding enzyme also reduced some films’ strength in laboratory tests. Read the full explanation of the experiment and its limits.

More Detail, Sources and Data

Four Materials, the Same Treatment

Field exposure ran June 6–November 3, 2023, at the BMS5-2 site off Misaki, Japan: 150 days, 757 meters deep, at 4.4°C. The added enzyme was a cutinase from the fungus Humicola insolens, loaded at 3% of film weight. Each plotted group contains four film samples.

All four materials with both an enzyme-treated and untreated comparison are shown. Two additional materials without an enzyme-treated comparison are not included. Each pair uses the same zero-to-100% scale. Dots are separated vertically only to keep equal values visible; their horizontal positions are the actual measurements.

Mass Lost After 150 Days, as % of Initial Film Mass
Material and treatmentMeanFour individual samples (%)n
PBS, no added enzyme1.2%0.81, 1.48, 1.00, 1.594
PBS, with enzyme97.3%94.68, 100.00, 99.12, 95.494
PBSA, no added enzyme3.4%2.86, 3.74, 3.79, 3.044
PBSA, with enzyme100.0%100.00, 100.00, 100.00, 100.004
PCL, no added enzyme2.5%2.50, 2.16, 2.81, 2.674
PCL, with enzyme100.0%100.00, 100.00, 100.00, 100.004
PBAT, no added enzyme1.1%1.18, 1.10, 1.07, 1.134
PBAT, with enzyme12.1%10.30, 8.71, 15.58, 13.994

Source and Calculation

Huang et al., Nature, published October 21, 2026, Figure 3a source workbook and Supplementary Tables 8 and 13. Vastkind recalculated each arithmetic mean from the four original observations and rounded displayed means to one decimal place. No experiments were rerun. The raw points are shown instead of inferred error bars: the supplementary PBS summary’s reported spread does not match either common standard-deviation calculation from the workbook.

Source data reused under CC BY 4.0. This is an original Vastkind visualization, not a reproduction of the authors’ figure. The chart measures film mass loss, not complete conversion to carbon dioxide, harmless residues or ecosystem safety.

Robotics

Braking Guidance Cut Simulated Car Contacts From 13 to Zero

Braking guidance reduced contacts in the same 24 tests.

Runs with stopped-car contact · 24 selected simulations · October 2026 preprint

Always-on guidance with perfect obstacle information. Simulations, not real-world accident rates.

Source: Cercola et al., Table I and the 24-scene analysis · Chart: VastkindPreprint submitted: 20 Oct 2026 · No road-safety claim

Recognizing a stopped car is not the same as generating a path that stops before it. The same driving model made contact in 13 of 24 selected simulations without added braking guidance, and in none with an always-on stopping constraint and exact knowledge of the obstacle’s geometry.

The intervention changed how the model generated a path, not its training or camera. It shows what a constraint can achieve under ideal obstacle information, not that a car is ready to drive more safely on public roads.

More Detail, Sources and Data

Same Model, Same Selected Scenes

The researchers selected 24 braking-required scenes from 38 dynamic scenarios. These were simulated reconstructions of recorded drives using the frozen Alpamayo-1.5 model. The added guidance was always on and used privileged information: the obstacle’s exact geometry. It was not a test of an imperfect obstacle detector or of deciding when braking guidance should activate.

Runs With Stopped-Car Contact
ConfigurationRuns with contactRuns tested
Without added braking guidance1324
Always-on guidance, exact obstacle geometry024

The bars use one zero-to-24 scale. The first bar occupies 13 ÷ 24, or 54.17%, of the track; the zero bar has zero length. These are reported test outcomes, not estimated real-world crash probabilities. No simulator was rerun by Vastkind.

A Constraint Is Not the Whole Driving System

The useful result is narrower than a safer self-driving car: under ideal obstacle information, the model could obey a stopping constraint that it otherwise failed to enforce. Obtaining reliable obstacle information and deciding when to activate that constraint remain separate problems.

Source, Version and Reuse

Cercola et al., Perceives but Doesn’t Brake, arXiv:2610.28184v1, submitted October 20, 2026, Table I and the explicit 13-of-24 comparison. This is a preprint, not a peer-reviewed road-safety validation. Vastkind’s original visualization uses the paper’s reported counts under CC BY 4.0; it does not reproduce the authors’ graphic or JavaScript. Other experiments in the paper have different conditions and are not mixed into this comparison.

Energy

Most Hybrids Sold in the U.S. Don’t Plug In

Most hybrids sold don’t plug in.

Share of all new vehicle sales · U.S. · July–September 2026

New cars and light trucks; rounded estimates.

Source: Omdia via U.S. EIA · Chart: VastkindData: Q3 2026 · Published: 19 Oct 2026

A hybrid can have a battery and an electric motor without ever using a charging cable. These no-plug hybrids accounted for 16% of new U.S. car and light-truck sales in the third quarter of 2026, compared with 0.7% for plug-in hybrids.

That is roughly 23 no-plug hybrids sold for every plug-in hybrid. Battery-only vehicles were a separate 6% of sales. The distinction matters when judging how many new vehicles will use chargers.

More Detail, Sources and Data

That distinction matters: a conventional hybrid uses a battery and an electric motor but does not charge from the electricity grid. A battery-electric vehicle does; a plug-in hybrid can use both grid electricity and fuel. Calling all three “electric vehicles” conceals very different demands on chargers and power supplies.

Which Cars Need a Charging Cable?

Which new cars can charge from the grid?

U.S. new car and light-truck sales · July–September 2026

Hybrid, no plug16%

Does not charge from the grid

Battery-only electric6%

Charges from the grid

Plug-in hybrid0.7%

Can use grid electricity and fuel

0%5%10%15%20%

Share of all new light-duty vehicles sold.

Omdia estimates, via U.S. EIA, 19 Oct 2026. Rounded shares; selected categories, not a 100% breakdown. Vastkind visualization.

Share of all new U.S. car and light-truck sales, Q3 2026
Vehicle typeSales share
Hybrid, no plug16%
Battery-only electric6%
Plug-in hybrid0.7%

The chart compares each category with all new cars and light trucks sold in the United States in Q3 2026. It is not a breakdown of only electric vehicles, or the entire fleet already on the road. Each bar starts at zero; the categories are separate, and the figures are rounded estimates.

The EIA reports a combined share of about 23%. The displayed components add to 22.7% because they are rounded at different precision. No exact remainder or unpublished prior-year hybrid value is reconstructed.

For context, the battery-only vehicle sales share was 10% in Q3 2025 and 6% a year later. That is a decline of about 4 percentage points, not proof that the number of electric cars sold fell 40%. Total vehicle sales also affect that calculation.

New sales are only one slice of the fleet. They cannot be substituted for the share of cars on the road, or for the electricity used by those cars. The separate Vastkind chart on U.S. EV electricity use measures the latter.

The clearest signal in this quarter's mix is the gap between electrification and plugging in: the largest of these three categories never needs a charging cable.

Source and method

Source: U.S. EIA, 19 October 2026, reporting Omdia estimates. Measurement period: 1 July–30 September 2026. Vastkind redrew three explicitly published aggregate percentages; this is not an independently audited Omdia microdata release. No full historical series was digitized. The graphic and accessible table use the same CSV values. EIA's published rounded numbers, not its graphic design, are reproduced with attribution. No open licence for the underlying proprietary database is asserted.

Biotech

Most Enzyme Designs Cut 10% or Less of Their Target

Most enzyme designs cut 10% or less of their target.

135 selected designs · One 12-hour lab test at 37°C · 2026

Eight had no cutting measurement: their proteins were the wrong size or would not dissolve.

Source: Chen et al. · Zenodo · Chart: VastkindStudy: 14 Oct 2026 · Screen: 135 selected designs

Scientists designed enzymes to act as molecular scissors. In a 12-hour lab test, 78 of 135 designs cut no more than 10% of the target material; 49 cut more.

Eight produced protein of the wrong size or protein that would not dissolve. They have no cutting measurement. A small amount of cutting is still activity: 10% is not a universal pass mark. This selected batch does not show how useful the enzymes would be in practice.

More Detail, Sources and Data
What happened to the 135 selected designs?12-HOUR LAB TEST AT 37°C · SELECTED DESIGNS ONLY
Selected lab-screen outcomes: 49, 78, 4 and 4 of 135 designs Selected lab-screen outcomes: 49, 78, 4 and 4 of 135 designs

49 of 135 selected enzyme designs (36.3%) crossed the study's cutoff: cutting more than 10% of the target material in a 12-hour laboratory test. Another 78 cut no more than 10%; four produced the wrong-size protein and four would not dissolve. In the chart, ‘cis’ means the enzyme and target are attached. ‘Cleavage’ means cutting, ‘insoluble’ means unable to dissolve, and ‘non-numeric’ means no cutting measurement, not zero cutting.

SOURCE: CHEN ET AL., NATURE; ORIGINAL SCREEN DATA ON ZENODO. VASTKIND RECALCULATION.Study published 14 October 2026 · Selected laboratory designs, not all computer-generated candidates

The laboratory is a different filter

The researchers designed enzymes intended to cut a protein target. In this screen, the enzyme and its target were attached to each other. After 12 hours at 37°C, the team measured how much cutting had occurred, correcting for the background signal without zinc.

The chart follows all 135 designs selected and ordered for that laboratory screen. It does not show every candidate generated on a computer. Earlier computational filters had already removed designs, and the supplementary methods say the discarded sequences were not retained.

A cutoff is not a verdict on every design

A design below 10% is not necessarily inactive. The threshold separates measured amounts of cutting; it does not turn a continuous result into a universal pass or fail. Protein concentrations were not standardized across the screen.

The two design groups also went through different selection procedures. Their combined outcome is a useful account of this selected screen, not a general AI success rate, a controlled comparison of two methods or evidence of a medical treatment.

The useful question is not just how many promising designs emerged. It is what happened to the whole selected batch when it reached the laboratory.

Sources, Method & Data

The total is 135 designs selected and ordered for testing, listed in Supplementary Table S3. Cutting was measured for 127 of them. The strict threshold is greater than 0.10, using unrounded original values. ZnO4 is printed as 0.10 in the supplementary table but recalculates to 0.1026764266, so it belongs above the cutoff. This test did not measure health effects or cutting a separate target molecule; the enzyme and its target were attached.

All 135 selected designs in the laboratory screen
Screen outcomeDesignsShare of 135
Cut more than 10%4936.3%
Cut 10% or less7857.8%
Wrong-size protein43.0%
Protein would not dissolve43.0%

Rounded percentages sum to 100.1%. Counts sum to 135. Original data: Chen et al., Zenodo record 22654831, CC BY 4.0. This is a new Vastkind visualization and recalculation, not a reproduction of a publisher figure. The software archive's MIT license is distinct from the dataset license.

Nature paper · Supplementary methods and Table S3. Public abstract, figure caption, complete supplementary methods and original numerical data checked; subscription main text not claimed as independently read.

Energy

Electric Vehicles Used Less Than 1% of U.S. Electricity

Electric vehicles use less than 1% of U.S. electricity.

Share of total electricity use · U.S. · January–June 2026

Electric cars, light trucks and plug-in hybrids; national electricity use, not the power needed locally at one moment.

Source: U.S. EIA · Chart: VastkindData: Jan–Jun 2026 · Release: September 2026

Charging electric cars and light trucks, including plug-in hybrids, used an estimated 0.67% of U.S. electricity in the first half of 2026. Out of every 100 units of electricity used nationwide, less than one went to these vehicles.

Their share of U.S. electricity more than doubled from the same months of 2023, while remaining a small national share. Local grids can still face pressure when many vehicles charge in the same place at the same time.

More Detail, Sources and Data
Growing quickly. Still below 1%.JANUARY–JUNE EACH YEAR · EIA ESTIMATES

The axis ends at 1%, not 100%. These national estimates cover battery-only and plug-in hybrid cars and light trucks, not every electric vehicle or a local grid's peak load.

Data: U.S. EIA · Monthly Energy Review, Table 7.6 · Calculation and chart: VastkindPeriod: Jan–Jun 2023–2026 · Release: September 2026 · Retrieved: 13 Oct 2026

Charging Grew Fast but Remained a Small National Share

Cars can draw more electricity without becoming a large part of the national power bill. EIA estimates that U.S. light-duty electric vehicles used 13.7 terawatt-hours in the first half of 2026, up from 5.9 TWh in the same months of 2023. Total U.S. electricity end use in the latest half-year was about 2,043.3 TWh. Divide the two and EVs account for 0.67%.

The comparison uses January through June in every year, rather than mixing winter and summer halves. It starts in 2023 because EIA changed how it estimates vehicle miles traveled that year; earlier observations are not strictly comparable.

The country is not one charging socket

A national energy total cannot show whether a neighborhood transformer can serve several new chargers, or whether charging coincides with a local evening peak. Terawatt-hours measure energy over a period; those constraints also depend on where electricity is drawn and how much power is needed at once.

The useful distinction is scale: EV charging grew rapidly from a small national base. That does not make every local connection easy, but it does keep a growing load from being mistaken for a dominant share of U.S. electricity use.

Sources, Method & Data

Calculation. We sum January–June monthly observations in EIA's September 2026 Table 7.6 and downloadable data. Share = estimated EV electricity use (series ESVHPUS) / total electricity end use (ELTCPUS) × 100. Both series are published in million kWh; 1,000 million kWh equals 1 TWh. Ratios use unrounded source values. Display rounding is not a measure of model certainty.

Coverage. The charging total estimates electricity used by on-road battery-only and plug-in hybrid cars and light trucks up to 8,500 lb gross vehicle weight, including cars and light trucks. It is not a census of charger meters and excludes heavy trucks, buses, off-road vehicles and gasoline burned by plug-in hybrids. The geography is the 50 U.S. states and District of Columbia. EIA's estimation method combines vehicle stock, travel and energy-use assumptions.

Total Electricity Use and a Change in the Estimates. Electricity end use is retail sales plus direct use, not power-generation capacity, electricity generation or total primary energy. Vehicle charging is already included in total electricity use and is not added again. This is not a claim to measure every kilowatt-hour produced and used on site by rooftop solar. EIA notes a vehicle-miles-traveled methodology change beginning in 2023, so no earlier-year growth comparison is drawn.

First-half estimates; electricity in TWh
PeriodVehicle chargingTotal end useCharging share
Jan–Jun 20235.91,899.30.31%
Jan–Jun 20248.21,967.60.42%
Jan–Jun 202510.92,021.30.54%
Jan–Jun 202613.72,043.30.67%

Source version and rights. New explanation of existing data, not an October demand reading. Latest observations: June 2026. EIA's public aggregate data are reused with attribution under its reuse policy. Original chart by Vastkind; no EIA logo or third-party microdata reproduced.

Energy

Battery Storage Nearly Doubled. Reported Incidents Barely Changed.

Battery storage nearly doubled. Reported incidents barely changed.

Global stationary battery storage · 2024 versus 2025

GWh means storage for a million kilowatt-hours. EPRI, an electricity-research institute, reports incomplete incident counts, not the risk at a particular site.

Source: EPRI · Chart: VastkindData: 2024–2025 · Report: October 2026

Installed battery storage capacity grew from 343 to 634 gigawatt-hours between 2024 and 2025. Over the same period, EPRI’s report recorded 26 incidents, then 24.

That much larger capacity makes incidents per unit of storage fall by about half, even though the incident count barely changes. It is a useful distinction when reading a falling rate; the figures do not predict the risk at a particular site.

More Detail, Sources and Data
Almost the same incident count. Much more capacity.2024 VS 2025 · EPRI OCTOBER 2026 REPORT

These are report-snapshot values, not a complete count of every battery failure. Annual events divided by installed GWh are not an individual site's failure probability. The public CSV has 23 events for 2025, versus 24 in the report; the versions are not combined.

Data: EPRI · October 2026 report, pp. 6–7 · Calculation and chart: VastkindMeasured: 2024–2025 · Report: October 2026 · Retrieved: 10 Oct 2026

Storage Grew Much Faster Than the Incident Count Changed

For a community considering a large battery installation, a falling incident rate sounds reassuring. The key is to compare the number of incidents with the amount of installed storage.

In EPRI's October 2026 report, the recorded annual count falls from 26 to 24, a decline of 7.7%. The published capacity figure rises from 343 to 634 gigawatt-hours, up 84.8%. Divide one by the other and the result falls from 7.58 to 3.79 incidents per 100 GWh, a 50.1% decline. The large change in that ratio is not a halving of the incident count.

A useful ratio is not a site-risk forecast

The report tracks major internal thermal incidents in stationary lithium-ion storage, including grid-scale, commercial and industrial installations. It is not a census of every malfunction, electric-car battery or home battery. Reporting is incomplete, and each incident counts once regardless of its severity.

The comparison divides incidents by stored-energy capacity. One gigawatt-hour (GWh) means storage for one million kilowatt-hours. It does not measure electricity generated, power at one moment or years of operation. Newly installed capacity can enter the annual stock without having operated for a full year. We also could not independently verify that the proprietary capacity series covers every installation type in the incident count.

The fair reading is narrower than “batteries are half as dangerous”: in this report, recorded incidents fell slightly while the amount of installed storage rose sharply between these two years. A decision about one installation still depends on its design, operation and consequences of a failure.

Sources, Method & Data

Snapshot and calculation. Four aggregate values come from printed pages 6–7 of EPRI report 3002034920, dated October 2026. Its PDF was created on October 8; the exact first-publication day was not established. This chart is a new explanation of 2024–2025 data, not a new incident count for October 2026. Each ratio is annual recorded events / cumulative installed capacity in GWh × 100. Changes compare 2025 with 2024. Input capacities are already rounded; two decimal places on the ratio do not imply extra measurement precision.

Source versions do not fully reconcile. The public stationary-storage CSV retrieved on October 10 carried a September 29, 2026 Last-Modified header. It contains 26 events dated 2024 and 23 dated 2025, versus 26 and 24 in the newer report. Some regional and confidential-case totals also differ. A version difference is plausible, but no case-by-case revision history resolves it. We reproduce the report aggregates, not a reconstructed current global total; no cases have been invented or added to the CSV.

What the Incident and Capacity Totals Cover. EPRI includes major thermal events caused by internal battery-system failures and excludes external causes such as wildfires. Some cases are confidential. The report's source figure cites Benchmark Mineral Intelligence's BESS Forecast Q1 2026, China Energy Storage Alliance and Korea Electrical Safety Corporation. The underlying capacity records were not independently accessible. Matching of utility, commercial/industrial and backup-power-system coverage between the incident count and the storage-capacity total remains unverified. This annual division does not account for how long each unit of storage operated or how severe each incident was. It is not the chance of an individual installation failing, or the report's separate analysis of equipment age.

EPRI October 2026 report snapshot; own division
YearRecorded incidentsCapacity (GWh)Incidents per 100 GWh
2024263437.58
2025246343.79

Graphic and rights. Own calculation and original graphic using four published aggregate facts; no EPRI figure or layout reproduced. EPRI's report is © 2026, all rights reserved, not an open-licensed dataset. The accessible CSV is linked for source-version checking, not substituted for the report.

AI

AI Solved Many Cases at Least Once. Far Fewer Every Time.

AI solved many cases at least once. Far fewer every time.

Cases solved out of 97 · Four runs each · August 2026

Simulated customer support; different settings for how much work the models could do before answering. Two examples, not a model ranking or a prediction of the next run.

Source: Sierra Research · Chart: VastkindRuns: 3–4 Aug 2026 · Submitted: 4 Aug 2026

An AI can solve a customer’s problem once and still be unreliable. Across 97 simulated banking-support cases, Claude Opus 5 solved 64 at least once in four runs. It solved only 31 on all four runs.

Qwen 3.8 Max showed the same gap: 71 cases solved at least once, but 34 every time. These results show why repeated success is a harder test than a single good demonstration.

More Detail, Sources and Data
Repeat the same case. The result can change.97 SIMULATED CASES PER MODEL · FOUR RUNS EACH

Numbers inside the bars count cases, not individual runs. “Every time” means success in all four recorded runs, not four chances to succeed once. Two examples, not a model ranking.

Data: Sierra Research · τ-knowledge · Chart: VastkindRecorded: 3–4 Aug 2026 · Submission: 4 Aug 2026 · Retrieved: 7 Oct 2026

Getting it right once is the easier bar

Imagine asking customer service to replace a lost card. A successful interaction is encouraging. The harder question is whether the same request still works when the conversation takes a different path.

In these published simulations, Claude Opus 5 succeeded at least once on 64 of 97 cases; Qwen 3.8 Max did so on 71. Require success in all four runs, and those counts fall to 31 and 34. The middle segment shows cases that sometimes worked and sometimes did not.

A Simulated Bank, Not Real Customer Service

These are recorded evaluations in a fictional banking environment, with an AI playing the customer. Vastkind grouped the original scores; we did not run the agents again or independently regrade their conversations. The chart measures neither real customer satisfaction nor financial safety.

Both agents received the same embedded case definitions and tool access, but used different settings for how much work the models could do before answering. These two examples do not establish a general model ranking or equal-cost comparison. Four runs per case are a small sample, not a guarantee about the next attempt.

One obstacle to repeating this test: 15 of the 97 embedded task definitions differ from the code revision named in the run files. They match between the two files analyzed here. The reason for the repository discrepancy is unresolved.

Sources, Method & Data

What we counted. The τ-knowledge banking_knowledge AllTools run files contain 97 case definitions and 388 runs per model, four distinct trial IDs per case. We treated a logged reward within 0.000001 of 1 as success, matching the project metric, then grouped cases by zero, one to three, or four successes. Both selected files contain no runs stopped by a technical system error. Bar widths are group counts divided by 97; labels give the counts.

Settings and timing. Qwen 3.8 Max used xhigh reasoning; Claude Opus 5 used max. Both used GPT-5.2 with low reasoning as the customer simulator and a 200-step limit. The metadata evaluation date is August 3, 2026; raw timestamps span August 3–4 without a stated time zone. The submission date is August 4; the hosted files also carry August 4 modification timestamps. This is a new explanation of that snapshot, not a new October test or a comprehensive current leaderboard.

Cases and Individual Attempts Are Different Counts. Individual-run success was 189/388 for Claude and 214/388 for Qwen. That average is not the first attempt. “At least once” counts cases with one or more successes; “every time” counts cases with four. We make no independence assumption or statistical-significance claim.

Original records. Repository snapshot; MIT code license. Own aggregation and graphic, not a copied source figure. The hosted conversation files carry no separate license notice; full dialogs are not republished here.

The Saved Tasks Differ From the Named Code Version. The run files name code commit fc0055dc4e0a316c3f83133267fbd6faaa770992. Fifteen embedded task definitions differ from that revision and from the repository snapshot above. Our counts describe the scores in the published run files. They do not establish an exact reproduction from the named code revision. Other model results were not included because their comparability was not fully verified.

Recorded outcomes: cases, not runs
Model0 of 41 of 42 of 43 of 44 of 4Total
Claude Opus 5331112103197
Qwen 3.8 Max26145183497

Related: Can AI spot an IKEA assembly mistake? A different capability test, with its own limits.

AI

AI’s Furniture-Checking Score Rose From 28% to 80%

AI’s furniture-checking score rose from 28% to 80%.

Furniture photo test score · Model releases: Nov 2025–Sep 2026

60 photos from three IKEA builds; test scores, not physical assembly or a count of mistakes found.

Source: Epoch AI · Chart: VastkindReport: 23 Sep 2026 · Models: Nov 2025–Sep 2026

Given an assembly photo, a manual and inspection tools, the AI had to find mistakes and explain them. The best score among the tested models rose from 28.3% for Claude Opus 4.5 to 80% for GPT-6 Astra.

That is a large improvement on this test’s 60 photos from three IKEA builds. It measures photo analysis, with some grading by another AI. It does not establish reliable performance on every piece of furniture.

More Detail, Sources and Data
Can AI judge an IKEA assembly photo?BEST TESTED SCORE: 28% → 80%

The best test score rose from about 28% to 80%. The line tracks each new highest score; gray dots show other tested models. “Benchmark” means this specific test of 60 photos from three furniture builds, with a manual and tools. Past model results compared; no physical assembly was tested.

Data: Aiden Ament & Greg Burnham · Epoch AI · CC BY 4.0 · Chart: VastkindReport: 23 Sep 2026 · Model releases: Nov 2025–Sep 2026 · Accessed: 6 Oct 2026

Not a promise to fix your furniture

The sample covers only 60 photos from three builds and is partly graded by another AI. It does not establish reliability on other furniture, real-time repairs or physical assembly. The score is not simply the proportion of flawed builds flagged, and the tested models are not a complete historical leaderboard.

Sources, Method & Data

The 21 source rows are ordered by model release date as reported by Epoch, not test date. The line retains each new record among those models; it is not a complete history of all available AI. GPT-5 and Gemini 3 are absent. Published scores are plotted unchanged ×100. The endpoints are 28.3333% and 80%, a gain of 51.6667 percentage points. Source standard errors are 5.8665 and 4.9289 percentage points; all model standard errors are in the table below. A standard error estimates statistical uncertainty in a score; it is not a 95% confidence range or a guarantee about new photos.

The dataset has 42 intentionally flawed and 18 correct photos from STÄLL, TONSTAD and GULLABERG builds. Models receive instructions and image/Python tools. Correct step identification and an adequate error description matter; another AI grades the descriptions with lenient scoring rules. The score is not simply the fraction of flawed builds flagged. Some exported scores imply half cases when multiplied by 60; the public method does not explain that granularity, so no exact “x out of 60” count is inferred. No human-performance comparison. Accessed October 6, 2026.

How the test is scored · Original chart dataset

All 21 tested models; original Epoch scores
ModelRelease date (Epoch)Score (%)Score uncertainty (percentage points)
Claude Opus 4.52025-11-2428.33335.8665
GPT-5.22025-12-1138.33336.3298
Claude Opus 4.62026-02-0528.33335.8665
Gemini 3.1 Pro2026-02-1926.66675.6332
GPT-5.42026-03-0537.50006.2465
Claude Opus 4.72026-04-1633.33336.1372
Kimi K2.62026-04-2021.66675.2301
GPT-5.52026-04-2344.16676.4101
Claude Opus 4.82026-05-2842.50006.3807
Claude Fable 52026-06-0935.83336.1859
GPT-5.6 Sol2026-07-0956.66676.3409
GPT-5.6 Terra2026-07-0954.16676.4321
GPT-5.6 Luna2026-07-0942.50006.3807
Kimi K32026-07-1634.16676.1170
Gemini 3.6 Flash2026-07-2123.33335.3766
Claude Opus 52026-07-2460.83336.1859
Gemini 3.7 Flash2026-08-1326.66675.6332
Claude Fable 5.12026-09-0170.00005.7244
Qwen3.8 Max2026-09-0120.00005.2076
Gemini 3.8 Flash2026-09-0231.66675.9383
GPT-6 Astra2026-09-0380.00004.9289
Energy

China’s Wind and Solar Growth Beat the U.S., EU and India Combined

China’s wind and solar growth beat the U.S., EU and India combined.

Increase in annual electricity generation · 2020–2025 · TWh

TWh means a billion kilowatt-hours. This is the change in yearly electricity output, not five years added together or installed capacity. The 2025 estimates may change.

Source: Ember · Chart: VastkindData: 2020–2025 · Report: 21 Apr 2026

China’s annual wind and solar electricity generation grew by about 1,583 terawatt-hours from 2020 to 2025. The United States, European Union and India together added about 874 TWh—just over half as much.

This compares the increase in electricity produced in a year. China’s coal generation also rose over the period, so rapid growth in wind and solar did not mean fossil-fuel output was falling.

More Detail, Sources and Data
1. Energy for Power, Transport, Heating and Industry2025 · EJ, A UNIT OF ENERGY

Power, transport, heat and industrial uses. Fossil means oil, gas and coal; non-fossil means nuclear and renewables. EJ is a unit of energy. The Energy Institute does not fully count traditional fuels such as wood burned outside commercial markets. More supply is not automatically greater efficiency or independence.

Data: Energy Institute · Statistical Review 2026 · Chart: VastkindData: 2025 · Comparison: 2020–2025 · Released: 30 Jun 2026
2. What is driving power growth?CHANGE IN YEARLY ELECTRICITY OUTPUT · 2020 → 2025 · TWh

+ more / − less annual generation. China adds 1,583 TWh of wind and solar output; the other three together add 874 TWh. Other includes energy from biomass, other renewables and other fossil sources. “Net change” is the overall increase after adding gains and losses. TWh is a billion kilowatt-hours.

Data: Ember · Global Electricity Review 2026 · CC BY 4.0 · Chart: Vastkind2025 generation estimates may be revised · Report: 21 Apr 2026 · Accessed: 6 Oct 2026

Different sources of growth

The U.S. adds solar and gas. India adds coal. In the EU, wind and solar rise as other output falls. China’s wind and solar gain is larger than the other three combined, but its coal generation grows too.

The comparison concerns changes in annual output, not the cumulative electricity generated over five years.

Energy is bigger than electricity

The first panel includes fuel use beyond power stations. The second examines electricity generation only. The two panels use different measures and cannot be added. Higher energy supply alone is not a ranking of technology, efficiency or energy independence.

2020 was a pandemic year. These are 2020–2025 comparisons, not forecasts or claims about 2026 growth.

Sources, Method & Data

Panel 1 measures energy supply in exajoules (EJ); one EJ is a billion billion joules. It uses the Energy Institute’s Statistical Review 2026 and its published accounting method. Fossil = oil + gas + coal. Non-fossil = nuclear + hydro + solar + wind + other renewables, as covered by EI; commercially recorded energy including modern renewables, not complete coverage of traditional/noncommercial biomass. Energy supply is not the energy finally delivered as useful work or heat, nor a measure of domestic production or independence. Wind/solar enter as electricity output; thermal sources include conversion losses, so source shares are not equivalent to electricity shares. China means mainland China; U.S. territories are excluded. The official Total EU aggregate covers the current 27 members. 2025 annual values; 2020–2025 change is a ratio of annual totals, not EI’s leap-year-adjusted growth series. Numbers rounded for display. Direct EI downloads were blocked; unchanged EI CSV/XLSX snapshots archived by Our World in Data were used, with matching source checksums.

Panel 2 uses Ember’s separate dataset for electricity produced before power stations deduct their own use. A terawatt-hour (TWh) is a billion kilowatt-hours, the energy unit used on electricity bills. 2025 values are estimates based on monthly generation data and may be revised. Each cell is annual 2025 output minus annual 2020 output, in TWh, not cumulative generation over five years. Other = bioenergy + other renewables + other fossil. The net row is the total-generation difference; displayed rounded cells may not add exactly. An output increase is not the same as installed capacity: weather, plant use and outages matter. Renewables and thermal energy have different conversion losses. The four regions are selected comparisons, not a world total or a country ranking of technological quality.

China’s wind/solar change: 913.83 + 668.95 = 1,582.78 TWh. Other three: 384.55 + 303.08 + 186.63 = 874.26 TWh. EU net output grows 43.80 TWh; wind/solar grows 303.08 while fossil output falls 200.09. Sources published June 30 and April 21, 2026; accessed October 6, 2026.

EI data and methodology · Ember source and methods

2025 energy supply and change since 2020
Region2025 EJFossil share %2020 EJChange %
China162.19387.55135.448+19.75
USA93.82983.1786.185+8.87
EU2752.19373.2053.458-2.37
India39.09992.9930.055+30.09
Energy supply by source, 2025 (EJ)
RegionOilGasCoalNuclearHydroSolarWindOther renewables
China33.84515.91092.2395.2935.0124.2244.0611.611
USA36.42432.8848.7309.0120.8791.4141.6892.797
EU2721.76611.8524.5897.1031.1551.3401.7282.660
India10.8922.39623.0710.5870.6420.6070.3740.530
Electricity change: annual 2025 minus annual 2020 (TWh)
SourceChinaU.S.EU-27India
Solar+913.83+258.10+222.44+143.08
Wind+668.95+126.45+80.64+43.55
Coal+835.09-36.24-95.54+330.37
Gas+81.92+183.17-93.24-26.25
Nuclear+121.98-5.10-31.10+9.22
Hydro+77.57-38.25-20.72+14.03
Other+113.60-11.52-18.68+4.17
Compute

Training Llama 3.1 Used 81 Million Times as Much Computing as AlexNet

Training Llama 3.1 used 81 million times as much computing as AlexNet.

Two AI models · Estimated training computation · 2012 versus 2024

Two selected models with different tasks. More computing work does not mean the same increase in intelligence.

Source: Epoch AI · Chart: VastkindModels: 2012 and 2024 · Estimates

Training Llama 3.1 405B in 2024 used roughly 81 million times the computing work estimated for AlexNet in 2012. AlexNet helped machines recognise images; Llama is a language model.

The comparison makes the change in training scale tangible. It does not measure an equivalent change in intelligence: these are two different kinds of model, and both computation figures are estimates.

More Detail, Sources and Data
Computing Work Used to Train Four AI ModelsCALCULATION COUNT · EACH STEP UP IS 100×
10¹⁸10²⁰10²²10²⁴10²⁶2012201620202024AlexNet: 4.7e+17 FLOPAlexNetGPT-3 175B: 3.14e+23 FLOPGPT-3 175BPaLM 540B: 2.5272e+24 FLOPPaLM 540BLlama 3.1-405B: 3.8e+25 FLOPLlama 3.1-405BYear the Model Was Published10¹⁸10²⁰10²²10²⁴10²⁶2012201620202024AlexNet: 4.7e+17 FLOPAlexNetGPT-3 175B: 3.14e+23 FLOPGPT-3 175BPaLM 540B: 2.5272e+24 FLOPPaLM 540BLlama 3.1-405B: 3.8e+25 FLOPLlama 3.1-405BYear the Model Was Published

Selected historical models · estimates. Each vertical step is 100×.

Data: Epoch AI · Chart: VastkindModels: 2012–2024 · Dataset Accessed: 5 Oct 2026

What It Shows

The full chart places GPT-3 and PaLM between the two endpoints. It compares the work needed to train four selected models, not how much each model can do.

What It Does Not Show

Four selected historical models are not a complete trend or a ranking of today’s frontier. Epoch classifies all four records as “Confident”; its guidance still allows roughly a factor of three of uncertainty. This label covers the least certain populated quantity, not a separately measured error bar for each compute value.

Sources, Method & Data

FLOP counts arithmetic operations, the small calculations used in training. Points use the dataset’s publication dates and estimates. The vertical scale uses equal spacing for equal multipliers: each labeled step up is 100 times as much work. No fitted trend line is shown. Selection and visualization by Vastkind.

Data: Epoch AI, Data on AI Models, accessed 5 October 2026. CC BY 4.0. Dataset Documentation · Original CSV.

Values Used in This Chart
ModelPublication DateTraining calculations (FLOP)
AlexNet2012-09-304.7e+17
GPT-3 175B2020-05-283.14e+23
PaLM 540B2022-04-042.5272e+24
Llama 3.1-405B2024-07-233.8e+25
Energy

Onshore Wind Cost a Quarter Less Than Large-Scale Solar

Onshore wind cost a quarter less than large-scale solar.

Global average generation cost · New projects · 2025 USD/MWh

MWh means 1,000 kilowatt-hours. These are lifetime generation costs, not bills; grids and storage add costs.

Source: IRENA · Chart: VastkindData: 2025 · Report: July 2026

New onshore wind projects in 2025 averaged $33 to generate a megawatt-hour of electricity over their lifetimes. Large-scale solar averaged $44. Using those global averages, wind cost a quarter less.

The figures help compare the cost of producing electricity. A household’s bill also pays for networks and other costs; the wider system may need storage or backup when wind and sunlight are unavailable.

More Detail, Sources and Data
Average Cost of Producing Electricity2025 US DOLLARS PER MWh · NEW PROJECTS
0255075100125Onshore Wind33Solar panels44Hydropower62Offshore Wind78Energy from biomass86Earth’s heat89Solar mirrors1150255075100125Onshore Wind33Solar panels44Hydropower62Offshore Wind78Energy from biomass86Earth’s heat89Solar mirrors115

Average cost over a project’s life. Different technologies supply power at different times.

Data: IRENA · Chart: VastkindData: 2025 · Report: Jul 2026

What It Shows

Onshore wind and utility-scale solar had the lowest global average generation costs among these renewable technologies in 2025. Their position is striking, but an average is not a quote for a specific project.

What It Does Not Show

This measure spreads a project’s costs over its expected lifetime electricity output. One megawatt-hour (MWh) is 1,000 kilowatt-hours, the unit used on electricity bills. It is neither a household tariff nor a measure of the full electricity system, including storage and grid needs. Geography, financing and when power is available all matter.

Sources, Method & Data

Values from Figure S1, page 6, of IRENA’s Renewable Power Generation Costs in 2025 (2026). Units are 2025 US dollars per megawatt-hour. Bars begin at zero and are sorted from lowest to highest. “Solar mirrors” uses mirrors to focus sunlight and produce heat for electricity. Data © IRENA 2026; selected and visualized by Vastkind.

Values Used in This Chart
Technology2025 USD/MWh
Onshore Wind33
Solar panels44
Hydropower62
Offshore Wind78
Energy from biomass86
Earth’s heat89
Solar mirrors115
Robotics

China Installed More Factory Robots Than the Rest of the World

China installed more factory robots than the rest of the world.

Share of new industrial robot installations · Worldwide · 2025

Installations by country; includes robots made abroad.

Source: IFR · Chart: VastkindData: 2025 · Report: September 2026

China installed more factory robots in 2025 than all other countries combined: 59% of new installations, against 41% elsewhere. This shows where new factory automation is concentrated, not where the robots were made or how many each worker has.

More Detail, Sources and Data
Six industrial robot installation marketsNEW INSTALLATIONS · 2025 · CHANGE FROM 2024

China: 59% of global installations. “YoY” means change from the previous year; “k” means thousand. EU-27 means the European Union’s 27 member countries. Installation locations, not manufacturing origins. Other markets is the rounded total left after subtracting the named markets; its growth rate is not available.

Data: IFR · World Robotics 2026 · Chart: VastkindData: 2025 · Source: September 2026 · Expanded: 6 Oct 2026

China accounts for almost six in ten

The U.S. overtook Japan. The EU contracted. China alone installed more industrial robots than the rest of the world combined in 2025. It contributed about 59,000 of the 61,000 net additional installations worldwide, using IFR’s consistently rounded time-series figures.

More Installations Does Not Mean More Robots per Worker

These are annual industrial robot installations, not humanoids, all robots already operating or robots per worker. The EU-27 is not all Europe. Market sizes do not adjust for the size of manufacturing.

Updated October 6, 2026: expanded the original three-market comparison to six groups, with year-over-year changes and explicit residual rounding. Original URL and publication date retained.

Sources, Method & Data

2025 annual installations, with IFR’s reported percentage changes versus 2024. China 354,200, EU-27 60,500, U.S. 38,400, Japan 36,219 (displayed as 36.2k), South Korea 30,200. Most values are rounded in IFR’s public releases or graphics. “Other markets” is the global total of roughly 603,000 minus the five named groups: 83,481 calculated by subtracting the named markets from an already rounded world total, displayed as approximately 83k to avoid false precision. This includes European countries outside the EU. It is not the larger “rest of the world excluding China” total of approximately 249,000. A comparable growth rate for the residual group is not supplied; no rate is reverse-engineered from rounded percentages. Bars show installation counts on one common linear scale from zero; the separate percentage column is not a second bar scale.

China’s share is about 354,200 / 603,000 = 58.7%, displayed as 59%. The growth contribution uses the separate, consistently rounded original time-series graphics: global 542k to 603k and China 295k to 354k, hence approximately +61k globally, +59k in China, +2k elsewhere. No employment, robot-density, total-stock or humanoid inference. Measurement year 2025; release September 2026; accessed October 6, 2026.

2025 installation markets
MarketInstallations (thousands)Change from 2024
China354.2k+20%
EU-2760.5k−11%
United States38.4k+12%
Japan36.2k−19%
South Korea30.2k−1%
Other markets≈83kn/a

Explore Vastkind

Explore topics
For another moment

Your reading list

Saved on this device. No account needed.

Something catch your eye? Use the bookmark beside an article to keep it here.