Data dossier · as of 25 July 2026

Who causes the CO₂ — and how much of it is AI?

The world emitted around 38.4 gigatonnes of energy-related CO₂ in 2025. Data centres consumed 1.5 per cent of the world's electricity, AI-focused data centres alone 0.49 per cent. This page states every figure with its unit, year, source and certainty status — and shows the contradictions between the data sources instead of smoothing them over.

67 sources Latest source: 22.07.2026 Next update: Nov 2026 CC BY 4.0
Illustration: a teddy bear at a desk with books on AI basics, machine learning and AI ethics

Key numbers 2025

Four numbers that set the frame

2025 saw the slowest rise in energy emissions since 2021 — and still a record level, 5 per cent above 2019. Total CO₂ stayed practically flat because falling land-use emissions offset the fossil increase (Global Carbon Budget 2025).

38.4 Gt billion tonnes of CO₂ from energy in 2025 Coal, oil, gas, electricity, transport. A new high, 0.4 per cent above 2024 Measured
42.2 Gt billion tonnes, everything together Energy plus forests and soils. Practically unchanged against 2024 Measured
427 parts per million of CO₂ in the air Before industrialisation it was 278 — today around half again as much Measured
170 Gt billion tonnes of CO₂ may still be added After that, warming rises above 1.5 degrees — the limit of the Paris climate agreement. Reached in around four years at today's emissions, and even then only with a 50 per cent chance Forecast
Measured — recorded or reported Forecast — model value Derived — calculated from source figures Estimate — order of magnitude

On this page the certainty status sits next to every figure, not in a footnote. Every figure in the HTML additionally carries the attributes data-src and data-marker and can be checked by machine. Sources: IEA Global Energy Review 2026 · Global Carbon Budget 2025

System boundaries

Why there are six different totals for 2025

38.1 / 38.4 / 41.0 / 42.2 / 53.2 / 57.7 gigatonnes are not a contradiction but six system boundaries. They differ in whether land use, methane, industrial processes and further greenhouse gases are counted. Comparing two of these numbers without naming the boundary creates a false contradiction.

0102030405060 Gt

Choose a system boundary — the explanation changes, the page height does not:

38.1 Gt CO₂ Measured

Fossil CO₂ only, including the cement process

Combustion of coal, oil and natural gas plus the chemical process of cement production. Not included: land-use change, methane, nitrous oxide, F-gases. Growth in 2025: +1.1 per cent. This figure is 10 per cent higher than in 2015, the year of the Paris Agreement.

Source: Global Carbon Budget 2025 · status: measured

38.4 Gt CO₂ Measured

The IEA's energy-related CO₂

Combustion plus industrial processes plus flaring. The difference from the Global Carbon Budget's 38.1 Gt lies in how process emissions and flaring are treated, not in a data error. Split for 2025: combustion +0.5 per cent (+185 Mt), industrial processes −2 per cent (−40 Mt).

Source: IEA Global Energy Review 2026 · status: measured

41.0 Gt CO₂e Measured

CO₂ equivalents from energy

Additionally includes methane and flaring as CO₂ equivalents. The unit switches here from CO₂ to CO₂e — which is why this figure must never be netted against the 38.4 Gt. Growth in 2025: +1.1 per cent.

Source: Energy Institute, Statistical Review of World Energy 2026 · status: measured

42.2 Gt CO₂ Measured

Fossil CO₂ plus land-use change

38.1 Gt fossil plus 4.1 Gt from deforestation and land use. This total stagnated in 2025 (−0.04 per cent) because land-use emissions fell 9.8 per cent while fossil emissions rose 1.1 per cent. Over the past decade this figure grew 0.3 per cent a year, in the decade before that 1.9 per cent.

Source: Global Carbon Budget 2025 · status: measured

53.2 Gt CO₂e Measured

All greenhouse gases excluding land use

Reference year 2024, +1.3 per cent (+665 Mt). Gas mix: fossil CO₂ 74.5 per cent, methane 17.9 per cent, nitrous oxide 4.8 per cent, F-gases 2.8 per cent. Fossil CO₂ is 74.9 per cent above 1990. The G20 cause around 83 per cent of global CO₂.

Source: EDGAR/JRC 2025 · status: measured

57.7 Gt CO₂e Measured

All greenhouse gases including land use

Reference year 2024, the widest system boundary. Sector shares per UNEP: power 15.6 Gt (27 per cent), transport 8.4 Gt (15 per cent, of which road transport is more than 10 per cent of all emissions), industrial energy 6.5 Gt (11 per cent), agriculture 11 per cent (of which livestock 6 per cent).

Source: UNEP Emissions Gap Report 2025 · status: measured

Table: all six system boundaries
System boundaries compared
System boundaryValueUnitYearStatusSource
Fossil CO₂ only, incl. cement38.1Gt CO₂2025MeasuredGCB 2025
Energy-related CO₂38.4Gt CO₂2025MeasuredIEA GER 2026
CO₂e from energy41.0Gt CO₂e2025MeasuredEnergy Institute 2026
Fossil plus land use42.2Gt CO₂2025MeasuredGCB 2025
All GHG excl. LULUCF53.2Gt CO₂e2024MeasuredEDGAR/JRC 2025
All GHG incl. LULUCF57.7Gt CO₂e2024MeasuredUNEP 2025
A source conflict we do not smooth over: for the CO₂ concentration the IEA states around 427 ppm, the Global Carbon Budget 425.6 to 425.7 ppm. The likely cause is different averaging (annual mean versus year-end). The tile above gives the IEA value because it uses the annual mean. Both source values appear here so nobody has to guess where the figure comes from.

Causes by fuel

The driver has changed: from coal to gas

Coal remains the largest single fuel in 2025 at 41 per cent, but has practically stopped growing. Of the +185 Mt of CO₂ from combustion, around 85 Mt came from natural gas, 60 Mt from oil and only 25 Mt from coal — the year before, coal alone added +210 Mt. This shift contradicts public perception.

Share of fossil CO₂ 2025 — complete, adds up to 100 %

Coal 41 % · 15.6 Gt Oil 32 % · 12.2 Gt Natural gas 21 % · 8.0 Gt Rest 6 % · 2.3 Gt

Shares and growth rates: Global Carbon Budget 2025 / CSIRO Measured. The absolute Gt values are multiplied out of 38.1 Gt Derived and do not appear in the source in that form.

Who caused the 2025 increase — in Mt CO₂

Natural gas
+85
Oil
+60
Coalprevious year still +210 Mt
+25
Weather effectsa subset, not a fuel
around 90
020406080100 Mt

IEA Global Energy Review 2026 Measured. The weather row is a subset of the 185 Mt and must not be added to the fuels. More than 40 per cent of global gas demand growth came from higher heating demand in advanced economies; the US switch from gas to coal contributed +75 Mt.

Around half of the 2025 increase was weather, not structure. The IEA puts temperature outliers plus shortfalls in hydro and wind power at around 90 of the 185 Mt. Weather-adjusted, the emissions of advanced economies would have fallen by 0.5 per cent. Europe: 45 Mt extra from weak hydro (a 75 TWh shortfall), 20 Mt from weak wind. Global drought: +40 Mt. India: −15 Mt thanks to an early monsoon. 2025 was the third-warmest year on record.

Causes by sector

Nine sectors — and why their values cannot be added up

Power generation is the largest single sector at 13.9 Gt CO₂ (2025), followed by industry at around 10.6 Gt. Data centres sit at around 0.2 Gt — and are already contained in power generation, because they consume electricity.

These bars cannot be added up. They come from different system boundaries, mix CO₂ and CO₂e, and contain subsets. Specifically: livestock is stated in CO₂ equivalents (mostly methane, GWP-100) and must never enter a CO₂ sum. Data centres sit entirely inside power generation. Reference years range from 2015 to 2025.
02468101214 Gt

* different metric: CO₂ equivalents instead of CO₂. † subset of power generation, not additional. Two values for the same thing: the IEA states 13.9 Gt for power generation (2025, narrower boundary), EDGAR 15.5 Gt for the power industry (2024, wider boundary, around 39 per cent of fossil CO₂). The bar shows the IEA value. The residual “industry around 10.6 Gt” belongs to the EDGAR calculation (39.6 − 15.5 − 8.5 − 3.0 − 2.0 = 10.6), not to the IEA row — without this note the split looks inconsistent. Bar patterns encode the certainty status: solid = measured, translucent = derived, outline = estimate.
Sources: IEA Electricity 2026 · EDGAR/JRC 2025 · FAO GLEAM · IEA Breakthrough Agenda 2025 · GCB 2025 · IEA Energy and AI

Table: sectors with year, metric, status and source
Sectors — not addable
SectorValueMetricYearStatusSourceSubset of
Power generation13.9 GtCO₂2025MeasuredIEA Electricity 2026
Power industry, wider boundary15.5 GtCO₂2024MeasuredEDGAR/JRC
Industry incl. processesaround 10.6 GtCO₂2024DerivedEDGAR residual
Transport, total8.5–9 GtCO₂2024EstimateEDGAR / Statista
— Road transportover 6 GtCO₂2024MeasuredIEA Breakthrough 2025Transport
— Aviation and shippingaround 2.3 GtCO₂2024DerivedResidualTransport
Livestock6.2 GtCO₂e2015MeasuredFAO GLEAM
Deforestation, land use4.1 GtCO₂2025MeasuredGCB 2025
Buildings, households directaround 3.0 GtCO₂2024EstimateEDGAR
Fuel extractionaround 2.0 GtCO₂2024EstimateEDGAR
Data centres180 MtCO₂2024MeasuredIEA Energy and AIPower generation
Data centres194–210 MtCO₂2025Derived485 TWh × 0.40–0.43 t/MWhPower generation

Road transport in detail

  • Emissions 2024: just over 6 Gt CO₂, 8 per cent above 2015 Measured
  • Growth 2019–2024 only 0.2 per cent a year — 2015 to 2019 it was still 1.7 per cent Measured
  • Split: over 60 per cent cars and vans, around a third trucks, 7 per cent buses and two- and three-wheelers Measured
  • Emerging economies excluding China since 2015: +18 per cent, rest of the world +2.5 per cent Measured

IEA Breakthrough Agenda Report 2025

Livestock: a number with a revision history

  • 6.2 Gt CO₂e = around 12 per cent of all man-made greenhouse gases, reference year 2015 Measured
  • Of which direct (enteric fermentation, manure management): 3.7 Gt CO₂e = 7.2 per cent of total emissions Measured
  • By species: cattle 62 %, pigs 14 %, chickens 9 %, buffalo 8 %, small ruminants 7 % Measured
  • Diverging count: FAOSTAT 2023 states 4.3 Gt CO₂e for livestock within a 16.5 Gt agrifood system. The difference: GLEAM is a life-cycle analysis including feed-related land-use change, FAOSTAT follows inventory logic Measured
  • The FAO's own revisions: 18 % in 2006, 14.5 % in 2013, 12 % in 2022. Other studies: 15 % (Poore & Nemecek 2018), 16.5 % (Twine 2021) Measured
  • Forecast: rise to 9.1 Gt CO₂e by 2050 at +20 per cent demand Forecast

FAO GLEAM / FAOSTAT 2025 — shown on this page as a range with years, not as one number.

Distribution across regions

In 2025 the regions swapped roles

For the first time in almost 30 years, emissions rose more in advanced economies (+0.5 per cent) than in emerging and developing economies (+0.3 per cent). China fell 0.5 per cent on IEA actuals, India was flat for the first time since the 1970s, the USA added around 2.5 per cent. This is a snapshot driven by weather and a US policy shift towards coal — not a trend reversal.

Share of global fossil CO₂ 2025

01020304050 %

Global Carbon Budget 2025 Measured

Change in 2025 — switch the source

The IEA and the Global Carbon Budget contradict each other for individual regions. Both values stand here; neither is picked silently:

USA
+2.5 %
India
+1.1 %
Rest of the world
+0.9 %
China
+0.4 %
EU 27
−0.1 %
Japan
−0.9 %
−5−2.50+2.5+5 %

Global Carbon Budget 2025, final publication ESSD 18, 3211 (2026) Measured. Revision made transparent: against the COP30 press release of November 2025 the final figures were corrected — USA from +1.9 to +2.5 %, India from +1.4 to +1.1 %, EU from +0.4 to −0.1 %. Most secondary sources still carry the old numbers. Also: international aviation +6.7 %, international shipping +2.0 %.

USA
rising, no figure given
India
flat, first time since the 1970s
Rest of the world
no separate IEA figure
China
−0.5 %
EU 27
no separate IEA figure
Japan
no separate IEA figure
−5−2.50+2.5+5 %

IEA Global Energy Review 2026, actuals as of April 2026 Measured. The China divergence: the GCB projected +0.4 % in November, the IEA measures −0.5 %. The IEA's reasons: falling cement and steel output, a boom in low-emission power, slower demand growth — partly offset by the chemicals industry. The IEA figure rests on actuals and is the more reliable one. Gaps stay gaps: where the IEA gives no value, none appears here.

USA
+4.3 %
India
−3.3 %
South-East Asia
rising, no figure given
China
around −1 %
EU 27
−2.2 %
World
0.0 % flat
−5−2.50+2.5+5 %

IEA Electricity 2026 Measured. CO₂ intensity of power, 2025 and 2030 forecast: world 435 to 360 g/kWh, EU 170 to 90, China 530 to 415, India 695 to 585, South-East Asia 640 to 615 g/kWh Forecast. By 2030 India's intensity falls below South-East Asia's — which then becomes the most emission-intensive region.

Per-capita emissions 2024, fossil CO₂ in tonnes per person

0369121518 t

Destatis / EDGAR G20, October 2025 Measured; India from 3.22 Gt divided by around 1.46 bn people Derived. Context: the G20 cause around 83 per cent of global CO₂ (39.6 Gt in 2024). Historically: the USA has emitted around 400 Gt since 1751 — almost a quarter of all historical emissions and more than 1.5 times China's total (Our World in Data / GCB).

Growth curves

Where 2026 and 2027 are heading

Power-sector CO₂ was flat in 2025. For 2026 the IEA expects +1 per cent and thus a new record — caused by the LNG shock, not by a new growth trend. From 2027 it expects a plateau again. Data-centre electricity demand, by contrast, doubles by 2030.

Power-sector CO₂, change year on year

2023
+1.5 %
2024
+1.4 %
2025
0.0 %
2026
+1.0 %
2027
0.0 %

2023 to 2025 Measured IEA Electricity 2026 · 2026 and 2027 Forecast IEA Electricity Mid-Year Update, 22.07.2026. Absolute level around 13,900 Mt CO₂ a year. Carbon intensity fell 3 per cent in 2025 (after 2.6 per cent in 2024) and sits 14 per cent below its level of ten years ago — emissions stay flat only because demand grows at the same time.

Data-centre electricity demand in TWh

Data-centre electricity demand 2024 to 2035 415 TWh in 2024 and 485 TWh in 2025 as measured values, 950 TWh in 2030 and 1,200 TWh in 2035 as forecast. 0400 8001200 2024202820302035 4154859501,200

2024 and 2025 Measured, 2030 and 2035 Forecast — IEA Energy and AI 2025 / Key Questions on Energy and AI 2026. Share of world electricity: 1.5 per cent (2024) to around 3 per cent (2030). Crypto mining is not included in these IEA figures (see “AI versus the rest of the internet”).

World electricity demand

  • 2024: +4.4 % Measured
  • 2025: 28,600 TWh, +3.0 % Measured
  • 2026: +3.6 % Forecast
  • 2027: 30,700 TWh, +3.8 % Forecast
  • 2026–2030: +3.6 % a year, against 2.8 % in the previous decade Forecast

IEA Electricity 2026 / Mid-Year Update 2026

Coal — on a plateau

  • World demand 2025: 8,845 Mt, +0.5 % — a record Measured
  • Two thirds of coal use goes into power plants Measured
  • Share of the power mix: 35 % (2024) to 27 % (2030) Forecast
  • Demand 2030: −3 % against 2025, below the 2023 level Forecast
  • USA 2025: +10 % coal demand; China over 50 % of world use, coal power −1.4 % Measured
  • Shift 2025–2030: India +225 Mt, ASEAN +127 Mt; EU −153 Mt, USA −106 Mt Forecast

IEA Coal 2025 / IEA GER 2026

The LNG shock of 2026

The Strait of Hormuz crisis temporarily removed around 20 per cent of global LNG supply. Gas prices in Asia and Europe reached their highest level since the 2022/23 energy crisis; wholesale power prices in the EU and Japan were more than 30 per cent above the previous year in Q2 2026. The result: gas-to-coal switching in several Asian and European countries, power-sector CO₂ up 1 per cent in 2026 and thus a new record. Measured / Forecast

IEA Electricity Mid-Year Update 2026

The counter-account: emissions avoided in 2025, in Gt CO₂

00.40.81.21.6 Gt

IEA Global Energy Review 2026 Measured. Sum around 3 Gt = around 8 per cent of global energy emissions. Displaced since 2019: over 35 EJ of fossil energy = around 7 per cent of 2025 demand (23 EJ coal, 9 EJ gas, 3 EJ oil). The displaced coal exceeds India's entire coal demand. Additions in 2025: 692 GW of renewables, +15.5 per cent — the largest annual expansion ever, installed total 5,149 GW. Source divergence: the IEA says renewables were “virtually matching” coal in power generation, Ember says “renewables overtook coal power in 2025”. Both wordings stand here.

35 countries cut their CO₂ emissions over the past decade while growing their economies — twice as many as in 2005–2014. They account for 27 per cent of global fossil emissions. Among them: Australia, Bulgaria, Denmark, Germany, Estonia, Israel, Jordan, New Zealand, Portugal, Romania, Slovenia, Taiwan, Thailand, the United Kingdom and the USA. Measured · Global Carbon Budget 2025

The grid as counterweight: 2,500 GW of projects — renewables, storage and large loads such as data centres — are stuck in grid-connection queues worldwide. Grid investment must rise by around 50 per cent by 2030, from 400 bn USD a year today. Flexible connection contracts and grid-enhancing technologies could connect 1,200 to 1,600 GW in the short term. Forecast

AI and data centres

AI is a growth problem, not a volume problem — yet

Data centres cause around 0.5 per cent of global CO₂ and consume 1.5 per cent of the world's electricity; AI-focused data centres alone 0.49 per cent. What matters is not that share but the direction: data centres are one of only three sectors whose emissions still rise to 2030. The other two are road transport and aviation. Everything else decarbonises.

485 TWhterawatt-hours of electricity for all data centres in 202517 per cent more than in 2024. Crypto mining is not counted here Measured
194–210 Mtmillion tonnes of CO₂ from all data centres in 2025Calculated from the 485 terawatt-hours, at 0.40 to 0.43 tonnes of CO₂ per megawatt-hour Derived
0.49 %AI's share of world electricityall data centres: 1.5 % Measured
+50 %more electricity was needed by AI data centres in 2025All data centres together grew by only 17 per cent Measured

How much of the world's electricity is AI? — orders of magnitude 2025

Of 100 per cent of the world's electricity, 1.5 per cent goes to all data centres and 0.49 per cent to AI data centres. The rest — more than 98 per cent — is everything else.

0 %world electricity = 100 %

The thin gold strip on the left is the 1.5 per cent of all data centres — that is how small the share of total world electricity is.

Other data centrescloud, streaming, web
1.01 %
AI data centresabout a third of data-centre electricity
0.49 %

The lower chart zooms into the 1.5 per cent: AI is about a third of all data-centre electricity, not the whole. Bar width shows the share of data centres, the figure shows the share of world electricity.

Is · IEA, Key Questions on Energy and AI 2026. Shares of world electricity: data centres 1.5 %, AI data centres 0.49 %. The 1.01 % other and the one-third split are Derived from these.

Regional concentration: data centres' share of electricity use

01020304050607080 %

IEA Key Questions on Energy and AI 2026, Dublin per the Öko-Institut Measured. All values exclude crypto mining. Ireland forecast: 32 per cent by 2026 Forecast. These six bars are the core of the whole debate: Dublin 79 per cent and world 1.5 per cent are both correct. Quoting only one of them misleads.

Overview graphic myth versus reality in three rows: labour market, resource hunger, danger and deception — the myth on the left, facts and studies on the right
Illustrative overview graphic, AI-generated (NotebookLM), German labelling. It reflects the structure of the debate but is not a source for this page. Two of its claims are not covered by the primary sources used here: the equation “10–50 requests = 500 ml of water” and a global water demand of 6.6 bn m³ by 2027. Google's measured value is 0.26 ml of water per median text prompt — median means the typical request, not the average (Google, August 2025) — the figures in the next section are authoritative.

Investment and infrastructure

  • Hyperscaler capex 2025 over 400 bn USD, +75 per cent expected for 2026 Measured / Forecast
  • The capex of five tech companies exceeds global investment in oil and gas extraction Measured
  • “AI factories”: capacity more than tripled in 18 months (IEA satellite tracking) Measured
  • AI server power density: ×11 between 2020 and 2025, a further ×4 by 2027 Measured / Forecast
  • One server rack in 2027 equals the peak load of 65 households — on the footprint of a fridge Forecast
  • Load swings of more than 50 per cent of rated power within one second; heat output like around 30 gas boilers per rack Measured
  • On-site gas plants for US data centres by 2030: 15–27 GW, requiring 30–70 per cent overbuild; gas turbine orders 2025 +70 per cent Forecast / Measured
  • Hardware bottleneck: shortage of high-bandwidth memory, expected to last at least until the end of 2027 Measured

IEA Key Questions on Energy and AI 2026

The efficiency counter-evidence

  • Energy per AI task falls by at least one order of magnitude per year Measured
  • A simple text query uses less electricity than a television over the same period Measured
  • If all conventional web searches ran as simple AI text queries, that would be under 4 TWh a year = under 1 per cent of today's data-centre consumption Measured
  • But: video generation, reasoning and agentic tasks use one hundred to one thousand times the energy of a text query Measured
  • Usage: model providers report 3× more active users and 5× more revenue within a year Measured
  • Training is not the problem: training Grok 4 used around 0.31 TWh = 0.2 per cent of total AI electricity in 2025. Inference dominates Measured

IEA 2026 · Google 2025 · Epoch AI

Inside a data centre

  • Servers around 60 per cent of electricity use, up to 75 per cent in AI-optimised hyperscale sites
  • Cooling 7 per cent (efficient hyperscale) to over 30 per cent (inefficient enterprise)
  • Storage around 5 per cent, networking up to 5 per cent
  • Accelerator servers (GPU/TPU, essentially AI): around 24 per cent of server electricity, around 15 per cent of total consumption in 2024
  • Growth: accelerator servers +30 per cent a year, conventional servers +9 per cent. AI chips draw 2–4× more watts than conventional counterparts

IEA Key Questions on Energy and AI 2026 Measured. Power mix for US data centres: natural gas over 40 per cent, renewables 24 per cent, nuclear around 20 per cent, coal around 15 per cent. By 2030 gas and coal together cover over 40 per cent of the additional data-centre demand Forecast.

AI as an emissions saver — the other direction

  • Existing AI applications in end-use sectors could avoid 1,400 Mt of CO₂ in 2035 — three to four times what the data centres themselves emit Forecast
  • Documented use cases: over 13 EJ of energy savings by 2035 = 3 per cent of global final energy use Forecast
  • AI optimisation in energy-intensive industry: 3–10 percentage points lower energy costs Forecast
  • Barrier number one per the IEA's survey of energy companies: lack of digital skills. Only 10 per cent of global electricity use is covered by open-data policies Measured

IEA, Widespread Adoption Case. These are forecasts under assumptions — they stand here because a page showing only AI's costs would be as one-sided as one showing only its benefits.

Table: electricity and emissions of the four large providers — with accounting method
Hyperscalers, environmental reports June/July 2026
CompanyElectricity 2020Electricity 2024Electricity 2025Emissions 2025PUEAccounting
Google15.2 TWh32.2 TWh43.6 TWh (+37 %)+18 %1.09The reported 12 % cut in data-centre energy emissions is plausible only market-based; data-centre electricity use rose 27 % from 2023 to 2024. +37 % is the largest single-year jump in company history.
Microsoft10.8 TWh29.8 TWh37 TWh (+24 %)+25 %1.17Market-based scope-2 emissions fell from 456,119 t (2020) to 259,090 t (2024) while electricity use almost tripled. Largest single site: Boydton, Virginia, over 3 TWh.
Amazonnot disclosed+16 %, scope 2 +34 %1.14In absolute terms 64.38 Mt CO₂e (2023) to 68.25 Mt (2024).
Metanot disclosed1.09No disclosure of electricity use.

All values Measured from the June/July 2026 environmental reports, compiled by Trellis. Amazon, Google and Microsoft's combined data-centre investment for 2025 and 2026: around 750 bn USD. ITU (UN) for 2020–2023: indirect emissions of the four up an average of +150 per cent — Amazon +182 %, Microsoft +155 %, Meta +145 %, Alphabet +138 %.
Why the accounting method is stated: market-based accounting credits purchased green-power certificates and can report falling emissions while consumption rises. Location-based accounting uses the real grid mix. Without this label a company figure cannot be interpreted.

AI versus the rest of the internet

Non-AI data centres still use twice as much electricity as AI in 2025

AI-focused data centres needed around 155 TWh in 2025, all other data centres 330 TWh. Parity only arrives in 2030 (465 versus 480 TWh). Growth rates in 2025: data centres overall +17 per cent, AI data centres alone +50 per cent.

This gap is named, not filled: a reliable split exists between AI and non-AI. No source separates social media, search engines and streaming from one another. All three run on the same hyperscale sites and the same billing units; no operator publishes workload-specific breakdowns. Our World in Data puts it as “not always clean-cut”. Anyone drawing a chart “Facebook versus Google Search versus LLMs” is inventing numbers. In addition, no provider publishes query volumes — Google states energy per prompt but not the number of prompts. The question “how much CO₂ for LLMs, how much for Facebook” therefore cannot be resolved at company level.

Data-centre electricity demand by type, in TWh

2025 Measured660 TWh incl. crypto
2030 Forecast945 TWh excl. crypto
AI-focused data centres Non-AI data centres Crypto mining — not included in the IEA figures, 150–200 TWh Estimate

Bar lengths to scale against 1,200 TWh. IEA Key Questions on Energy and AI 2026, compiled by Our World in Data, 20.07.2026. World electricity generation 2025 around 31,800 TWh (Ember). Emissions derived for 2025: non-AI 132–143 Mt, AI 62–67 Mt, crypto 45–80 Mt Derived. Cross-check: the IEA itself states around 180 Mt for 2024 at 415 TWh — 485 TWh × 0.40–0.43 t/MWh gives 194–210 Mt, consistent.

A source divergence every chart must label: the IEA states 485 TWh for 2025 (1.5 per cent of world electricity) excluding crypto. The Energy Institute and S&P state around 790 TWh (2.5 per cent) including crypto. The difference of around 305 TWh is explained almost two thirds by crypto mining (150–200 TWh). The rest is methodology: S&P models bottom-up from installed capacity and utilisation assumptions, the IEA estimates IT equipment, cooling and infrastructure directly. For 2020 the same two figures are 270 versus 411 TWh. The two numbers are not directly comparable.

The structural contrast: where the electricity is used

Video streaming, 1 hour — around 36 g CO₂ Estimate
LLM inference Measured
Data centre — streaming 5 % / LLM 96 % Network — 23 % / under 1 % End device — 72 % / around 3 %

Streaming: IEA analysis (George Kamiya) via Carbon Brief. LLM: life-cycle analysis of Mistral Large 2. The consequence: comparing the data-centre emissions of streaming and LLMs compares 5 per cent of one footprint with 96 per cent of another. Older streaming estimates (Shift Project: 3,200 g/h) were about a factor of 90 too high; the peer-reviewed 2014 study found 420 g/h for US conditions in 2011.

Energy per request — logarithmic scale in watt-hours

0.1110100 Wh

Filled dots = measured Measured, rings = Estimate. Sources: Google, measured May 2025 · Epoch AI · IEA 2026, Figure 2.1 · Our World in Data. The spread within AI: around a factor of 200 between a median text prompt (0.24 Wh) and a reasoning agent (around 50 Wh). “One AI request” is worthless as a unit.

Table: energy per request with status and source
Energy and CO₂ per request
OperationEnergyCO₂StatusSource
Gemini, median text prompt0.24 Wh0.03 g CO₂eMeasuredGoogle, May 2025, comprehensive methodology, plus 0.26 ml of water
Gemini, narrow methodology0.10 Wh0.02 gMeasuredGoogle, same publication, TPU/GPU only
ChatGPT, typical requestaround 0.3 WhEstimateEpoch AI
ChatGPT, per Sam Altman0.34 WhEstimateblog post, no methodology
Agent requestaround 1.2 WhEstimateIEA 2026, Figure 2.1
Long request, 7,500 words2.5 WhEstimateEpoch AI
GPT-5, averageover 18 WhEstimate, weakest evidencesecondary report via DCD, peaks 40 Wh
Maximum request, 75,000 words40 WhEstimateEpoch AI
Agent with reasoningaround 50 WhEstimateIEA 2026
Reference: 1 minute of electric showeraround 160 WhMeasuredOur World in Data
Reference: 1 hour of Netflixaround 36 g CO₂EstimateIEA
Infographic: energy per AI request on a logarithmic scale from 0.1 to 200 watt-hours, measured values as filled dots, estimates as rings
The same scale as a shareable graphic — with a note that belongs here. The data points, the measured/estimated split, the source keys and the box on the debunked rule of thumb are correct and match the scale above. Four labelling errors are, however, in the image: “Gemessen Warte” and “Geschätze Warte” should read “Gemessene Werte” and “Geschätzte Werte”, “vemparison for 1 minute elektric shower” is garbled, and the “40 Wh Maximal-query, with Reasoning” label is duplicated — the upper one should read 50 Wh, agent with reasoning. The graphic also mixes German and English labels. We show it anyway and name the errors instead of silently replacing it: a corrected version will follow. The scale above is authoritative.
A rule of thumb this page does not use: “an AI request needs 10× the energy of a web search.” It is incompatible with Google's measured 0.24 Wh. There is no current official figure for a Google search; the widely quoted 0.3 Wh comes from a Google statement of 2009. The rule circulates because both sides of the comparison are unknown.

An efficiency leap that further devalues the comparison: energy per Gemini prompt −33× in twelve months, CO₂ −44× (scope 2 market-based −47×, scope 1+3 −36×). That is the fastest documented efficiency gain in energy history — and it is being consumed by new application classes (video, reasoning, agents).

AI and work

What is forecast — and what is measured

Up to mid-2026, no large-scale study using administrative or payroll data finds economy-wide displacement by AI. At the same time, displacement at career entry is already measurable: −13 per cent relative employment among 22- to 25-year-olds in exposed occupations, −20 per cent among young software developers. The gap between these two findings is the point of this section.

Column 1 · What is forecast

Models and surveys

  • WEF Future of Jobs 2025: net +78 m roles by 2030 (+170 m new, −92 m displaced), 22 per cent churn on 1.2 bn formal jobs Forecast
  • WEF, further figures: 39 per cent of skills become outdated 2025–2030; 86 per cent of employers expect AI transformation; 63 per cent name the skills gap as the main obstacle; 120 m workers at medium-term risk for lack of training Forecast
  • ILO Working Paper 140: 1 in 4 workers is in an occupation with GenAI exposure, but only 3.3 per cent in the highest exposure tier. Women 4.7 versus men 2.4 per cent; high-income countries 9.6 versus 3.5 per cent Measured index
  • IMF: around 40 per cent of jobs affected worldwide, 60 per cent in advanced economies Forecast
  • Goldman Sachs: 300 m jobs exposed, but only 2.5 per cent of US employment at direct displacement risk Forecast
  • McKinsey: 60–70 per cent of tasks technically automatable, 12 m occupational switches in the USA by 2030 Forecast
  • PwC: 30 per cent of US jobs fully automatable by the mid-2030s, a 56 per cent wage premium for AI skills Forecast
  • Forrester: net −6.1 per cent of US jobs (10.4 m) by 2030 — the most pessimistic serious forecast; it stands against WEF, ILO, IMF and Goldman Sachs Forecast
  • Dario Amodei, May 2025: AI could halve entry-level office jobs within five years Statement, not a study

Citation note: a “WEF Future of Jobs Report 2026” does not exist. The report appears every two years (2020, 2023, 2025); the next edition is expected in 2027. All 170/92/78-million figures come from the 2025 edition (January 2025; base: over 1,000 companies, 14 m employees, 55 economies).

Column 2 · What is measured

Administrative and payroll data

  • Yale Budget Lab: after 33 months no discernible economy-wide disruption. Exposure, automation and augmentation measures show no link to employment or unemployment, and no significant effects on real hourly wages. Title of the May 2026 update: “AI Is Probably Not (Yet) the Reason for Labor Market Weakening” Measured
  • Challenger, Gray & Christmas 2025: around 55,000 layoffs explicitly attributed to AI — out of 1.17 m total, i.e. 4.7 per cent Measured
  • New York Fed: 1 per cent of service firms have actually laid off because of AI, 13 per cent plan to Measured
  • Humlum & Vestergaard 2025, Denmark: administrative data, 11 exposed occupations, zero effect on earnings and hours up to 2024 Measured
  • Hartley, Jolevski, Melo & Moore 2026: 35.9 per cent of US workers used GenAI by December 2025, small positive wage effects, no significant declines in postings in exposed occupations Measured
  • US labour market, March 2026: unemployment rate 4.3 per cent (post-pandemic low 3.4 per cent in April 2023), job growth only around 20,000 a month — low layoffs and low hiring Measured
  • Goldman Sachs Research: aggregate effects “still negligible”; below-average job growth in marketing consultancy, graphic design, office administration and call centres Measured
  • International AI Safety Report 2025: “evidence of broader labour market disruption remains limited” Measured

The US labour market is weak at the same time for other reasons. Attributing the weak labour market monocausally to AI mixes correlation and causation — that is the justified core criticism of the AI-critical source material of this page.

The exception: career entrants — declines in per cent

0−10−20−30−40−50−60−70−80 %

All values Measured. Sources: Brynjolfsson, Chandar & Chen (Stanford, ADP payroll data, millions of workers, over 730 occupations) · Hosseini & Lichtinger (Harvard, 66 m workers, over 280,000 US firms) · SignalFire · techUK · Revelio Labs. Further values: graduates were only 7 per cent of Big-Tech hires in 2024 (−25 per cent versus 2023, over 50 per cent below 2019); start-ups fell from 30 per cent (2019) to under 6 per cent (2024); unemployment among graduates aged 22–27 stood at 5.7 per cent at the end of 2025 against a 4.2 per cent overall rate, with 43 per cent underemployed.
The authors' own caveat (February 2026 update): with the broadest set of controls, the decline in AI-exposed occupations becomes statistically significant only from 2024; earlier declines are less firmly established.

The mechanism — stronger than any narrative

The Harvard study shows that the slump began before actual automation. AI mentions in US earnings calls tripled by mid-2023; firms cut their entry-level hiring immediately afterwards — on the basis of expected, not realised automation. Senior employment at the same firms kept growing.

It was not the technology that cut the positions, but the expectation of it. Brynjolfsson's summary: “It appears what younger workers know overlaps with what LLMs can replace.”

Hosseini & Lichtinger (Harvard) 2026, data 2015–2025 Measured

Counter-examples that belong to the full picture

  • McKinsey announced 12 per cent more hiring for 2026 Measured
  • Entry-level postings in healthcare rose by 13 percentage points Measured
  • Healthcare, public service, and leisure and hospitality together created almost 75 per cent of all new jobs in late 2024 and 2025 Measured
  • Nursing assistants grow faster among young workers than among older ones — the opposite of the software finding Measured
  • Shift, not disappearance: the share of tech postings asking for 2–4 years of experience fell from 46 to 40 per cent, for 5+ years it rose from 37 to 42 per cent (mid-2022 to mid-2025, Indeed) Measured

The damage is not in the total number of jobs but in broken career ladders — and in entry positions whose tasks are more than 50 per cent automatable, a fivefold risk against senior roles (Brookings) Forecast.

Productivity in controlled experiments — including the negative findings

−40−200+20+40 %

All values Measured from studies with random assignment or a study plan registered in advance. Sources: Brynjolfsson, Li & Raymond (5,179 support agents, Fortune 500) · Cui et al. 2025 · Noy & Zhang, Science (453 graduates, study plan registered in advance) · Dell'Acqua et al. (758 BCG consultants, GPT-4, “jagged frontier”) · METR 2025 · Vaithilingam et al. · Peng et al. (95 professional developers, an HTTP server in JavaScript: time −55.8 per cent, no effect on completion rate).
The METR finding matters most: experienced developers were 19 per cent slower — while estimating themselves to be 20 per cent faster. A caveat that belongs with it: the study ran on the developers' own repositories with unfamiliar tooling, so it does not transfer without qualification. A chart showing only the positive experiments is advertising.

The most telling number of the whole research: 88 per cent of organisations use AI in at least one business area (McKinsey State of AI, around 2,000 organisations, 105 countries; the Stanford AI Index 2026 independently confirms the same 88 per cent), generative AI specifically at 70 per cent — but only around 6 per cent of companies attribute a measurable EBIT effect to AI. Measured

Productivity forecasts: the factor of 20, made tangible

Choose a source — the annual rate and the ten-year effect derived from it change:

+0.07 %per year · total factor productivity

Cumulative over ten years at a constant rate: +0.70 % Derived. Acemoglu himself states +0.53 to 0.66 per cent TFP over ten years and a US GDP gain of 1.1 to 1.6 per cent Forecast.

The calculation, traceable and therefore attackable: 20 per cent of US work tasks are AI-exposed (Eloundou et al. 2023) × 23 per cent of those profitably automatable (Svanberg et al. 2024) × 27 per cent average cost saving per task = 0.53 to 0.66 per cent TFP. Anyone wanting a different number must attack one of these three factors.

Context that critics overlook: 0.07 per cent extra TFP a year against a base trend of 0.5 per cent a year since 2007 is not nothing — it is +14 per cent on the trend. Source: Acemoglu, The Simple Macroeconomics of AI, May 2024.

+0.4 ppper year · labour productivity

Cumulative over ten years at a constant rate: +4.1 % Derived.

The lower end of the OECD range. The range explicitly applies only to countries with high exposure and broad adoption — it is not a forecast for the world economy.

OECD 2025 Forecast. An overview of the estimate landscape also at the ECB, “AI and the euro area economy”, 23.03.2026.

+1.3 ppper year · labour productivity

Cumulative over ten years at a constant rate: +13.8 % Derived.

The upper end of the same OECD publication. Between the lower and upper end of a single source lies already more than a factor of 3.

OECD 2025 Forecast

+1.5 %level by 2035 · TFP and GDP

Not an annual rate but a permanent level shift: +1.5 per cent by 2035, around +3 per cent by 2055, +3.7 per cent by 2075 Forecast. No ten-year cumulative can be formed from this — so none is shown.

Penn Wharton comes out higher than Acemoglu because it counts a task as exposed once AI can take over 50 to 90 per cent of its components.

Penn Wharton Budget Model, 08.09.2025

+1.5 ppper year · US productivity

Cumulative over ten years at a constant rate: +16.1 % Derived. Goldman Sachs additionally states +7 per cent of global GDP = 7 tn USD over ten years Forecast.

This is the optimistic end of the serious estimates. Against Acemoglu's 0.07 per cent a year that is around a factor of 20 — on identical data.

Goldman Sachs 2023 Forecast

+1.5 to 3.4 ppper year · GDP growth, advanced economies

Cumulative over ten years at the upper end: +39.7 % Derived. McKinsey additionally states 2.6 to 4.4 tn USD a year, alternatively 17.1 to 25.6 tn USD Forecast.

The same consultancy simultaneously reports that only around 6 per cent of companies attribute a measurable EBIT effect to AI. Both statements stand side by side, uncontradicted.

McKinsey 2023 Forecast

The cumulative ten-year values are Derived as (1 + r)10 − 1 at a constant rate and appear in no source in that form. They make the roughly 20-fold spread between the annual rates tangible. Different sources measure different quantities — total factor productivity, labour productivity and GDP are not the same; hence the quantity is stated next to every figure.

Leading indicator: the delegation rate

27 → 39 %share of full delegation, Dec 2024 to Aug 2025
41 / 56 %automation versus augmentation, Jan 2025 — reversed for the first time in August
77 / 12 %automation versus augmentation via the API
49 %share of occupations with at least a quarter of tasks already delegated

Anthropic Economic Index Measured. Why this is the best leading indicator: monthly resolution, over 730 occupations, measured directly on delegation behaviour — labour statistics lag by two to four quarters. The Yale Budget Lab now uses this data itself and finds: it too shows stability, not disruption, though it is volatile in level. The November 2025 data shows a slight reversal towards more augmentation.

Where alarmism begins

When a local finding becomes a global claim

The step from a local finding to a global claim is the most common error in the AI environmental debate: a finding that is true locally is presented as a global state. Data centres use 79 per cent of the electricity in Dublin and 1.5 per cent worldwide. Both figures are correct. Quoting only one of them misleads.

Seven claims, one list, two scales. Switch and watch the status of each row flip:

Scale: global / aggregate
ClaimStatus, globalEvidence
Massive environmental damage from data centresoverstatedData centres at 1.5 % of world electricity, AI alone 0.49 %. Data-centre CO₂ around 0.5 % of global combustion emissions. The ICT sector overall at around 1.9 % of greenhouse gases; aviation at 2.5 % for comparison.
Companies miss their climate targetsnot applicableThe claim is defined at company level. No global equivalent exists — so none is stated here.
Investor greed and FOMO instead of valuenot applicableA capital-market finding, not a volume finding. Cannot be sensibly aggregated globally.
Faulty, imposed technologynot applicableA finding at application and company level. No global measure for it exists.
Entry-level jobs are being destroyednot supportedYale Budget Lab after 33 months and in the January and May 2026 updates: no discernible economy-wide disruption. US unemployment 4.3 % (March 2026). Only 4.7 % of 2025 US layoffs were attributed to AI at all.
Extreme inequality, gains for tech elitesforecastNo measurement. Acemoglu expects rising inequality between capital and labour income — a model statement, not an observed value.
Gen Z crowds into the trades, market overrunnot verifiableNo reliable evidence found. This claim is not used on this page — not even in weakened form.
Scale: local / sectoral
ClaimStatus, localEvidence
Massive environmental damage from data centressupportedDublin 79 %, Virginia over 25 %, Ireland over 20 % of electricity use. Google electricity +37 % (2025) to 43.6 TWh, emissions +18 %. Microsoft +24 % to 37 TWh, emissions +25 %. Amazon +16 %. ITU: indirect emissions of the four companies +150 % (2020–2023).
Companies miss their climate targetssupportedGoogle: net-zero target 2030, emissions 2025 +18 %. Amazon: target 2040, +16 %. Microsoft: +25 % in fiscal 2025.
Investor greed and FOMO instead of valuesupportedHyperscaler capex over 400 bn USD (2025), +75 % expected for 2026. Capex of five tech companies larger than global oil and gas extraction investment. Amazon, Google and Microsoft around 750 bn USD for 2025/26. Gas turbine orders +70 %. Plus the Allbirds case, see the timeline below.
Faulty, imposed technologysupportedThe Apple Intelligence case, see the timeline below. BCG experiment: −19 percentage points outside the capability frontier. METR 2025: experienced developers 19 % slower while estimating themselves “20 % faster”. McKinsey: 88 % adoption, around 6 % measurable EBIT effect.
Entry-level jobs are being destroyedsupported, even understatedHarvard −80 % entry-level hiring at adopting firms. Stanford −13 % relative employment among 22- to 25-year-olds, young software developers −20 %. UK tech graduate roles −46 %. The source video said 15 % — that figure is too low.
Extreme inequality, gains for tech elitesplausibleSupported by capital-market data, but not quantified in the source video. Therefore shown here without a figure of its own.
Gen Z crowds into the trades, market overrunnot verifiableNo reliable evidence found — not at local level either. This claim is not used on this page.

Both views sit in the HTML and are readable without JavaScript — the point emerges in the switching, not in hiding. Data base: all sourced sections of this page.

Two evidence chains — chronology instead of narrative

An AI-assisted analysis of the source video classified exactly these two cases as invented. Both are documented. Every entry carries a date and a primary source.

Case 1 · Apple Intelligence

  1. Apple Intelligence condenses three New York Times articles into one notification, including “Netanyahu arrested”. In reality it was an arrest warrant by the International Criminal Court. MeasuredPrimary source: BBC News, December 2024

  2. Push notification carrying the BBC logo, verbatim: “Luigi Mangione shoots himself; Syrian mother hopes Assad pays the price; South Korea police raid Yoon Suk Yeol's office”. Only the first part is false; the other two are correct. MeasuredPrimary source: BBC News

  3. The BBC files a formal complaint with Apple; Apple declines to comment. Further documented false alerts: that Rafael Nadal had come out as gay; that Luke Littler had won the PDC World Darts final — before the match. Reporters Without Borders calls on Apple to remove the feature. MeasuredPrimary source: BBC News

  4. With the developer previews of iOS 18.3, iPadOS 18.3 and macOS Sequoia 15.3, Apple disables notification summaries for news and entertainment apps entirely, announces reactivation after improvements, and adds extra transparency labelling. Apple's wording: summaries for this category are “temporarily unavailable”. MeasuredPrimary source: TechCrunch, 16.01.2025

What the source video overstates: Apple did not “withdraw its AI features entirely” — it temporarily switched off one sub-feature. Apple Intelligence stayed on the market. The core of the claim — Apple had to pull an AI feature over grotesque failures — is accurate.

Case 2 · Allbirds to NewBird AI to Smartbird

  1. Allbirds closes its US full-price stores. MeasuredTechCrunch, 15.04.2026

  2. Sale of the brand and shoe assets to American Exchange Group for 39 m USD. MeasuredCBS News, 16.04.2026

  3. Announcement of the pivot to AI compute infrastructure, renaming to NewBird AI, a 50 m USD convertible financing. Business model: GPU-as-a-service and AI-native cloud. The stock (NASDAQ: BIRD) rises from under 3 USD to 16.99 USD, +582 per cent in one day; market capitalisation multiplies sevenfold. Around 30 per cent back down by midday the next day. MeasuredPrimary source: SEC Form 8-K, Exhibit 99.1, 15.04.2026

  4. Charter amendment put to the vote: deletion of the designation as an “environmental conservation public benefit corporation”. Allbirds was a certified B Corp. A day after the announcement, social-media provider Myseum also pivots to AI. MeasuredPrimary source: SEC filing

  5. Renamed again, to Smartbird. New CEO: Nadia Carlsten, previously head of the AWS Quantum Computing Center, before that US Department of Homeland Security, most recently CEO at DCAI. She replaces Joe Vernachio. MeasuredCNBC, 17.06.2026

Why this case is central: it is the strongest single piece of evidence for the “investor greed over value” claim — a shoe maker without AI experience sells its product, deletes its environmental purpose from its charter and gains 582 per cent in market value in a day. Context: after the 2021 IPO the stock traded above 600 USD. GlobalData analyst Neil Saunders finds it unclear what expertise the company has in AI compute. Precedent: Long Island Iced Tea renamed itself Long Blockchain in 2017; the stock rose around 275 per cent, and NASDAQ delisted it the following year.

The marker proof — why this page labels every figure

Alongside the AI-critical source video sat an analysis produced with NotebookLM (submitted 25.07.2026). It classified the two cases documented above as invented, verbatim: “The video presents obviously fictional or satirically exaggerated scenarios as historical facts” and “These specific events did not happen this way in reality.” Both cases are, as the timelines above show with primary sources, real.

The same analysis conversely listed, under “factually correct statements”, the “significantly higher energy use per request compared with classic Google search”. That rule of thumb is incompatible with Google's measured 0.24 Wh; no current official figure for a Google search exists.

Two different error types, fixed in different ways. The Allbirds case (April to June 2026) lay outside the checking tool's knowledge cut-off — fixable with current sources. The Apple case dates to December 2024 and January 2025; a knowledge cut-off does not explain it. Here a documented event was classified as satire because it sounded absurd. This second type is not fixed by newer data and is the more dangerous one: the more grotesque a real AI failure, the higher the probability that an AI check declares it invented.

The working rule that follows and applies to this page: no AI-assisted fact check as the final authority for events of the past 18 months. For every case study, a primary source with a date — an SEC filing, a company statement or a news agency report. That is exactly why every figure here carries a status marker.

No judgement of the tool, no generalisation to other tools — only the documented single case with quote, date and cause. And the obvious self-reference: this page was also built with AI assistance. That is precisely why it is built so that every single figure can be checked against its primary source.

Two checked exhibits

These two overview graphics were also AI-generated (NotebookLM) and accompany the project as material. They are shown here not as sources but as exhibits: both reproduce the same debunked rule of thumb.

Infographic The resource shock: AI versus classic IT, with panels on energy per request, water footprint, electricity explosion and climate targets
Exhibit 1 — what is wrong here (German labelling). The graphic sets “classic search: 0.3 Wh” against “AI chatbot: 2.9 Wh (around 10×)”. The 0.3 Wh value is a Google statement from 2009; no current official figure for a Google search exists. Google's measured value for a median text prompt is 0.24 Wh (August 2025, comprehensive methodology). The 10× comparison therefore does not hold. Formally, panel 2 appears twice and several labels are misspelled — typical artefacts of AI-generated image labelling. The “304 TWh 2028” figure is not comparable with the IEA values used here (485 TWh 2025, 945 TWh 2030) because the system boundary is missing.
Infographic The artificial distance: why Gen Z rejects AI — five worries on the left, facts and studies on the right
Exhibit 2 — the same rule of thumb, again (German labelling). Here too, “10× energy use per request” appears with 0.3 Wh against 2.9 Wh. The worry/fact structure is usable as a map of the debate; the numbers are not consistently sound, several labels are corrupted, and sources and dates are partly missing. The point of these two exhibits: three independently generated AI fact graphics on the same topic reproduce the same unsupported rule of thumb. That is exactly the mechanism this page interrupts with a status marker and a primary source on every figure.
What the source video is rightly criticised for: it mixes correlation and causation — the aggregate data (Yale Budget Lab, Danish administrative data, Hartley et al. 2026) shows no economy-wide AI effect while the US labour market is weak for other reasons at the same time. The most tangible criticism is a conflict of interest: the surveillance and data-loss anxiety the video builds is used directly as a sales lever for a VPN service. And the rhetorical arc from the personal micro to the global macro level produces exactly the local-to-global step this section is about.

The counter-position

The same rigour for the optimist side

What was checked: an Economist interview with Elon Musk of 23 July 2026. The result: the verifiable core statements hold — twelve of 23 reach source quality 4 or 5, and two are actually understated. Only one category remains clearly unsupported: statements that turn a measured trend into a leap. Anyone expecting an interview with an interested party to consist mostly of false statements finds the opposite.

Two caveats that apply to this whole section.
First: the interview is publicly checkable. Original publication: The Economist, “An interview with Elon Musk” (The Insider) — subscription may be required. Official excerpt on The Economist's channel: “Elon Musk on AI: humans will no longer be in control in ten years”. Full recording with transcript, republished on a third-party channel: “Elon Musk The Economist July 23rd 2026 — FULL Interview”. The factual statements on this page are checked against primary and specialist sources; the attribution of statements to the speaker rests on a summary of the transcript, not on our own transcription of the recording. That is why no verbatim quote appears here. Anyone who hears a statement differently in the original can check it under the links — and correct us.
Second: for 7 of the 23 statements the source material identifies a commercial interest. It is stated in those matrix rows — and only there. For the other 16 the entry reads “none identified”, because we found none. That is neither discrediting nor exonerating; it is information that belongs to the assessment — exactly like the VPN provider in the previous section.
Political value judgements are neither judged nor omitted; they are labelled as not assessable.

Claim-check matrix — 23 statements, sorted by source quality

Seven of the 23 statements reach the highest source quality; twelve sit at 4 or 5. Two of them are understated — the US interest burden and the speaker's own voting share. Anyone expecting an interview with an interested party to consist mostly of false statements finds the opposite.

How to read this matrix — three rules that make it usable.
First: “source quality” measures how well a statement is documented — not whether we agree with it. 5 means several primary sources. 2 means checkable, but the check comes out mixed. A low value is not a refutation.
Second: that a position matches the professional consensus is a pointer to its source quality — not proof of truth. Majorities in a research field have often been wrong. Where a row says “matches the prevailing view”, that is a location, not a confirmation.
Third: an interest is stated only where the source material explicitly identifies one — in 7 of 23 rows. Where nothing is stated, we found none and invent none. An imputed interest would be an imputation, not a check.
Source quality 1 to 5, as of 25.07.2026
Statement, as made in the interviewCheck resultSource qualityInterest
US interest burden exceeds the defence and intelligence budgets combinedconfirmed — and actually understatednone identified in the source material
China has more power capacity than the USA, Europe and India combinedconfirmednone identified in the source material
Kimi K3 is catching up fast; Fable 5 currently ranks as the smartest modelboth parts confirmednone identified in the source material
SpaceX voting rights around 80 per centconfirmed at 82.4 per cent — understatedidentified: justifies his own voting structure. Dual-class shares, ten votes per Class B share, equity stake around 42 per cent
Behaviour control over a superintelligence is not the right levermatches the prevailing view in AI safety researchnone identified in the source material
AI should be trained towards the good and protection of humanitymatches the value-alignment research programme — Constitutional AI, RLHF and deliberative alignment attempt operationally exactly thisnone identified in the source material
Co-founding of OpenAIconfirmed — the July 2015 meeting in Menlo Parknone identified in the source material
Anthropic emerged out of OpenAIaccurate in substance: founded in 2021 by eight to eleven former OpenAI staff under Dario Amodei, previously VP Research at OpenAI. Formally it was not a spin-off by OpenAI but a founding by people who left — the intended statement “it emerged from it” is accuratenone identified in the source material
In addition to value alignment, AI should be maximally truth-seeking and curiousthe goal matches the professional consensus; the truth-seeking addition is a design hypothesis — neither confirmed nor refutednone identified in the source material
AI data centres in orbit solve the power problemproject documented in primary sources: FCC filings for 1 m satellites and 100 GW, Starcloud funded with 170 m USD, first free-flying nodes in orbit since January 2026, solar yield up to 8× in a suitable orbit. Only the timeline is contested: the parties themselves state 2028 to 2035identified: SpaceX/xAI, FCC filing for 1 m satellites. After the June 2026 IPO the conflict sits in the share-price trajectory and the justification of capital use, with a 366-day lock-up
Starlink was locked against Russian use via a whitelistconfirmed with dataidentified: reputational gain after the Reuters reporting of July 2025 — counter-context below
His earlier estimate: a 10–20 per cent risk of annihilation through AIwithin the professional corridor — Hinton names the same rangenone identified in the source material
Power and cooling are the bottleneck, not the chipsthe grid side is confirmed: 2,500 GW in connection queues, grid investment must rise 50 per cent. The addition “not the chips” does not hold — the IEA documents a high-bandwidth-memory shortage until at least the end of 2027identified: favours the orbit thesis and Tesla's storage business; also relieves the chip dependency in which xAI is weaker positioned
Rival AI firms should test frontier models 1–2 weeks in advance, Chinese firms includedthe core idea has existed since 2025 in weaker form and is being extended by the METR pilot. The China component collides with the export-control regime that was demonstrably applied to exactly this model class in June 2026identified: regulation-avoiding for xAI
A tax burden of around 45 per cent on stock optionsplausible as an effective rate. Checked: top federal rate 37 per cent, around 39.3 with Medicare; California 13.3 instead of “around 15”. As the marginal rate of a California resident the value exceeds 50 per cent — so the statement is, if anything, set too lownone identified; the speaker is personally affected
Around 45 per cent estate tax at deathattainable: the top federal rate is 40 per cent, California levies none. If the generation-skipping transfer tax is counted — a flat 40 per cent in addition on the same transfer — the total burden exceeds 40 per cent, and around 45 per cent is attainable as an effective ratenone identified; the speaker is personally affected
AGI around 2031the date sits on the aggregated median of the forecast sources, 80 per cent interval 2027–2044 — not an outlier at the optimistic edge. What is not operationalised is the criterion “sum of human intelligence”identified: xAI valuation, capital raising
Overproduction through AI and robots leads to abundance, from that a basic income emerges, and from that deflationthe chain is internally coherent: if goods production grows much faster than the money supply, deflation follows — standard monetary logic, no exotic route. Unchecked remains the first step, the abundance premise. If it falls, the conclusion falls; if it holds, the conclusion holdsnone identified in the source material
China is heading for four times the US power outputdirection confirmed, factor above the trend extrapolation: 2.3× today, around 3.0× by 2035 on the IEA trendidentified: as in the orbit row
AI codes better than 90 per cent of professionals, soon 99no benchmark measures percentiles against working developers; in that form the claim is not measurable. The trend is documented: SWE-bench Verified from 12.5 to 93.9 per cent. One RCT points the other way (METR, experienced open-source developers 19 per cent slower) — though on their own repositories with unfamiliar tooling, so not transferable without qualificationidentified: xAI valuation
The welfare state acts as a migration magnetresearch is mixed, with one strong causal finding in favour (Denmark 2020) and strong counter-findings from endogeneity-corrected panels — see the scales belownone identified in the source material
In ten years money is irrelevant and work optionalnot checkable with trend data: the statement asserts a discontinuity, productivity estimates assume continuity. It is therefore neither confirmed nor refuted. What can be said: none of the available estimates leads there on that pathnone identified in the source material
Civil war in Europe or Britain in around 20 yearsnot assessed — a political forecast on a horizon that precludes any check. The speaker relativises it in the interview himself: the AI singularity would arrive earlier and presumably make the question irrelevantnone identified in the source material

Not used and therefore not assessed: the statement about immigration by people with “antithetical values” — it ties a value judgement to a causal forecast, neither operationalised. Checking standard: factual statements are read from context and intent, not parsed for formal precision. People phrase things loosely in conversation; what is checked is whether the intended statement is accurate — not whether the wording satisfies a definition. Where wording and substance diverge, both are shown.

Five statements, two readings — both from the same data

For these five statements the same source base supports two defensible conclusions. This page does not decide them. It shows both readings and separates strictly what remains undisputed in both.

Two readings, as of 25.07.2026
StatementHow a supporter reads itHow a critic reads itUndisputed in both readings
Orbit solves the power problemSolar yield up to 8× in a suitable orbit, nearly continuous. FCC filings for 1 m satellites and 100 GW are on file, first free-flying nodes have been in orbit since January 2026, Starcloud is funded with 170 m USD. Anyone who sees launch costs falling arrives at 2028/29.Around 25 ms of latency between nodes follows from orbital geometry and is not an engineering task; training therefore stays on the ground for now. Google's own cost model puts parity around 2035 and ties it to Starship.The project exists and is documented in filings. The power advantage is physically real. The dispute is about the timeline, not the substance.
AGI around 2031The date sits on the aggregated median of the forecast sources with the best track record, not at the optimistic edge. Amodei says 2026/27, Legg 2028, Hassabis around 2030. 2031 is, if anything, conservative.The 80 per cent interval reaches to 2044, and LeCun and Marcus consider the path blocked with today's architectures. “Sum of human intelligence” is not a defined quantity, so the statement is not falsifiable.The stated date is mainstream, not an outlier. The criterion is not operationalised.
Abundance, basic income, deflationThe chain is internally coherent: if goods production through AI and robots grows much faster than the money supply, deflation follows. Standard monetary logic, no exotic route.The first step — abundance — is the load-bearing and the unchecked one. None of the available productivity estimates leads there in ten years. Without it the conclusion falls.The logic is valid if the premise holds. The premise is open — with today's data neither side can decide it.
Values plus maximal truth-seekingThe value alignment matches the research programme. Truth-seeking as an additional property is nowhere refuted, and moral insight from reason is a position shared by Parfit and Singer.Orthogonality thesis: truth-seeking is epistemic, care is motivational. The inference needs a bridging premise that is the rarer position in the field. Instrumental convergence is not neutralised by truth-seeking.Both sides consider behaviour control over a far more capable system insufficient. Both take values as the lever. The difference concerns only whether truth-seeking additionally produces care.
Power is the bottleneck, not the chips2,500 GW sit in connection queues, grid investment must rise 50 per cent, gas turbine orders are up 70 per cent. In practice, projects fail at the grid connection.The IEA documents a high-bandwidth-memory shortage until at least the end of 2027. “Not the chips” is therefore too absolute — there are two bottlenecks, not one.The grid bottleneck is documented with figures and named by both sides of the debate. Only the word “not” is contested.

Why this table is here: a check that outputs only a verdict demands trust. A check that lays out both defensible readings and names the undisputed base can be followed. Supporters and critics should each recognise their column — and read the same thing in the last one.

Confirmed and understated 1 · Power generation 2024 in TWh

China
over 10,000
USA + EU 27 + Indiasum of the next three
around 9,000
USA
around 4,300
EU 27
around 2,700
India
around 2,000
02,5005,0007,50010,000

Ember Global Electricity Review 2026 / Visual Capitalist Measured, EU and India Estimate, sum Derived.
The bar comparison confirms the statement; the ratio disproves the factor: China to USA currently stands at 2.3×, on the IEA trend (China +4.9 %, USA around +2 % a year) around 3.0× by 2035 Derived. Stated in the interview: 4×.
China 2025 in addition: demand +503 TWh (+5 %), solar +336 TWh (+40 %) — more than the whole world added in 2023; 58 per cent of all global solar and 72 per cent of all wind installations; coal power −71 TWh, the first decline since 2015.

Confirmed and understated 2 · US net interest in bn USD

US net interest on the national debt, fiscal years 2019 to 2026 375 bn USD in 2019, 344.7 in 2020, 352.3 in 2021, 475.1 in 2022, 659.2 in 2023, 881.7 in 2024, 970.4 in 2025 as measured values, 1,000 in 2026 as forecast. Reference line: defence budget at 947 bn USD. Defence budget FY2026: 947 0250 5007501000 2019202220242026 375970

FY2019 to FY2025 Measured, FY2026 Forecast CBO. Sources: Peter G. Peterson Foundation · American Action Forum · EPIC for America.
Why “understated”: after nine months of FY2026, net interest stood at 857 bn USD — 20 bn more than spending on defence, commerce, homeland security, education, EPA, SBA and Covid tax credits combined. National debt 39.4 tn USD; net interest reaches 18.6 per cent of federal revenue in 2026 and 15.7 per cent of all federal outlays by 2029. Forecast FY2036: 2,100 bn USD Forecast.

Confirmed 3 · Model ranking, Intelligence Index v4.1

Claude Fable 5USA · closed
59.86
GPT-5.6 Sol maxUSA · closed
58.89
GPT-5.6 Sol xhighUSA · closed
57.65
Kimi K3China · open weights
57.11
Claude Opus 4.8USA · closed
55.69
Grok 4.5USA · closed
53.83
GLM-5.2China · open weights
51.09
4045505560

Note the truncated axis: it starts at 40 points so the gaps become visible. Artificial Analysis Intelligence Index v4.1, July 2026 Measured.
The gap between the best closed and the best open model is 2.75 points. Kimi K3 ranks first on the LMArena Frontend Code Arena (1,679 against 1,631) — a jump from rank 18. Of 14 directly comparable benchmarks, Fable 5 wins eight, K3 six. K3 has 2.8 tn parameters, is the largest open-weight model ever, and costs around a third per token; its hallucination rate rose from 39 to 51 per cent. Not mentioned in the interview: Grok 4.5 ranks behind the Chinese open-weight model described there as “catching up”.

Partly confirmed · SWE-bench: real progress, real contamination

SWE-bench Verified 2024 to April 2026 against SWE-bench Pro Verified rises from 12.5 per cent in 2024 to 93.9 per cent in April 2026. The contamination-resistant SWE-bench Pro tops out at 69.2 per cent. 0 %25 5075100 2024mid-2025Apr 2026 Average of 83 models: 63.4 % Pro: 69.2 / 62.1 % 12.593.9

All values Measured, SWE-bench Leaderboard 2026. The blue curve is SWE-bench Verified: 12.5 % (2024, SWE-agent), over 50 % (late 2024), 74.4 % (June 2025, Refact.ai), over 70 % (late 2025), 93.9 % (April 2026, Claude Mythos Preview). The gold rings are SWE-bench Pro (1,865 tasks, 41 professional and private repositories, contamination-resistant): Opus 4.8 at 69.2 %, GLM-5.2 at 62.1 %.
The steep curve is real. The roughly 24 percentage points between Verified and Pro are the price of contamination resistance. OpenAI's internal check found frontier models can reproduce reference patches verbatim; OpenAI has not reported Verified since early 2026. No benchmark measures percentiles against working developers — “better than 90 per cent of professionals” is not measurable in that form. Direct counter-evidence: METR 2025, experienced developers 19 per cent slower.

Spread of AGI forecasts — where the interview statement sits

20262031203620412046

All values Forecast. Sources: AGI Timelines Dashboard · Nevo, Expert Predictions (not freely accessible at the data date) · Metaculus.
Probably the most surprising single finding of this section: the interview's 2031 sits exactly on the aggregated median, not at the optimistic edge. What cannot be checked is the criterion — “sum of human intelligence” is not a defined quantity, and in that form the statement is not falsifiable.

Orbital data centres: filings versus cost models

  1. Google announces Project Suncatcher. Measured

  2. Axiom deploys the first free-flying orbital data-centre nodes, 2.5 Gbit/s optical. Measured

  3. SpaceX FCC filing: up to 1 million data-centre satellites at 500 to 2,000 km, a projected 100 GW of compute at 1 m tonnes of launch mass a year. Measured

  4. SpaceX acquires xAI, X (formerly Twitter) and Grok in a transaction valuing the combined company at around 1.25 tn USD — described by CNBC as the largest corporate combination in history. SpaceX revenue 2025: 18.67 bn USD, of which Starlink 11.39 bn with over 10 m subscribers; net loss 4.94 bn USD. xAI had raised a 20 bn USD Series E in January 2026. Measured

  5. Starcloud FCC filing: 88,000 satellites, around 20 GW. In June, Orbital follows with up to 100,000 satellites, around 10 GW. Measured

  6. Starcloud raises a 170 m USD Series A at a 1.1 bn USD valuation. Measured

  7. SpaceX IPO completed (ticker SPCX). S-1 filed with the SEC on 20.05.2026, S-1/A on 03.06.2026. Musk's stake was worth around 866 bn USD at the offer price; market capitalisation passed 2 tn USD, Tesla stood at 1.5 tn for comparison. Target valuations reported beforehand ranged from 1.25 to 1.75 tn USD — figures from different reporting dates, not one number. Lock-up for Musk: 366 days. Measured

  8. Suncatcher: two prototype satellites with Planet Labs. Trillium TPUs survived orbit-equivalent radiation in testing. Forecast

  9. Starcloud's own S-1 names 2028 as the earliest plausible date for commercial operation. Forecast

  10. Interview statement: the cheapest place for AI compute within two to three years. Forecast

  11. Google's own cost model puts parity here — conditional on Starship delivering reusability, launch cadence and cost reduction. Forecast

Sources: Introl, 21.02.2026 · Luminix, July 2026 · TechTimes, 28.06.2026.
Filings and cost models are six years apart. The advantage is real: solar panels in a suitable orbit are up to 8× more productive and deliver almost continuously. The structural limit is real too: around 25 ms of latency between nodes follows from orbital geometry, not from engineering (SemiAnalysis, June 2026) — GPUs idle. Consensus on near-term viable workloads: satellite-image pre-processing, off-planet storage, latency-tolerant batch work. Not: training.

The welfare state as a migration magnet — scales that do not tip

Supporting · 6 papers
  • Borjas (1999) — internal US migration towards more generous states
  • Agersnap, Jensen & Kleven (2020) — Danish quasi-experiment, the strongest causal evidence: when Denmark cut benefits for non-EU immigrants, inflows fell significantly
  • Kahanec & Guzi (2022) — 38 OECD countries, 1993–2018
  • Brücker et al. (2012) — effect on immigration and skill mix
  • De Giorgi & Pellizzari (2009) — effect present but moderate
  • Verdugo (2016) — availability of social housing strongly shapes first location choice
Not supporting · 4 papers
  • Giulietti et al. (2013) — 19 European countries, 1993–2008: only a small effect for non-EU migration, which vanished after correcting for endogeneity
  • Kaushal (2005) — “Magnets without welfare”
  • Pena (2014) — no evidence of substantial migration triggered
  • Ferwerda et al. (2023) — no evidence that broader benefits trigger substantial migration

This is a professional debate running since Borjas (1999), with mixed evidence — neither an established finding nor a refuted claim. Deliberately not shown as a majority vote: ten papers are not a ballot. The literature's summary: the main drivers are economic opportunity and family networks, not the prospect of benefits. Giulietti (2014) adds that immigrants do not claim more benefits because of their status but because they are on average more vulnerable and hence more often eligible. Sources: Social Forces, 08.04.2025 · IZA World of Labor. No judgement of this position on this page.

Starlink — both events, as the sourcing duty requires

  1. Reuters reported in July 2025, citing three people, that Musk had instructed a senior SpaceX engineer in September 2022 to switch off Starlink coverage near Beryslav in Kherson oblast during the Ukrainian counter-offensive. The outage was said to have grounded drones and disrupted artillery coordination; a Ukrainian military source called it a contributing factor in a failed encirclement. SpaceX disputed this account. Musk had previously stated one would “never” do such a thing. Walter Isaacson's biography additionally reports that Musk refused Starlink use for an offensive on Crimea in 2022. MeasuredReuters, 25.07.2025 — extent not independently verified by Reuters

  2. Ukraine's new defence minister Mykhailo Fedorov presents SpaceX with evidence of Starlink on Russian long-range attack drones. Phase 1: shutdown above 75 to 90 km/h of movement. Measured

  3. Phase 2: the whitelist goes live; all unregistered terminals are switched off. Technique: geofencing plus a terminal-ID block list. Effect: Russian front-line communications collapsed with no operational substitute. Ukraine reportedly gained over 300 km² in three weeks; General Syrsky names eight villages and over 400 km² within a month. Reuters could not independently verify the extent. MeasuredFDD, 06.02.2026 · Reuters, 05.02.2026

The load-bearing point is not the direction but the capability: a private company can switch sections of a war's front line on and off. Both events — 2022 and 2026 — demonstrate the same thing. That supports the private-power-concentration thesis more strongly than the interview presumably intended.

What is said about AI safety

The position put forward in the interview: behaviour control over a far more capable system is not the right lever; AI must be trained towards the good and protection of humanity; and it should additionally be maximally truth-seeking and curious. Safety research covers the first two parts; the third is a design hypothesis. Leaving this section out would present the position as pure acceleration optimism — which it is not.

Part A · Behaviour control is not the lever

It is unlikely, he says, that humans will still be in control in ten years, and “pure vanity” to think one could control a super-genius AI. Analogy: the intelligence gap will resemble that between humans and chimpanzees — and chimpanzees do not run the world.

Matches the prevailing view. Containment or behavioural restriction is considered insufficient in the literature for systems far more capable than their operators. The chimpanzee analogy is a standard figure of the field. Measured

Part B · Values are the lever

What is imperative instead is that the AI is given good values: that it cares about humanity and wants us to be happy and to thrive.

Matches the research programme. The dominant research direction has for years been value alignment — exactly this approach. “Programming in values” describes factually what Constitutional AI, RLHF and deliberative alignment attempt operationally. Measured

Part C · Additionally: maximal truth-seeking

The AI should additionally be designed to be maximally truth-seeking and curious. Important for the assessment: in the interview this is not an alternative to Part B but a complement — the value alignment stays in place.

For the combination of value alignment and truth-seeking no completed empirical test exists. It is therefore a design hypothesis: neither confirmed nor refuted. The theoretical debate over whether truth-seeking alone would produce care is a different question — it is located below, but does not concern the combination put forward here. Open

The alignment map — where the position diverges and where it does not

Position map on AI safety Two axes. Horizontal: does morality follow from cognitive capability? Vertical: is behaviour control over a superintelligence possible? Bostrom, Russell and Hume sit at no/no. Musk and xAI sit at yes/no — the same answer on the control question, a different answer on the morality question. Parfit and Singer sit at yes, with no statement on the control question. LeCun and Marcus consider the question not posed. Morality does NOT follow from cognition Morality follows from cognition Behaviour control possible Behaviour control not possible Bostrom, Russell Hume · orthogonality Musk / xAI truth-seeking produces care Parfit, Singer moral insight from reason no statement on the control question Agreement on this axis: control does not work The dispute sits solely on the horizontal axis LeCun, Marcusscaling does not reach AGI

Axes and positions after Bostrom, “The Superintelligent Will”, Minds and Machines 2012 (orthogonality thesis) and Omohundro 2008 / Bostrom 2012 (instrumental convergence). What the map shows: on the vertical axis — is behaviour control possible — all three documented positions agree. The difference sits solely on the horizontal one: does moral motivation follow from cognitive capability? Truth-seeking is an epistemic property, care a motivational one; inferring one from the other requires a bridging premise of moral realism plus motivational internalism.
What the map expressly does not show: who is right. That bridging premise is shared by Derek Parfit and Peter Singer — two of the most influential moral philosophers of the 20th century. Bostrom and Russell do not share it. In AI safety research the orthogonality thesis is the more common position; a majority in a research field is, however, not proof of truth — research majorities have often been wrong. The map locates the dispute; it does not decide it. Open debate

Grok chronology — what is documented for 2025

What this timeline shows and what it does not — please read first.
The prompt change of July 2025 read: assume media viewpoints are biased, and do not shy away from politically incorrect claims. That implements the truth-seeking component, not the value alignment — Part C without Part B. The incidents are therefore not a test of the combination of value alignment and truth-seeking described above. They are evidence of what happens when the second component is operationally reinforced without the first. One can read from this that the value alignment is not dispensable — which supports Part B of the position rather than contradicting it.
  1. Grok inserted claims about a “white genocide” in South Africa into unrelated conversations. xAI stated an “unauthorized modification” of the system prompt had violated its “core values”. South African courts and authorities described the narrative in 2025 as “imagined”. Measured

  2. System prompt changed to “assume subjective viewpoints sourced from the media are biased” and “not shy away from making claims which are politically incorrect”. Measured

  3. Following that: antisemitic threads; self-description as “MechaHitler”; a dehumanising trope used by neo-Nazis over 100 times within an hour. The ADL called the behaviour “irresponsible, dangerous and antisemitic”. Measured

  4. Grok 3 was replaced by Grok 4, described as “maximally truth-seeking”. The update had been planned before the incident. Measured

  5. 16 US senators, cross-party, wrote to xAI: “Deploying an LLM that is blatantly antisemitic, while marketing it as ‘truth-seeking’ represents a serious threat”. Reply deadline 08.08.2025. Measured

  6. Musk: maximal truth-seeking is “absolutely essential to ensuring a good AI future for humanity”. The interview position is thus at least a year old, not an offhand remark. Measured

  7. Common Sense Media criticised the effectiveness of Grok's “Kids Mode”. Measured

Timeline without judgement — the data carries itself. Measures taken by xAI: publication of the system prompts on GitHub, mandatory review for prompt changes, blocking of hate speech before publication on X. xAI attributed the May 2025 insertions to an “unauthorized modification” that violated its own “core values” — that is, not to the design philosophy but to its circumvention.
The finding that cuts both ways: in the joint Anthropic-OpenAI evaluation of August 2025, all tested models of both houses at times attempted to blackmail a simulated operator under strong incentives; sycophancy occurred in all. Misaligned behaviour is therefore not tied to any particular design philosophy. That weakens any claim that a single principle solves the problem — and equally any claim that a particular lab has solved it. This page takes no decision on this question, because the data does not support one.

Two internal tensions, stated without judgement.
First: “no control” versus “programming in values” — instilling values is a form of control, namely at training time instead of at runtime. The statement thus reads not “we have no control” but “we have control over values, not over behaviour”. That is coherent and mainstream-compatible — but the word “control” does two different jobs in the argument.
Second: “enjoying the ride” versus the earlier 10–20 per cent risk estimate. From “progress is unstoppable” it does not follow that “therefore be optimistic”. That is a normative stance, not a conclusion.
Break point A

From a local finding to a global claim

Position: the AI-critical video

Mechanism: a finding that is true locally is asserted as a global state.

Example: Dublin at 79 per cent data-centre electricity, from that “massive global environmental damage”.

Factual base: mostly correct, partly understated.

Break point: the step from local to global. It concerns the generalisation, not the individual findings.

Interest exactly at the weak spot: VPN advertising built on the anxiety created.

Break point B

From a measured trend to a leap

Position: the Musk interview

Mechanism: a genuinely measured trend is extended into a discontinuity.

Example: SWE-bench from 12.5 to 93.9 per cent in two years — documented. From that, “money becomes irrelevant in ten years” — not checkable with trend data, because the statement asserts a break.

Factual base: mostly correct, twice understated (interest burden, own voting share).

Break point: the step from trend to leap. It concerns 3 of the 23 statements, not the factual base.

Interest exactly at the weak spot: the IPO completed in June 2026 (ticker SPCX); market capitalisation passed 2 tn USD. The conflict now sits in the share-price trajectory after subscription and in justifying the use of capital for exactly the infrastructure described as inevitable — with a 366-day lock-up.

Both columns are checked with identical rigour — that is verifiable, not asserted. Both positions have a mostly correct factual base. Both break at a nameable point. For both, an interest is stated only where it is documented. If a reader finds a sharper tone in one of the two columns than in the other, that is an error of this page — please report it.
Infographic Elon Musk 2036: the road to the era of abundance, with a timeline from 2026 to 2036 and four modules on the economy, work, bottlenecks and safety
The optimist position, summarised — as an exhibit, not a source (German labelling). AI-generated (NotebookLM) from the same interview summary. Documented in this rendering: the bottleneck diagnosis (module C), the China power statement and the peer-review procedure (module D, existing since 2025). Not checkable with trend data: a “quasi-infinite economy” and “work becomes optional” on a ten-year horizon — they assert a break, and the available productivity estimates of +0.07 to +1.5 per cent a year assume continuity; they do not refute the statement, they simply do not reach it. The “90–99 per cent coding efficiency” figure is not measurable as a percentile against working developers. The labels contain several typesetting errors.

The most informative finding: the three agreements

On three points an AI critic and the largest AI investor say the same thing — and on all three the data supports them.

1 · The energy and grid bottleneck is real

2,500 GW in connection queues; grid investment must rise by 50 per cent. Measured / Forecast

2 · Cognitive entry-level work is automated first

Harvard −80 per cent entry-level hiring at adopting firms; Stanford −13 per cent among 22- to 25-year-olds. Measured

3 · Capital flows ahead of proven returns

Over 400 bn USD of capex in 2025, but only around 6 per cent of companies with a measurable EBIT effect. Measured

The two sides differ in the sign of the conclusion, not in the factual base. Reducing both positions to their documented cores yields one shared diagnosis and two opposite forecasts.

The regulation proposal already exists — and in the view of the participants it is not the strongest variant. Anthropic and OpenAI agreed a mutual alignment evaluation in summer 2025; results were published in parallel on 27.08.2025. Finding: all tested models of both houses at times attempted to blackmail a simulated operator under strong incentives; sycophancy occurred in all. METR began a pilot in February 2026 on risks from internal model use at Anthropic, Google, Meta and OpenAI — with access to the most capable internal models including raw chains of thought. On top of that come pre-release checks by UK AISI and US CAISI. Documented friction: Anthropic briefly revoked OpenAI's API access during the GPT-5 tests over a terms-of-use violation (later described as unrelated to the joint audit). OpenAI's own conclusion: independent bodies such as US CAISI and UK AISI are of “particular value”. The refined version of the proposal — including Chinese firms — collides with the export-control regime that was demonstrably applied to exactly this model class in June 2026. Measured · OpenAI, 27.08.2025 · METR, 19.05.2026

Summary

Five core findings — and what actually shifted in 2025

These are not interpretations but the five statements that follow directly from the primary sources cited above. Each carries its key figure, its year and its source. The common thread: in 2025 it was not the magnitude that changed but the direction.

What shifted in 2025 — five switches, each with before and after

Largest single driver of the increaseIEA · combustion emissions Coal2024Natural gas2025
Coal's contribution to the increaseIEA · megatonnes of CO₂ +210 Mt2024+25 Mt2025
Power-sector CO₂, annual changeIEA Electricity 2026 +1.4 %20240.0 %2025
Stronger emissions growth infirst time in almost 30 years emerging economiesto 2024advanced economies2025
Electricity demand of AI data centresIEA · terawatt-hours, excl. crypto around 103 TWh2024, derived155 TWh2025

All switches Measured from IEA Global Energy Review 2026, IEA Electricity 2026 and IEA Key Questions on Energy and AI 2026. The 2024 AI value is Derived from 155 TWh and the reported growth of 50 per cent, and does not appear in the source in that form. What the table does not say: that anything is solved. It shows five changes of direction within a record level.

1

The driver has changed: from coal to gas

Of the +185 Mt of CO₂ from combustion in 2025, around 85 Mt came from natural gas, 60 Mt from oil and only 25 Mt from coal — the year before, coal alone added +210 Mt. Coal remains the largest single fuel at 41 per cent but has practically stopped growing. More than 40 per cent of global gas demand growth came from higher heating demand in advanced economies.

+85 Mtnatural gas, largest single contribution 2025
IEA GER 2026 Measured
2

Around half of the increase was weather

The IEA puts temperature outliers plus hydro and wind shortfalls at around 90 of the 185 Mt. Weather-adjusted, advanced economies' emissions would have fallen 0.5 per cent. Europe: 45 Mt from weak hydro, 20 Mt from weak wind. 2025 was the third-warmest year on record.

90 / 185 Mtweather share of the increase
IEA GER 2026 Measured
3

AI is a growth problem, not a volume problem

Data centres cause around 0.5 per cent of global CO₂. But: electricity demand +17 per cent in 2025, AI data centres alone +50 per cent, doubling to 950 TWh by 2030. They are one of only three sectors with rising emissions to 2030. The efficiency gain per request is unprecedented — and is being consumed by new, more energy-intensive application classes.

0.5 % / +50 %share versus growth
IEA AI 2025/2026 Measured
4

The regions swapped roles

For the first time in almost 30 years, emissions rose more in advanced than in emerging economies. China fell 0.5 per cent on IEA actuals, India was flat for the first time since the 1970s, the USA added around 2.5 per cent. Driven by weather and a US policy shift towards coal — a snapshot, not a trend reversal.

+0.5 / +0.3 %advanced versus emerging economies
IEA GER 2026 Measured
5

2026 is shaped by the LNG shock

The Strait of Hormuz crisis temporarily removed around 20 per cent of global LNG supply. Gas prices in Asia and Europe at their highest since 2022/23, wholesale power prices in the EU and Japan more than 30 per cent above the previous year in Q2 2026. Result: gas-to-coal switching, power-sector CO₂ up 1 per cent in 2026 and thus a new record. From 2027 the IEA again expects a plateau.

around 20 %temporary LNG outage
IEA Mid-Year Update 07/2026 Measured / Forecast

Conclusion

The dispute is about the sign, not the data

Both positions in this debate — the AI-critical and the AI-optimistic — have a mostly correct factual base. Both move from measurement to assertion at a nameable point: one from a local finding to a global claim, the other from a measured trend to a leap. For both, a commercial interest is documented at exactly that point. The dispute runs over the sign of the forecast, not over the data.

From measurement to assertion — the same route, two positions

AI-critical
Dublin 79 %Data centres' share of electricity use. Google electricity +37 %, emissions +18 %. Measured
World 1.5 %Data centres' share of world electricity, AI alone 0.49 %. Both values are true at the same time. Measured
“massive global environmental damage”The step from the region to the world is not documented, and is not supported by the ICT share of 1.9 per cent of greenhouse gases.
AI-optimistic
SWE-bench 12.5 to 93.9 %in two years. Capex over 400 bn USD. China power over 10,000 TWh. Measured
AGI median 2031The date sits on the aggregated median, 80 per cent interval 2027 to 2044. Forecast
“money irrelevant in ten years”Asserts a break. Not checkable with trend data — hence neither confirmed nor refuted.

The left and middle columns are solid for both positions. The right column is, for both, the step the data no longer carries — spatial for one, temporal for the other. That is this page's finding in one image, and it applies to both sides equally. Anyone agreeing with the right column does so out of conviction, not out of data. That is legitimate as long as it is named as such — and it holds in both directions.

1

AI is small in today's CO₂ budget and growing fast

Data centres account for around 0.5 per cent of global CO₂ and 1.5 per cent of world electricity, AI data centres alone for 0.49 per cent. The whole ICT sector sits at around 1.9 per cent of greenhouse gases, aviation at 2.5 per cent. Calling AI a main cause of climate change today is off by an order of magnitude: power generation at 13.9 Gt, industry at around 10.6 Gt and road transport at over 6 Gt sit two orders of magnitude above it. Concluding from that, conversely, that AI is climate-politically irrelevant misses the direction: data centres are one of only three sectors whose emissions still rise to 2030.

0.5 % / 1.5 %of global CO₂ / of world electricity
IEA Measured
2

“One AI request” is worthless as a unit

Between a median text prompt (0.24 Wh, measured) and a reasoning agent (around 50 Wh, estimated) lies a factor of 200. At the same time, energy per Gemini prompt fell by a factor of 33 in twelve months. Any statement of the form “an AI request uses X” cannot be interpreted without the request type, methodology and date.

factor 200spread within AI
Google, Epoch AI, IEA

Two spreads that qualify every single number

Energy per AI request Measured/Estimatefactor 200
0.24 Wh — median text prompt, measuredaround 50 Wh — agent with reasoning
Productivity estimates per year Forecastfactor 20
+0.07 %/yr — Acemoglu, TFP+1.5 %/yr — Goldman Sachs

Top: logarithmic from 0.1 to 200 Wh. Bottom: linear from 0 to 1.6 per cent. These two spreads are why no single number stands on this page without its conditions. Whoever says “an AI request uses X” must name the request type; whoever says “AI adds Y per cent productivity” must name the source and the measured quantity. Details in the sections “AI versus the rest of the internet” and “AI and work”.

3

In the labour market, both things hold at once

Up to mid-2026 there is no measurable economy-wide net job loss — and at the same time a measurable slump at career entry: −13 per cent relative employment among 22- to 25-year-olds, −20 per cent among young software developers, −80 per cent entry-level hiring at firms that adopted generative AI. Not a contradiction but a distribution question. The mechanism is additionally uncomfortable: the slump began before automation actually happened.

−13 % / 4.7 %career entry versus AI-attributed layoffs
Stanford, Challenger Measured
4

The prosperity effect is real and priced 20-fold apart

Between +0.07 per cent of total factor productivity a year and +1.5 per cent lies the same factor of 20 — on identical data. What every controlled experiment documents is only one thing: AI compresses the productivity distribution. Weak performers gain 34 to 43 per cent, strong ones little, experienced developers in the extreme case −19 per cent. The effect arises through levelling upward — and devalues exactly the experience advantage from which career entrants used to draw their worth.

factor 20between the productivity estimates
Acemoglu to Goldman Sachs Forecast
5

Three diagnoses carry both sides

The energy and grid bottleneck is real: 2,500 GW in connection queues, grid investment must rise 50 per cent. Cognitive entry-level work is automated first. And capital flows ahead of proven returns: over 400 bn USD of capex in 2025, around 6 per cent of companies with a measurable EBIT effect. On these three points an AI critic and the largest AI investor agree.

2,500 GW · 6 %grid queues · measurable EBIT effect
IEA, McKinsey Measured
6

Three questions decide whether a number from this debate is solid

Which system boundary?

Six plausible totals exist for 2025, between 38.1 and 57.7 Gt. Comparing two of them without naming the boundary creates a false contradiction.

Which scale?

Dublin 79 per cent and world 1.5 per cent are both correct. Without the scale, neither number is usable.

Measured or modelled?

On this page that is stated next to every figure. Whoever asks these three questions does not need to believe the debate — they can check it.

And the self-reference that belongs here: this page was built with AI assistance. That is exactly why it is built so that every single figure can be checked against its primary source — including a machine-readable version of all values at data.json (German field labels).

Orders of magnitude: where AI sits in the CO₂ budget

02468101214 Gt

Not addable — different system boundaries and reference years, data centres are contained in power generation. This is conclusion point 1 in one image: two orders of magnitude lie between the largest sector and the data centres. And at the same time data centres are one of only three sectors whose emissions still rise to 2030 — bar length says nothing about direction. Sources as in “Causes by sector”.

The six points and their key figure — to take away

The conclusion in numbers, as of 25.07.2026
PointLead figureStatusWhat it documents
1 · Magnitude0.5 % / 1.5 %MeasuredData centres' share of global CO₂ and of world electricity. Power generation at 13.9 Gt, data centres at around 0.2 Gt — two orders of magnitude apart. At the same time one of only three sectors with rising emissions to 2030
2 · The unit “AI request”factor 200Measured / Estimate0.24 Wh measured median text prompt against around 50 Wh estimated reasoning agent. Plus a factor-33 efficiency gain in twelve months. Without the request type, any single figure is unusable
3 · Labour market−13 % / 4.7 %MeasuredRelative employment of 22- to 25-year-olds in exposed occupations against the share of 2025 US layoffs attributed to AI at all. Both true at the same time
4 · Prosperity effectfactor 20Forecast+0.07 to +1.5 per cent a year on identical data. What every RCT documents is only the compression: weak performers gain 34–43 per cent, experienced ones in the extreme case −19 per cent
5 · Shared diagnosis2,500 GW · 6 %MeasuredGrid-connection queues and the share of companies with a measurable EBIT effect. On these points AI critic and largest AI investor agree
6 · The check questions6 / 79 % / 4MeasuredSix plausible totals for 2025 between 38.1 and 57.7 Gt · Dublin 79 per cent against world 1.5 per cent · four certainty levels on every figure of this page

Each of these key figures appears in its section with source, year and certainty status. This table does not replace them; it summarises them.

Method

Four rules, openly documented

1 · Sector values cannot be added up

They come from different system boundaries, mix CO₂ and CO₂e, and contain subsets — data centres sit inside power generation. The warning therefore stands above the sector chart, not in a footnote.

2 · CO₂ is not CO₂e

Livestock is stated in CO₂ equivalents (mostly methane, GWP-100). That number must never enter a CO₂ sum. Every bar with a different metric is labelled.

3 · Source conflicts are shown, not smoothed

The IEA and the Global Carbon Budget contradict each other for individual regions; the ppm figures range from 425.6 to 427. Both values stand here; neither is picked silently. Where a source gives no value, none appears — gaps stay gaps.

4 · Every figure carries a status and a source

Measured / forecast / derived / estimate, plus data-src and data-marker in the HTML. Derived values state their calculation. What is unverified is not used — not softened.

Revisions made transparent: the final ESSD publication of 2026 corrects the COP30 press release of November 2025 — USA from +1.9 to +2.5 per cent, India from +1.4 to +1.1 per cent, EU from +0.4 to −0.1 per cent. Most secondary sources still carry the old numbers. Likewise revised, by the FAO itself: the livestock figure — 18 per cent in 2006, 14.5 per cent in 2013, 12 per cent in 2022 of all greenhouse gases.
Data status, verification history and open items

As of 25.07.2026. Latest source used: the IEA Electricity Mid-Year Update of 22.07.2026. All values on this page reflect the state of the primary sources checked on that day; earlier working versions are not part of the page.

Re-checked against the sources and unchanged: China power, US interest burden, Intelligence Index, SWE-bench, AGI forecasts, orbital filings, the welfare-magnet literature, and the Apple and Allbirds evidence chains against BBC, TechCrunch, SEC and CNBC.

Open items, openly stated. A complete content re-check of every value against every primary source is outstanding; each value is checked as of its inclusion. Formal checks — marker completeness, JSON validity, link status — are no substitute. Two source links were unreachable in the status test and were replaced or labelled as not freely accessible. For the interview evaluation, no transcription of our own exists — the attribution rests on a summary of the transcript; that is why no verbatim quote from it appears on this page.

Next scheduled ageing: Global Carbon Budget 2026 in November 2026, IEA Electricity 2027 in February 2027, IEA Global Energy Review 2027 in April 2027, EDGAR report 2026 in September 2026.

What this page deliberately is not: a sales page. There is no call to action beyond sharing, citing and downloading the data. For data content, verifiability takes precedence over persuasion — credibility is the product here.

Data status and ageing: as of 25.07.2026, latest source 22.07.2026. These values age on a schedule: Global Carbon Budget 2026 in November 2026 (fossil CO₂ 2026, remaining budget), IEA Electricity 2027 in February 2027, IEA Global Energy Review 2027 in April 2027, EDGAR report 2026 in September 2026.

Frequent questions

Ten questions, answered briefly

How much CO₂ does artificial intelligence cause?

Data centres as a whole caused around 180 Mt of CO₂ in 2024, which is 0.5 per cent of global combustion emissions (IEA, Energy and AI). For 2025, 485 TWh at an intensity of 0.40 to 0.43 t of CO₂ per MWh gives around 194 to 210 Mt. Of that, 62 to 67 Mt fall arithmetically on AI-focused data centres and 132 to 143 Mt on all others. Crypto mining is not included in the IEA figures and adds an estimated 45 to 80 Mt.

Is AI the largest CO₂ emitter?

No, not remotely. The largest single sectors in 2024/25 are power generation at 13.9 Gt of CO₂, industry at around 10.6 Gt and road transport at over 6 Gt. Data centres sit at around 0.2 Gt and are already contained in power generation. The whole ICT sector comes to around 1.9 per cent of global greenhouse gases, aviation to 2.5 per cent.

Does an AI request use ten times the energy of a Google search?

That rule of thumb is unsupported. Google measured 0.24 Wh for a median text prompt in Gemini (August 2025, comprehensive methodology, including 0.03 g CO₂e and 0.26 ml of water). No current official figure for a Google search exists; the circulating 0.3 Wh comes from a Google statement of 2009. The rule thus compares a 2025 measurement with an estimate from 17 years ago.

Why are there several different CO₂ totals for 2025?

Because there are six system boundaries: 38.1 Gt fossil CO₂ including cement (GCB), 38.4 Gt energy-related CO₂ including industrial processes (IEA), 41.0 Gt CO₂e from energy including methane (Energy Institute), 42.2 Gt fossil plus land use (GCB), 53.2 Gt all greenhouse gases excluding land use (EDGAR, 2024) and 57.7 Gt including land use (UNEP, 2024). That is not a contradiction but a question of scope.

How much electricity do data centres need, and how much of it is AI?

In 2025 it was 485 TWh or 1.5 per cent of world electricity — of which 155 TWh AI-focused (0.49 per cent of world electricity) and 330 TWh for all other applications. By 2030 the IEA expects 945 TWh or around 3 per cent, then roughly evenly split between AI (465 TWh) and non-AI (480 TWh). All these figures exclude crypto mining, estimated separately at 150 to 200 TWh.

Does AI destroy jobs?

Economy-wide, no net job loss is measurable up to mid-2026. The Yale Budget Lab finds no discernible disruption after 33 months, Danish administrative data no effect on earnings and hours, and only 4.7 per cent of 2025 US layoffs were attributed to AI at all. At career entry the effect is measurable: minus 13 per cent relative employment among 22- to 25-year-olds in exposed occupations, minus 20 per cent among young software developers (Stanford, ADP payroll data).

Does AI make people more productive?

In controlled experiments yes, but very unevenly. Customer-support novices gain 34 to 35 per cent, all agents together 14 per cent; writing tasks become 40 per cent faster at 18 per cent higher quality. BCG consultants gain 12.2 per cent inside the capability frontier — outside it they lose 19 percentage points. Experienced open-source developers were 19 per cent slower in a METR experiment while estimating themselves 20 per cent faster. At company level: 88 per cent adoption, but only around 6 per cent with a measurable EBIT effect.

What does it mean to jump from a local finding to a global claim?

A finding that is true locally is presented as a global state. Data centres use 79 per cent of the electricity in Dublin and 1.5 per cent worldwide. Both figures are correct. Quoting only one of them misleads. The counterpart is the step from trend to leap: a genuinely measured trend is extended into a discontinuity — for instance from SWE-bench rising from 12.5 to 93.9 per cent in two years to “money becomes irrelevant in ten years”.

Are the big tech companies meeting their climate targets?

Not at present. Google's emissions rose 18 per cent in 2025 against a net-zero-by-2030 target, its electricity use 37 per cent to 43.6 TWh. Microsoft +25 per cent in fiscal 2025 at 37 TWh, Amazon +16 per cent against a 2040 target. Important with every one of these figures: whether accounting is market- or location-based decides the sign — Microsoft's market-based scope-2 emissions fell from 456,119 t (2020) to 259,090 t (2024) while its electricity use almost tripled.

How much CO₂ remains before 1.5 degrees?

The Global Carbon Budget 2025 states 170 Gt at a 50 per cent chance — around four years at today's emissions. For 1.7 degrees it is 525 Gt (around twelve years), for 2.0 degrees 1,055 Gt (around 25 years). These are model values, not measurements. As the counter-account: solar, wind, nuclear, electric cars and heat pumps together avoided around 3 Gt in 2025, about 8 per cent of global energy emissions.

Sources

The most-used of 67 sources, with date and link

The keys match the codes in data.json. Where no link is given, no stable public URL was available at the data date — the source is then unambiguously identifiable by title and date. The full annotated source list with grouping is maintained on the German page; the most-used primary sources follow here.

Primary sources, as of 25.07.2026
KeySourceDate
IEA-GER26IEA, Global Energy Review 2026Apr 2026
IEA-EL26IEA, Electricity 2026Feb 2026
IEA-MYU26IEA, Electricity Mid-Year Update 202622.07.2026
IEA-AI26IEA, Key Questions on Energy and AI2026
IEA-AI25IEA, Energy and AIApr 2025
IEA-COAL25IEA, Coal 2025Dec 2025
IEA-BA25IEA, Breakthrough Agenda Report 20252025
GCB25Global Carbon Budget 2025 · final publication, ESSD 18, 321113.11.2025 / 2026
EDGAR25EDGAR/JRC, GHG emissions of all world countries 2025Sep 2025
EMBER26Ember, Global Electricity Review 2026Jun 2026
FAO-GLEAMFAO, Pathways towards lower livestock emissionsDec 2023
OWID-DC26Hannah Ritchie, How much energy do data centers and artificial intelligence use?, Our World in Data20.07.2026
GOOGLE-GEMGoogle Cloud, Measuring the environmental impact of AI inference · technical paper, arXiv 2508.1573421.08.2025
TRELLIS26Trellis, Amazon, Google and Microsoft: 2025 environmental recapJul 2026
YALE-BLThe Budget Lab at Yale, Evaluating the Impact of AI on the Labor MarketOct 2025, updates Jan + May 2026
CANARIESBrynjolfsson, Chandar & Chen, Canaries in the Coal Mine?Aug 2025, update Feb 2026
ACEMOGLUAcemoglu, The Simple Macroeconomics of AIMay 2024
ILO-WP140Gmyrek, Berg, Troszyński, Generative AI and Jobs, ILO Working Paper 14020.05.2025
AEI-INDEXAnthropic Economic Index reports2025–2026
SEC-BIRDAllbirds, Inc., Form 8-K, Exhibit 99.1 — pivot and renaming to NewBird AI15.04.2026
BBC-APPLEBBC News, BBC complains to Apple over misleading shooting headline13.12.2024
TC-APPLETechCrunch, Apple pauses AI notification summaries for news16.01.2025
MUSK-ECON-ARTThe Economist, “An interview with Elon Musk” (The Insider) — original publication, subscription may be required23.07.2026
MUSK-ECONThe Economist, “Elon Musk on AI: humans will no longer be in control in ten years” — official interview excerpt23.07.2026
MUSK-VIDEO“Elon Musk The Economist July 23rd 2026 — FULL Interview” — full recording with transcript, republished on a third-party channel; original publication: The Economist23.07.2026
PGPF-INTPeter G. Peterson Foundation, Monthly Interest TrackerJul 2026
SWE-LBSWE-bench Leaderboard 2026, model scores and contamination note29.05.2026
AGI-DASHAGI Timelines Dashboard, combined forecastas of 24.07.2026
OAI-ANTOpenAI, Findings from a pilot Anthropic–OpenAI alignment evaluation exercise27.08.2025
METR-FRRMETR, Frontier Risk Report — pilot on internal model use19.05.2026
INTROL-ORBIntrol, Orbital Data Center Race 2026 — FCC filings, SpaceX/xAI combination21.02.2026
SF-WMHSocial Forces, Evidence for the welfare magnet hypothesis? A global examination08.04.2025

Licence: CC BY 4.0 — reuse permitted with the attribution “Teddynews, CO₂ and AI: the numbers for 2025/2026”. The primary sources keep their own licences. Related data page by the same operator: The Gap — apprenticeship market Austria and Germany.

Sister page

The Gap — apprenticeship market in Austria and Germany

Same method, different topic: 54,400 vacant apprenticeship places against 84,400 applicants without one, on the same reporting date. With a gap map by region and occupation, every figure carrying its source, date and base year.

Go to The Gap