From experiment to everyday use.
AI is now used in at least one business function by a large majority of surveyed organisations. But adoption is not the same as scaling the technology across the enterprise.
Share, %
Survey. The definition has changed; 2018–2022 are omitted, not interpolated. Historical dip to 50%.
View data points and sources
| Period | Value | Source |
|---|
| 2017 | 20 Share, % Survey. The definition has changed; 2018–2022 are omitted, not interpolated. Historical dip to 50%. | Original ↗ |
| 2023 | 55 Share, % Survey. The definition has changed; 2018–2022 are omitted, not interpolated. Historical dip to 50%. | Original ↗ |
| 2024 | 78 Share, % Survey. The definition has changed; 2018–2022 are omitted, not interpolated. Historical dip to 50%. | Original ↗ |
| 2025 | 88 Share, % Survey. The definition has changed; 2018–2022 are omitted, not interpolated. Historical dip to 50%. | Original ↗ |
A habit for over a billion people.
ChatGPT’s public milestones show that demand extends far beyond the tech industry. This is usage at scale — not a measure of paying customers.
Million WAU
Public milestones, not continuous monthly measurements. Company figures reported in the press.
View data points and sources
| Period | Value | Source |
|---|
| Mar 25 | 500 Million WAU Public milestones, not continuous monthly measurements. Company figures reported in the press. | Original ↗ |
| Aug 25 | nearly 700 Million WAU Public milestones, not continuous monthly measurements. Company figures reported in the press. August 2025: nearly 700 million, not an exact value. | Original ↗ |
| Oct 25 | 800 Million WAU Public milestones, not continuous monthly measurements. Company figures reported in the press. | Original ↗ |
| Feb 26 | 900 Million WAU Public milestones, not continuous monthly measurements. Company figures reported in the press. | Original ↗ |
| Jul 26 | 1,000 Million WAU Public milestones, not continuous monthly measurements. Company figures reported in the press. | Original ↗ |
| Sep 26 | 1,200 Million WAU Public milestones, not continuous monthly measurements. Company figures reported in the press. | Original ↗ |
Same capability. A fraction of the price.
Inference costs at GPT-3.5-level capability fell more than 280-fold. Cheaper tasks unlock applications that previously did not make economic sense.
Million tokens / USD
GPT-3.5-level MMLU performance. Price: $20.00 → $0.07 per million tokens. Rounded values.
View data points and sources
| Period | Value | Source |
|---|
| Nov 2022 | 0.05 Million tokens / USD GPT-3.5-level MMLU performance. Price: $20.00 → $0.07 per million tokens. Rounded values. | Original ↗ |
| Oct 2024 | 14.3 Million tokens / USD GPT-3.5-level MMLU performance. Price: $20.00 → $0.07 per million tokens. Rounded values. | Original ↗ |
Demand is starting to pay.
Anthropic’s quarterly revenue makes the bull case more tangible. But rising revenue is not profit, and company accounting needs careful comparison.
Billion USD
Quarterly revenue, not ARR or profit. Official fact sheet. No interpolation between quarters.
Not financial advice. Revenue and investments are not promises of returns.
View data points and sources
| Period | Value | Source |
|---|
| Q1 25 | 0.376 Billion USD Quarterly revenue, not ARR or profit. Official fact sheet. No interpolation between quarters. | Original ↗ |
| Q2 25 | 0.787 Billion USD Quarterly revenue, not ARR or profit. Official fact sheet. No interpolation between quarters. | Original ↗ |
| Q3 25 | 1.382 Billion USD Quarterly revenue, not ARR or profit. Official fact sheet. No interpolation between quarters. | Original ↗ |
| Q4 25 | 2.04 Billion USD Quarterly revenue, not ARR or profit. Official fact sheet. No interpolation between quarters. | Original ↗ |
| Q1 26 | 4.731 Billion USD Quarterly revenue, not ARR or profit. Official fact sheet. No interpolation between quarters. | Original ↗ |
| Q2 26 | 11.555 Billion USD Quarterly revenue, not ARR or profit. Official fact sheet. No interpolation between quarters. | Original ↗ |
| Q3 26 | 16.956 Billion USD Quarterly revenue, not ARR or profit. Official fact sheet. No interpolation between quarters. | Original ↗ |
Capability keeps pushing the frontier.
Stanford reports major progress on Terminal-Bench. We show the published series rather than the plan’s approximate SWE-bench figures. Test success still does not guarantee robust everyday performance.
Terminal-Bench, %
Terminal-Bench according to Stanford 2026. Benchmark results are not enterprise ROI; test versions and tasks may change.
View data points and sources
| Period | Value | Source |
|---|
| Feb 2025 | 20 Terminal-Bench, % Terminal-Bench according to Stanford 2026. Benchmark results are not enterprise ROI; test versions and tasks may change. | Original ↗ |
| Early 2026 | 77.3 Terminal-Bench, % Terminal-Bench according to Stanford 2026. Benchmark results are not enterprise ROI; test versions and tasks may change. | Original ↗ |
Revenue on a different scale.
OpenAI’s CFO describes ARR exceeding $20 billion during 2025. A measure of the current revenue pace — not booked annual revenue or profitability.
ARR, billion USD
Annualized recurring revenue, not booked annual revenue. 2025: more than $20 billion; the bar shows the lower bound.
Not financial advice. Revenue and investments are not promises of returns.
View data points and sources
| Period | Value | Source |
|---|
| 2023 | 2 ARR, billion USD Annualized recurring revenue, not booked annual revenue. 2025: more than $20 billion; the bar shows the lower bound. | Original ↗ |
| 2024 | 6 ARR, billion USD Annualized recurring revenue, not booked annual revenue. 2025: more than $20 billion; the bar shows the lower bound. | Original ↗ |
| 2025 | > 20 ARR, billion USD Annualized recurring revenue, not booked annual revenue. 2025: more than $20 billion; the bar shows the lower bound. | Original ↗ |
Infrastructure shows up in the reports.
Nvidia’s revenue shows that demand for compute is physical and measurable. Data centres generated $193.7 billion in FY2026. This does not establish returns for end users.
Billion USD
The company’s fiscal year, not the calendar year. FY2026 ended 25 January 2026. Data centre revenue: $193.7 billion in FY2026.
Not financial advice. Revenue and investments are not promises of returns.
View data points and sources
| Period | Value | Source |
|---|
| FY2025 | 130.5 Billion USD The company’s fiscal year, not the calendar year. FY2026 ended 25 January 2026. Data centre revenue: $193.7 billion in FY2026. | Original ↗ |
| FY2026 | 215.9 Billion USD The company’s fiscal year, not the calendar year. FY2026 ended 25 January 2026. Data centre revenue: $193.7 billion in FY2026. | Original ↗ |
Individual benefits are documented.
NBER’s customer-service study shows more resolved cases per hour, especially for less experienced workers. The Copilot study found 55.8% faster work on one JavaScript task — not across all software development.
More cases / hour, %
Customer-service study, NBER 2023. Approximately 34% for less experienced workers. Context-specific results, not general promises of returns.
View data points and sources
| Period | Value | Source |
|---|
| All | 14 More cases / hour, % Customer-service study, NBER 2023. Approximately 34% for less experienced workers. Context-specific results, not general promises of returns. | Original ↗ |
| Less experienced | 34 More cases / hour, % Customer-service study, NBER 2023. Approximately 34% for less experienced workers. Context-specific results, not general promises of returns. | Original ↗ |
The buildout is also a risk.
Planned investment by four technology giants has grown sharply. The bull sees demand. The bear sees capital that must earn a return. Both are reading the same bars.
Billion USD
Four hyperscalers. More than $380 billion expected in 2025; at least $720 billion planned in 2026. Guidance, not final outcomes. Lower bounds. The 2025 compilation mixes calendar years with Microsoft’s FY2026.
Not financial advice. Revenue and investments are not promises of returns.
View data points and sources
| Period | Value | Source |
|---|
| 2025 | > 380 Billion USD Four hyperscalers. More than $380 billion expected in 2025; at least $720 billion planned in 2026. Guidance, not final outcomes. Lower bounds. The 2025 compilation mixes calendar years with Microsoft’s FY2026. | Original ↗ |
| 2026 | ≥ 720 Billion USD Four hyperscalers. More than $380 billion expected in 2025; at least $720 billion planned in 2026. Guidance, not final outcomes. Lower bounds. The 2025 compilation mixes calendar years with Microsoft’s FY2026. | Original ↗ |
Capital follows the technology wave.
Global AI investment reached historic scale in 2025. But large capital flows show the willingness to build — not that every project or investor will win.
Billion USD
Global, 2025. Private investment is included in corporate investment; the bars must not be added together.
Not financial advice. Revenue and investments are not promises of returns.
View data points and sources
| Period | Value | Source |
|---|
| Corporate | 581.7 Billion USD Global, 2025. Private investment is included in corporate investment; the bars must not be added together. | Original ↗ |
| Private | 344.7 Billion USD Global, 2025. Private investment is included in corporate investment; the bars must not be added together. | Original ↗ |
AI meets the physical world.
The IEA expects data centre electricity use to nearly double by 2030. Energy per task is falling fast, but more and heavier tasks can still increase the total.
TWh per year
2030 is an IEA forecast, not a measured outcome. All data centres, not just AI. Energy per AI task has fallen by at least an order of magnitude per year.
View data points and sources
| Period | Value | Source |
|---|
| 2025 | 485 TWh per year 2030 is an IEA forecast, not a measured outcome. All data centres, not just AI. Energy per AI task has fallen by at least an order of magnitude per year. | Original ↗ |
| 2030* | 950 TWh per year 2030 is an IEA forecast, not a measured outcome. All data centres, not just AI. Energy per AI task has fallen by at least an order of magnitude per year. | Original ↗ |
Swedish firms went from one in ten to one in three.
According to Statistics Sweden, 35.0% of Swedish enterprises with 10+ employees used AI in 2025, up from 9.9% in 2021. Among large firms the share was 71.9% in 2025 and 85% in 2026. “Uses AI” is broad, and 87% of AI-using firms in 2026 bought closed off-the-shelf software. The 2026 figure of 49% follows a series break and is shown separately.
Share of enterprises with 10+ employees, %
Comparable core 2021–2025. SCB warns of a series break between 2021 and 2023.
View data points and sources
| Period | Value | Source |
|---|
| 2021 | 9.9 Share of enterprises with 10+ employees, % Comparable core 2021–2025. SCB warns of a series break between 2021 and 2023. | Original ↗ |
| 2023 | 10.4 Share of enterprises with 10+ employees, % Comparable core 2021–2025. SCB warns of a series break between 2021 and 2023. | Original ↗ |
| 2024 | 25.1 Share of enterprises with 10+ employees, % Comparable core 2021–2025. SCB warns of a series break between 2021 and 2023. | Original ↗ |
| 2025 | 35 Share of enterprises with 10+ employees, % Comparable core 2021–2025. SCB warns of a series break between 2021 and 2023. | Original ↗ |
| 2026 (series break) | 49 Share of enterprises with 10+ employees, % Series break: new industries and NACE Rev. 2.1. Not a continuation of 35.0%. | Original ↗ |
The Nordics are at the top of the EU.
Denmark, Finland and Sweden were far above the EU average of 19.95% in 2025. For Swedish colour: Lovable reported annualized revenue of USD 600M in September 2026 — the company’s own run-rate, not audited revenue.
Share of enterprises with 10+ employees, %
Eurostat isoc_eb_ai, 2025. 2026 figures are not published until December 2026.
View data points and sources
| Period | Value | Source |
|---|
| Denmark | 42.03 Share of enterprises with 10+ employees, % Eurostat isoc_eb_ai, 2025. 2026 figures are not published until December 2026. | Original ↗ |
| Finland | 37.82 Share of enterprises with 10+ employees, % Eurostat isoc_eb_ai, 2025. 2026 figures are not published until December 2026. | Original ↗ |
| Sweden | 35.04 Share of enterprises with 10+ employees, % Eurostat isoc_eb_ai, 2025. 2026 figures are not published until December 2026. | Original ↗ |
| EU | 19.95 Share of enterprises with 10+ employees, % Eurostat isoc_eb_ai, 2025. 2026 figures are not published until December 2026. | Original ↗ |
AI has moved into the clinic.
The FDA has authorized more than 1,600 AI-enabled medical devices. 258 were added in 2025, but only 2.4% of devices with clinical studies were backed by randomized trials. Physicians report up to 83% less time on notes, yet only 5% of 500+ clinical AI studies used real clinical data.
FDA-authorized AI devices, cumulative
Official anchor points, not a yearly series. Nothing interpolated.
View data points and sources
| Period | Value | Source |
|---|
| 1995 | 1 FDA-authorized AI devices, cumulative Official anchor points, not a yearly series. Nothing interpolated. | Original ↗ |
| 2015 | 6 FDA-authorized AI devices, cumulative Official anchor points, not a yearly series. Nothing interpolated. | Original ↗ |
| 2023 | 223 FDA-authorized AI devices, cumulative Official anchor points, not a yearly series. Nothing interpolated. | Original ↗ |
| Sep 2026 | > 1,600 FDA-authorized AI devices, cumulative FDA's list is not exhaustive. Most devices are cleared via 510(k), not randomized trials. | Original ↗ |
A Nobel Prize and a research wave.
The 2024 Chemistry Prize went to David Baker for computational protein design and to Demis Hassabis and John Jumper for AlphaFold2, used by more than two million people in 190 countries. AI-related natural-science publications rose about 26% in a year. More papers are still not the same as more breakthroughs.
AI-related publications in the natural sciences, count
Approximate counts per Stanford. More papers do not equal more breakthroughs.
View data points and sources
| Period | Value | Source |
|---|
| 2024 | c. 63,547 AI-related publications in the natural sciences, count Approximate counts per Stanford. More papers do not equal more breakthroughs. | Original ↗ |
| 2025 | c. 80,150 AI-related publications in the natural sciences, count Approximate counts per Stanford. More papers do not equal more breakthroughs. | Original ↗ |
Data center revenue more than quadrupled in two years.
Nvidia’s Data Center segment grew from USD 47.5bn to USD 193.7bn between FY2024 and FY2026. These are fiscal, not calendar, years, and one vendor’s revenue — not a buy call or proof that customers earn a return.
Data Center segment, USD bn, fiscal year
Fiscal years ending late January, not calendar years. One segment, not a pure AI index.
View data points and sources
| Period | Value | Source |
|---|
| FY2024 | 47.5 Data Center segment, USD bn, fiscal year Fiscal years ending late January, not calendar years. One segment, not a pure AI index. | Original ↗ |
| FY2025 | 115.2 Data Center segment, USD bn, fiscal year Fiscal years ending late January, not calendar years. One segment, not a pure AI index. | Original ↗ |
| FY2026 | 193.7 Data Center segment, USD bn, fiscal year Fiscal years ending late January, not calendar years. One segment, not a pure AI index. | Original ↗ |
Faster diffusion than the PC and the internet.
Stanford summarizes that generative AI reached about 53% population adoption in three years — faster than the PC (1981) and the internet (1995) in the same kind of comparison. Estimated US consumer value rose from USD 112bn to 172bn a year. The US itself lags, ranking 24th.
Generative AI adoption in the population, %
Stanford AI Index 2026. The US ranks 24th. Adoption is defined differently across technologies.
View data points and sources
| Period | Value | Source |
|---|
| Singapore | 61 Generative AI adoption in the population, % Stanford AI Index 2026. The US ranks 24th. Adoption is defined differently across technologies. | Original ↗ |
| UAE | 54 Generative AI adoption in the population, % Stanford AI Index 2026. The US ranks 24th. Adoption is defined differently across technologies. | Original ↗ |
| US | 28.3 Generative AI adoption in the population, % Stanford AI Index 2026. The US ranks 24th. Adoption is defined differently across technologies. | Original ↗ |
After the dip came a wave of capital.
In constant prices, private AI investment fell from USD 145.4bn in 2021 to 92.8bn in 2023, then reached 290.1bn in 2025. The series excludes internal R&D, capex and listed giants; China is underestimated, and single mega-deals create spikes.
Private AI investment, USD bn in constant 2021 prices
Constant 2021 USD. Not comparable on one axis with Stanford’s nominal USD 344.7bn for 2025.
View data points and sources
| Period | Value | Source |
|---|
| 2020 | 77.3 Private AI investment, USD bn in constant 2021 prices Constant 2021 USD. Not comparable on one axis with Stanford’s nominal USD 344.7bn for 2025. | Original ↗ |
| 2021 | 145.4 Private AI investment, USD bn in constant 2021 prices Constant 2021 USD. Not comparable on one axis with Stanford’s nominal USD 344.7bn for 2025. | Original ↗ |
| 2022 | 104.6 Private AI investment, USD bn in constant 2021 prices Constant 2021 USD. Not comparable on one axis with Stanford’s nominal USD 344.7bn for 2025. | Original ↗ |
| 2023 | 92.8 Private AI investment, USD bn in constant 2021 prices Constant 2021 USD. Not comparable on one axis with Stanford’s nominal USD 344.7bn for 2025. | Original ↗ |
| 2024 | 130.9 Private AI investment, USD bn in constant 2021 prices Constant 2021 USD. Not comparable on one axis with Stanford’s nominal USD 344.7bn for 2025. | Original ↗ |
| 2025 | 290.1 Private AI investment, USD bn in constant 2021 prices Constant 2021 USD. Not comparable on one axis with Stanford’s nominal USD 344.7bn for 2025. | Original ↗ |
AI skills pay.
PwC finds jobs requiring AI skills are advertised at a 62% wage premium, up from 57% the year before, with 118% in some consumer sectors and 16% in the public sector. Such jobs grew 69% versus 9% overall. The most AI-exposed firms grew headcount 52% and productivity 34% since 2018, versus 36% and 24%. Yet other US entry-level roles fell 10%.
Average wage premium in job ads requiring AI skills, %
Advertised wages (Lightcast), not a causal wage gap.
View data points and sources
| Period | Value | Source |
|---|
| Previous report | 57 Average wage premium in job ads requiring AI skills, % Advertised wages (Lightcast), not a causal wage gap. | Original ↗ |
| 2026 report | 62 Average wage premium in job ads requiring AI skills, % Advertised wages (Lightcast), not a causal wage gap. | Original ↗ |
Open models caught up — then the gap reopened.
The closed-model lead was 15.2% in May 2023 and just 0.5% in August 2024. By March 2026 it was 3.4%, and six of the top ten models are closed. Open weights push down price and widen access, but Swedish firms mostly buy closed off-the-shelf tools.
Arena gap between best closed and best open-weight model, %
Crowdsourced benchmark. The gap reopened in 2025–2026.
View data points and sources
| Period | Value | Source |
|---|
| May 2023 | 15.2 Arena gap between best closed and best open-weight model, % Crowdsourced benchmark. The gap reopened in 2025–2026. | Original ↗ |
| Aug 2024 | 0.5 Arena gap between best closed and best open-weight model, % Crowdsourced benchmark. The gap reopened in 2025–2026. | Original ↗ |
| Mar 2026 | 3.4 Arena gap between best closed and best open-weight model, % Claude Opus 4.6 (1,503) vs GLM-5 (1,454). The same chapter also states 3.3%. | Original ↗ |
Americans are growing more concerned.
Pew: the share of Americans more concerned than excited about AI rose from 37% in 2021 to 50% in June 2025. Where benefits are concrete the picture is brighter: 44% view AI in healthcare positively versus 19% negatively. Opinion is not usage.
Share of respondents, %
Pew: share of Americans more concerned than excited.
View data points and sources
| Period | Value | Source |
|---|
| US 2021 | 37 Share of respondents, % Pew: share of Americans more concerned than excited. | Original ↗ |
| US Jun 2025 | 50 Share of respondents, % Pew: share of Americans more concerned than excited. | Original ↗ |
The world is growing more optimistic.
Stanford: the global share who believe AI offers more benefits than drawbacks rose from 52% to 59%, while 52% are also nervous. It is a different question and population from Pew's, so the series are shown separately.
Share of respondents, %
Stanford: global share optimistic about AI’s benefits.
View data points and sources
| Period | Value | Source |
|---|
| Global, prior | 52 Share of respondents, % Stanford: global share optimistic about AI’s benefits. | Original ↗ |
| Global 2026 | 59 Share of respondents, % Stanford: global share optimistic about AI’s benefits. | Original ↗ |
More incidents in the media — and the water.
The OECD counts an average of 324 media-reported AI incidents and hazards per month in 2025, versus 92 per month in 2022. These are media reports, not verified harms, and their share of all AI news fell. Stanford estimates that annual water use for GPT-4o inference may exceed the drinking-water needs of 1.2 million people, and that training Grok 4 emitted 72,816 tCO₂e.
Media-reported AI incidents and hazards, monthly average
Counts media reports, not verified harms. The share of all AI news fell from 3.2% to about 2.5%.
View data points and sources
| Period | Value | Source |
|---|
| 2022 | 92 Media-reported AI incidents and hazards, monthly average Counts media reports, not verified harms. The share of all AI news fell from 3.2% to about 2.5%. | Original ↗ |
| 2025 | 324 Media-reported AI incidents and hazards, monthly average Data through September 2025. | Original ↗ |