AI statistics in 2026: the scoreboard, checked number by number

The world's data centers will consume "more than 1,000 TWh of electricity in 2026 — the equivalent of Japan". The sentence appears in dozens of AI statistics roundups published this year, almost always in the present tense, almost never with a date. I went and checked. The number exists, it comes from the International Energy Agency — and the IEA itself has retired it. It was the ceiling of a range projected in January 2024 (620 to 1,050 TWh), which included cryptocurrencies in the count; the base case in the same report was "just over 800". Whoever rounded 800 up to "more than 1,000" was not a marketing blog: it was the agency's own executive summary. The IEA's current series, from April 2026, measures 485 TWh in 2025 — without crypto — and Gartner projects 565 TWh for 2026. The number in circulation is almost double the best available number, and the zombie was born inside the original document, in the distance between the report's body and the summary everyone quotes.

This is an AI statistics article — the genre that produces the most numbers and the fewest dates. There is only one ruler here: a number with no date, no measurer, and no scope is not a data point. I checked every figure in this text against the primary source, with date and scope noted. What follows is the scoreboard of what survived.

The collision that teaches the ruler: the "AI market" is worth less than one AI company

The 2025 "global AI market" was valued at US$ 390.9 billion (Grand View Research, Nov 2025). Anthropic, on its own, was valued at US$ 965 billion in its Series H round (official announcement, May 28, 2026). A company "worth" two and a half times the entire market it operates in should break any spreadsheet — and it does not, because the two figures do not measure the same thing. A market is estimated revenue; a valuation is a bet on future cash flow; investment is money that actually came in; a round is an event. The statistics roundups add the four categories into a single list, and that is how the feeling that "AI numbers never square" is born. They do not square because they are not addable:

Three observations the table hides. First: the OECD's US$ 258.7 billion (OECD) include the mega-rounds (deals above US$ 100 million account for 73% of the value) and exclude big tech's internal investment — the brief declares the cut on page 2. Second: Grand View's "market" figure is not auditable — the list of companies counted and the model behind the projected growth sit behind the paywall, and the firm's own material contradicts itself: the report page says a CAGR of 30.6% (2026-2033), the official press release announces 31.5% (2025-2033) — two growth rates for the same forecast, depending on the base year. Third: in the same firm's separate report, the 2025 "genAI market" is US$ 22.2 billion — 5.7% of the "AI market" in the middle of the generative era. Definition boundaries this opaque are the reason Precedence Research publishes US$ 4.2 trillion (by 2035) and MarketsandMarkets US$ 3.6 trillion (by 2033) for "the same" market. The spread between consultancies is, itself, the data point.

And there are the numbers that age in weeks. The xAI+SpaceX case is the living proof: July's roundups said it was "worth more than US$ 2.1 trillion". I dated every point: the IPO priced on Jun 12, 2026 at US$ 1.78 trillion; it closed the first day at US$ 2.1 trillion; four days later the stock hit its all-time high ~40% above the debut close; and in August 2026 the company is worth US$ 1.43 trillion. The same number was stale upward and downward within a window of eight weeks. A valuation without a date is not a data point — it is a memory.

At the opposite end of the age spectrum sit the fossil forecasts, present in nearly every 2026 roundup: "AI will add US$ 15.7 trillion to the economy by 2030" is from PwC in June 2017 — and the study's original URLs are dead; pwc.com today redirects both the page and the PDF to other products. McKinsey's "US$ 13 trillion" is from September 2018; the "genAI adds US$ 2.6-4.4 trillion per year" range, from June 2023. They are projections of potential, with methodologies that cannot be compared to one another, quoted nine, eight, and three years later as if they described 2026.

The scoreboard

I took the statistics that circulate most in "AI in 2026" roundups — including the German-language survey that inspired this story — and checked them one by one against the primary source. An honest legend: checks out (number, date, and scope match), half-truth (right number, scope or denominator omitted), fossil (real number, from another era, circulating as current), zombie (projection retired by its own source), does not check out.

The row that best sums up the scoreboard is the 53% one. The AI Index 2026 opens by saying that "generative AI reached close to 53% population adoption in three years — faster than the PC and the internet". The figure's footnote (4.3.9, p. 199) says what the headline does not: the sample is individuals aged 18 to 64 in the United States (the RPS/CPS surveys, by Bick, Blandin, and Deming). It is neither the world population nor the online population — and the contrast is in the report itself: by Microsoft's telemetry, two pages later, US adoption is 28.3%, 24th place in the world. Both numbers are "right"; they measure different things. The denominator is where AI statistics die — and this pattern will repeat in every section below.

None of this is news to anyone who reads the forums. In the discourse archive I compiled (2024-2026), Reddit's standard reply to an unsourced AI statistic has already become a meme — "Source: Trust me bro!" — and Gartner's Hype Cycle chart is cited by optimists and pessimists alike to prove opposite theses. The ambient skepticism is well calibrated; what is missing, almost always, is someone opening the PDF.

The funnel the 88% hides

The most quoted corporate statistic of 2026 — 88% of organizations use AI — is true and comes from McKinsey ("The state of AI in 2025", Nov 2025, 1,993 respondents in 105 countries): regular use in at least one business function, up from 78% a year earlier. What almost never circulates with it are the other two questions from the same survey:

The three steps, in the source's exact wording: 88% use AI in at least one function; "no more than 10 percent" of respondents say they are scaling AI agents in any given function (the per-company cut is more generous: 23% are scaling agents somewhere in the organization, and 62% are at least experimenting); and about 6% are what McKinsey defines as high performers — they attribute to AI an effect of 5% or more on EBIT and report "significant" value. Notice the caveats, because they are the data: everything is self-reported by survey respondents, the "6%" is a composite definition, and 39% attribute some EBIT effect to AI — most of them, below 5%. Use is easy to declare; profit that shows up on the balance sheet is another species of claim.

Two ghosts haunt this family of numbers. The "72% of organizations use genAI", present in several roundups, does not exist in McKinsey's current edition — it is contamination from past editions (72% was overall AI adoption in early 2024). And the AI Index itself trips over the data point: the chapter 4 highlight says "70% of organizations use genAI" while the body of the same chapter says 79% — in the same PDF, 21 pages apart.

The discourse archive records the community's reaction to this family of statistics: when the "95% of genAI pilots fail" made the rounds (the MIT NANDA report, Aug 2025), the technical critique pointed at the small sample and the cut — the report measured formal IT channels, not actual use — and the top-voted joke summed up the mature skepticism: "I'm surprised 5% are working."

What people use it for: the 4.2% that contradicts the feed

The best public data on actual use comes from the paper NBER w34255 ("How People Use ChatGPT", Sep 2025; Chatterji, Cunningham, Deming and others — with OpenAI's internal data): 1.1 million conversations sampled between May 2024 and Jun 2025. The three biggest uses — practical guidance, seeking information, and writing — add up to 77% of all conversations. Programming: 4.2% of messages. Mathematics, 3%; data analysis, 0.4%.

The 4.2% demolishes the impression that "everyone uses AI to code" — but read the scope before quoting it: the sample covers consumer plans only (Free, Plus, Pro). Enterprise, Education and — decisively — API and Codex are left out. The paper notes that programming migrated out of ChatGPT; the 4.2% measures the chat, not the category. The bubble of people who code with AI exists; it just does not live where the statistic was collected.

At work, the numbers that circulate together come from separate rulers — each with its own denominator: 52% of US workers use AI in their own role (Gallup, Q2 2026, n=22,573 — the series went 40% -> 45% -> 52% in one year); 87% of digital workers use AI and 75% say they are more productive (Glean Work AI Index 2026 — 6,000 respondents, US/UK/Australia only, and perceived, self-declared productivity). And the genre's two pet fossils: "46% of code is written by AI" is the GitHub blog post of Feb 14, 2023 (three and a half years ago, when the same metric had been 27% in Jun 2022); and Nadella's "20-30% of Microsoft's code" (LlamaCon, Apr 2025) carries a clause the summary always drops: "in some of our projects" — some projects, not the company's codebase.

Among teenagers, the fossil is more recent and more treacherous: "26% of American teens use ChatGPT for homework" has fieldwork from Sep-Oct 2024. The next wave of the same Pew survey (fieldwork Sep-Oct 2025, published Feb 2026) measured 54% — a broader category, chatbots in general — with 1 in 10 saying they do "all or almost all" of their homework with them. Quoting the 26% in 2026 is being wrong by half.

Energy: anatomy of the zombie number

The 1,050 TWh case deserves the full X-ray, because it teaches how a zombie forms in three steps. Step 1: in Jan 2024 the IEA publishes a range for 2026 — 620 to 1,050 TWh — whose declared scope includes data centers and cryptocurrency mining (~160 TWh of the base case). Step 2: the IEA's own executive summary rounds the base case of "just over 800" up to "more than 1 000 TWh… roughly equivalent to Japan". Step 3: the roundups quote the summary, cut the range, the scope, and the date — and three years later the ceiling had become "data center consumption in 2026".

The series that replaced the retired range (IEA, Key Questions on Energy and AI, Apr 16, 2026) measures 485 TWh in 2025 and projects ~950 TWh in 2030 — without crypto, which the agency now tracks separately. Gartner (Jun 2026) projects 565 TWh for 2026, with the decomposition that matters: AI-optimized servers are 175 TWh — 31% of the total. AI is not even the majority of data center consumption; it is the fastest-growing slice (95 -> 175 -> 258 TWh between 2025 and 2027, in the projection).

The country comparison moves with it: with the 1,050 TWh zombie, "data centers" would be the world's 5th largest electricity consumer, between Russia and Japan — with a 2% margin over Japan, sensitive to the series used. With Gartner's 565 TWh, they drop to 9th, between South Korea and Germany (Ember/OWID demand series, 2025 data). Still gigantic; no longer the headline that circulated.

Two neighbors of this section call for the same ruler. xAI's Colossus supercluster: the "almost 2 GW" is a post by Musk himself (Dec 2025), and the "555,000 GPUs" is a press aggregation built on top of his posts — there is no independent audit of either. And the water viral — "one email on GPT-4 uses a 519 mL bottle of water" (Washington Post/UC Riverside, Sep 2024) — was dismantled in May 2026 by Andy Masley, who recalculated the cost at 2.1-9.8 mL per prompt; the researcher who originated the figure, Shaolei Ren, then communicated a revision to ~15 mL per prompt (5 mL in the data center; the rest in power generation). There is no new public paper from Ren — the trail is the update on Masley's post plus the Jul 2026 coverage — so I record it as a communicated revision, not a publication. Even so: the viral number sits about 35x above the best current estimate, and it keeps circulating.

US × China: the two scoreboards tell opposite stories — on purpose

In money, an ocean: private AI investment of US$ 285.9 billion in the US against US$ 12.4 billion in China in 2025 — 23.1x (AI Index 2026, Quid data, p. 181). With the caveat the report itself carries and the citations cut: the data cannot see state funds — China's government guidance funds alone would have allocated ~US$ 184 billion to AI between 2000 and 2023, by the estimate cited in the report. The 23x compares private capital, not national effort.

In frontier output, the same edition counts 59 notable US models against 35 from China in 2025 (Epoch AI's manual curation). And in performance, the story flips: in March 2026, the gap between the best American model and the best Chinese one on the Arena leaderboard was 2.7% (Claude Opus 4.6, 1,503, against Dola-Seed-2.0 Preview, 1,464) — a year earlier, it had reached 0.4%. It is a snapshot of a volatile leaderboard, and the Index itself warns that Arena position can reflect adaptation to the platform; the honest reading is not "China has caught up", it is "the performance gap closed while the capital gap and the frontier-volume gap stay open". The two scoreboards together are the thesis; either one alone is cheerleading.

In public perception, the numbers that circulate glued together come from different surveys — and none of them is "global". The median of the 25 countries surveyed by Pew (Oct 2025): 53% trust the European Union to regulate AI, 37% the US, 27% China. And the Gen Z enthusiasm that "fell from 36% to 22%" is Gallup/Walton (Apr 2026) — US only, ages 14-29, with anger up 9 points over the same period. These are cuts; treating them as "the world thinks X" is the same mistake as the 53% adoption figure.

Chatbots: the web ruler measures sites, not use

The traffic share that circulates — ChatGPT 52.7%, Gemini 27.3%, Claude 8.9% (Similarweb, May 2026) — deserves the two notes that always fall off along the way. First: a year earlier, it was 76.4% / 8.9% / 1.6%. ChatGPT lost 24 points of share in 12 months; Claude multiplied its own by ~5.6 (the roundups say "3x" — they understate); and the Jun 2026 cut shows chatgpt.com traffic falling 3.29% month over month. Second, and more important: the ruler is web-only. It measures visits to the domains; it does not see native apps (an estimated undercount of 15-60% depending on the assistant), does not see the API, does not see the AI embedded in Google search — which undercounts precisely Gemini. It is a proxy for consumer web product, useful for trends, useless for "how many people use it".

The Brazil step: present where you count people, absent where you count capital

The German-language survey that inspired this story does not have one line about Brazil — and neither do the international rankings. I went looking for the country in the AI Index 2026 tables, and the absence is the first data point: Brazil does not appear in the top 15 for private AI investment in 2025 (15th place, Sweden, sits at US$ 0.97 billion — Brazil is below that), nor in the tables of newly funded companies, nor in the 2013-2025 cumulative total. Where the report counts infrastructure and people, the country exists: 197 data centers (against 5,427 in the US), 11,100 top AI authors and inventors — 10th place in the world — and 84% of Brazilian university students have already used genAI in their studies (Chegg survey in 15 countries), above the 67% of the US and the UK, and 32 points above 2023.

In the general population, the picture I assembled in the previous article in this series still stands: ~23% of the Brazilian population uses generative AI (derived from the 32% of internet users measured by Cetic.br in 2025), against a global ceiling of 29.2% that — as I showed there — counts accounts, not people. Brazil's 2026 picture fits in one sentence: people adoption near the frontier, capital off the map — 84% of university students using it, less than a billion dollars coming in.

What I would do with this

I will not pretend there is a secret, correct "AI statistic" that everyone else gets wrong. What exists is a pocket ruler, and it costs three questions:

Who measured, measuring what? Revenue, investment, VC, valuation, web visits, survey respondents — they are species that cannot be added together. The metrology table near the top is worth more than any individual number in this text.

What is the denominator? "53% of the population" was the US, ages 18-64. "50% of workers" was the US. "Global median" was 25 countries. "87% of knowledge workers" were digital workers from three English-speaking countries. In AI statistics, the omitted denominator is the rule, not the exception.

When is the snapshot from? The Copilot 46% is three and a half years old. PwC's US$ 15.7 trillion is nine years old. xAI's valuation aged twice in eight weeks. Every number in this article carries its date in the body of the text — not as a matter of style, but because the date is what decides whether the number is still a data point.

The rest I hand back to the reader: the next time a LinkedIn card says "AI is already a market of X trillion used by Y% of humanity", the three questions fit in a comment — and the primary source is almost always two clicks away. That is all this article did.

Sources

• Stanford HAI — AI Index Report 2026 · official PDF (investment, the 53%, US×China, Chegg, Brazil)
• McKinsey — The state of AI in 2025 (88%, agents, high performers)
• OECD — Venture capital investments in AI through 2025 (US$ 258.7 billion, 61%)
• Grand View Research — Artificial Intelligence Market · press release (US$ 390.9 billion; the two CAGRs)
• Anthropic — Series H · CNBC · NBC (May 2026 valuations)
• CNBC — SpaceX tops $2 trillion · companiesmarketcap — SPCX (xAI+SpaceX timeline)
• IEA — Electricity 2024 · Energy and AI · Key Questions on Energy and AI (the retired range and the current series)
• Gartner — press release, Jun 10, 2026 (565 TWh, decomposition)
• Ember/OWID — electricity demand by country (country ranking)
• NBER — Working Paper 34255 (77%; programming 4.2%)
• Gallup — AI use at work · Glean — Work AI Index 2026 (52%; 87%/75%)
• GitHub — Copilot for Business (Feb 2023) · The Register — Nadella at LlamaCon (the code fossils)
• Pew — teens and ChatGPT (2024) · Pew — How Teens Use and View AI (Feb 2026) · Pew — regulatory trust (Oct 2025)
• Gallup/Walton — Gen Z and AI (Apr 2026)
• Similarweb via ppc.land (May 2026) (chatbot share; web-only caveat)
• Andy Masley — the origin of the water error · Forbes (Jul 2026) (the ~15 mL revision)
• Elon Musk's posts about Colossus: Jul 2025 · Dec 2025 (claims without independent audit)
• PwC — Sizing the prize (June 2017; press release from the time; original URLs dead, checked via the Wayback Machine) · McKinsey Global Institute — Notes from the AI frontier (Sep 2018) · McKinsey — The economic potential of generative AI (Jun 2023)

This article's story was inspired by a survey from gradually.ai (in German). No number was inherited from it: every figure was checked independently against the primary sources above, on August 2, 2026 — official PDFs where available (AI Index, OECD, IEA), Wayback Machine snapshots of the official pages where the site blocks automated access (McKinsey, Gartner, Grand View), and the original text via a Common Crawl capture of July 14, 2026. Valuations of private companies have no auditable source beyond the companies' own announcements and the financial press — that is why they are dated in the body of the text. The community quotes are a 2024-2026 discourse archive, with the threads checked live, not a portrait of "right now". What I did not verify: xAI's unpublished internal numbers (GPUs, power), the primary document of Shaolei Ren's revision (communicated via third parties), and the proprietary decomposition behind Grand View's market figure.