What the AI boom is costing, and who is collecting
Three questions run through this page. Who is making money, and is the money real? What does it cost in electricity, water and carbon to build and run the machines? And what is it doing to work, and to daily life? The evidence is uneven and often comes from the companies themselves, so every number below carries a source, and the contested ones are marked as contested.
Six figures to hold in mind
As of September 2026. Click any caption for the source.
The largest private construction program in history
Four companies pay for most of it. Amazon, Alphabet, Microsoft and Meta spent roughly $410 billion on capital projects in 2025; by late July 2026 their combined 2026 guidance had reached $720 to 760 billion. The International Energy Agency (IEA) says the five largest data-center companies now outspend the entire U.S. oil and gas industry on production, with a further 75% rise expected in 2026. Oracle guided about $50 billion, up 136%. Investors are wary: Alphabet’s July 2026 report set off a sell-off over the size of the spending.
The clearest winner is the chip supplier. Nvidia became the first company worth $5 trillion on October 29, 2025, dipped in early 2026, and reached $5.5 trillion on May 13, 2026, the highest market value in history. The model makers are close behind on paper. OpenAI closed a $122 billion round at an $852 billion valuation on March 31, 2026, with Amazon, Nvidia and SoftBank the largest investors, and filed for a public listing on June 8, 2026 that its chief executive later said would slip to 2027. Anthropic raised $30 billion at $380 billion in February 2026 and, by its own announcement, $65 billion at $965 billion on May 28, 2026.
Big-four capital spending
Billions of dollars: 2025 spending versus 2026 company guidance (midpoint of range)
2026 bars are guidance, not results; hover for the ranges. The 2025 company-by-company breakdown is approximate and was not confirmed against filings; the combined 2025 total of about $410 billion is reported by Statista and CNBC.
Nvidia’s market value
Trillions of dollars at reported milestones
The $4 trillion milestone in July 2025 is widely reported but its exact date was not verified for this page. The line skips the dip of early 2026.
Private valuations
Billions of dollars, post-money, at each round or share sale
OpenAI’s $500 billion figure was an employee share sale in October 2025. Anthropic’s May 2026 round is per the company’s announcement; independent confirmation was thin at the time of writing.
ChatGPT weekly users
Millions of weekly active users, as reported
The July 2026 point is “approaching one billion,” per The Information, about seven months later than OpenAI had projected; growth slowed after the GPT-5 backlash.
Circular deals and the bubble question
OpenAI has lined up about $1.4 trillion in “potential commitments” for computing power; one tally of 2025–2035 deals runs to roughly $1.15 trillion, including Broadcom $350 billion, Oracle $300 billion, Microsoft $250 billion, Nvidia up to $100 billion and AMD $90 billion. Several suppliers are also investors, which is why analysts describe a loop: Nvidia invests in OpenAI, which buys Nvidia chips through Oracle and Microsoft, which buy from Nvidia. Nvidia’s $100 billion arrangement was reported to have stalled in 2026.
Is it a bubble? Asked in August 2025 whether investors are overexcited about AI, Sam Altman said, “My opinion is yes.” An MIT NANDA report the same week found about 95% of enterprise generative-AI pilots deliver no measurable profit-and-loss impact; tech stocks lost about $1 trillion over four days. Mark Zuckerberg called a “collapse” “definitely a possibility.” In November 2025, Michael Burry disclosed bets against Nvidia and Palantir, arguing the hyperscalers understate how fast their chips lose value, and 45% of fund managers in a Bank of America survey named an AI bubble the biggest tail risk, up from 11% two months earlier.
The counter-view is that the revenue is real. Anthropic reported a $14 billion annualized run-rate in February 2026 and said it had reached $30 billion by about April. OpenAI’s annualized revenue was estimated near $40 billion by mid-2026, though estimates vary widely and its 2025 loss reportedly exceeded revenue. ChatGPT is used by close to a billion people a week. Whether that pays for a trillion dollars of chips is the open question.
Most companies still do not use it
The Census Bureau found 17 to 20% of U.S. businesses using AI to produce goods or services between December 2025 and May 2026, and 37% of firms with 250 or more employees. McKinsey’s survey of larger organizations worldwide found 88% use AI in at least one function, but only 39% report any effect on earnings.
More than half of all venture money
Global venture deal value reached about $512 billion in 2025, with AI taking more than half of the dollars and nearly two-thirds in the U.S. CoreWeave listed in March 2025 at about $23 billion; Cerebras rose 68% on its May 14, 2026 debut to $95 billion; SpaceX, merged with xAI, raised $75 billion in June 2026, the largest IPO ever.
The Sora app lasted seven months
OpenAI’s video app launched September 30, 2025, drew copyright objections from Japanese studios and Hollywood, won a $1 billion Disney investment in December, and was shut down on April 26, 2026 after usage peaked near one million users at a reported cost of about $1 million a day. Disney exited the deal.
The physical bill
A data center is a building full of chips that must be powered and cooled around the clock. The numbers below come from the IEA, the U.S. Department of Energy’s Berkeley Lab, grid operators and the companies’ own reports.
Global data-center electricity
Terawatt-hours per year, IEA
2024 use was about 1.5% of world electricity; the 2030 base case is just under 3%, growing about 15% a year, four times faster than everything else combined. The 2025 figure (roughly 485 to 500 TWh, up 15 to 17%) comes from the IEA’s April 2026 update as reported by secondary sources.
U.S. data-center electricity
Terawatt-hours per year, Berkeley Lab reference case
2024 was 4.7% of U.S. electricity; the 2030 reference case is 11.8%, with an uncertainty range of 521 to 843 TWh (9.5 to 15.3%). Data centers would be a third of all U.S. load growth. The 2023 figure is from the earlier 2024 report and was not re-read for this page.
Who pays: the grid
In PJM, the grid serving 13 states from Illinois to Virginia, the auction price for guaranteed generating capacity rose from about $29 per megawatt-day for 2024/25 to $329.17 for 2026/27, roughly eleven times in two years. PJM’s independent market monitor attributed 40% of the $16.4 billion cost of the December 2025 auction to data centers, most of them not yet built. U.S. residential electricity prices rose 42% in the five years to March 2026, with the District of Columbia up 94% and Maryland up 74%. Utilities asked for $29 billion in rate increases in 2025 and $18.6 billion in the first half of 2026; about one in six households entered 2026 behind on utility bills.
The other side. Gas prices, wildfire costs and transmission upgrades are also pushing bills up; Fortune argues it is “not just about data centers.” PolitiFact rated a senator’s claim that data centers raised bills 267% Mostly False, because the figure referred to wholesale prices at grid nodes, not consumer bills. Google says its data centers use 83% less overhead energy than the industry average and a Gemini prompt uses 33 times less energy than a year earlier. Critics reply with the Jevons paradox: per-prompt efficiency is swamped by volume, and Google’s electricity use still rose 37% in 2025.
PJM capacity price
Dollars per megawatt-day, annual auction
Company emissions are rising
Change in reported total greenhouse-gas footprint, latest report
Google: 2025 versus 2024, and versus its 2019 baseline. Microsoft: fiscal 2025, 20.29 million tonnes, partly a bookkeeping change after it stopped counting unbundled renewable certificates. Amazon: 80.85 million tonnes in 2025, its largest annual increase since it began reporting.
| Estimate | Energy | Water | Carbon | Who says |
|---|---|---|---|---|
| Median Gemini text prompt | 0.24 Wh | 0.26 mL | 0.03 g CO₂e | Google, Aug 2025 (published methodology) |
| Average ChatGPT query | 0.34 Wh | 0.000085 gal (about 1/15 teaspoon) | not given | Sam Altman blog, June 2025 (no methodology) |
| Training Mistral Large 2 (whole run) | not given | 281,000 m³ | 20,400 t CO₂e | Mistral life-cycle assessment with Carbone 4 and ADEME, July 2025 |
What one prompt costs
These are company-reported medians for short text prompts. Reasoning models, long documents and image or video generation use far more, and independent researchers criticize the exclusion of training and the use of “market-based” carbon accounting that credits purchased clean energy. Mistral found training plus everyday use account for 85.5% of its model’s lifetime emissions and 91% of its water. A 2024 paper in Nature Computational Science estimated generative AI could add 1.2 to 5 million tonnes of electronic waste by 2030.
Water
U.S. data centers directly consumed 17.4 billion gallons of water in 2023, projected at 38 to 73 billion by 2028; a medium facility uses about as much as a thousand households. The Houston Advanced Research Center estimated Texas data centers would use 49 billion gallons in 2025 and up to 399 billion by 2030. In Fayette County, Georgia, a QTS data center drew about 29 million gallons without authorization while neighbors reported collapsing water pressure; Tucson rejected “Project Blue” in August 2025 after a contractor hauled some 650,000 gallons outside city limits for dust control. Google says replenishment projects returned about 78% of the freshwater it consumed in 2025; Microsoft says 90% of its 2025 fleet uses low- or zero-water cooling. The Information Technology and Innovation Foundation argues the problem is “soluble” with dry cooling, at the cost of more electricity.
Gigawatt campuses and the plants that will feed them
Hover a point. A gigawatt is roughly the output of one large nuclear reactor.
Stargate: a $500 billion commitment and more than 9 GW planned across seven sites by 2029; only Abilene was operating in April 2026, at 0.3 GW of a planned 1.2. Meta’s Hyperion in Louisiana was expanded in July 2026 to 5 GW and more than $50 billion, with Entergy building about 3.8 GW of new gas plants whose filings showed consumer costs would rise. Nuclear: Microsoft’s 20-year deal to restart Three Mile Island Unit 1 (835 MW, targeted 2027–28), Meta’s 1,121 MW contract at Clinton, Illinois, and Amazon’s deal for up to 1.92 GW from Susquehanna through 2042.
Two readings of the same labor market
Careful researchers disagree about whether AI is already displacing workers. Both cases are laid out here.
Signs of disruption
- Stanford’s “canaries.” Using ADP payroll data, Brynjolfsson, Chandar and Chen first reported (August 2025) a 13% relative employment decline for 22-to-25-year-olds in the most AI-exposed occupations; the August 2026 revision put the gap at 19%. The effect comes through reduced hiring rather than layoffs, and is concentrated where AI substitutes for work rather than complementing it.
- Recent graduates. The New York Fed reports 5.6% unemployment and 42% underemployment among recent college graduates in the second quarter of 2026.
- Employers say so. Challenger, Gray & Christmas found AI the leading stated reason for job cuts for the fourth straight month in July 2026, with 10,970 cuts attributed to it.
- Amazon. The memo announcing about 14,000 corporate cuts in October 2025 called this generation of AI “the most transformative technology we’ve seen since the Internet.”
- The warning. Anthropic’s Dario Amodei told Axios in May 2025 that AI could eliminate half of entry-level white-collar jobs within one to five years; his January 2026 essay lists mass labor displacement among four core risks.
Signs it is early
- No economy-wide shift. The Yale Budget Lab, tracking the mix of occupations since ChatGPT launched, reported in October 2025 that it found no discernible disruption to the overall labor market. (The report page could not be retrieved; the wording is as widely reported.)
- Hiring, not firing. The Stanford authors themselves find no economy-wide displacement, and the effect shows up in employment, not wages.
- Adoption is thin. Only 17 to 20% of U.S. businesses used AI in production as of spring 2026, per the Census Bureau; retail was near 14%.
- Productivity is uneven. A 2025 randomized study by METR found experienced open-source developers took 19% longer with AI tools while believing they had sped up. The best-known positive result, a 14% average gain for customer-support agents, was 34% for novices and near zero for veterans.
- Earnings effects are rare. In McKinsey’s survey, 88% of organizations use AI but only 39% see any effect on earnings, and about 6% a meaningful one.
What controlled studies found
Change in measured productivity with AI tools, percent
Customer support: 5,179 agents, issues resolved per hour (Brynjolfsson, Li and Raymond, 2023). Developers: 16 experienced open-source contributors on 246 real tasks with Cursor and Claude 3.5/3.7 (METR, July 2025); they expected a 24% speedup and believed afterward they had gained 20%.
| Estimate | Finding | Source |
|---|---|---|
| Jobs “exposed” | The equivalent of 300 million full-time jobs worldwide exposed to automation; global GDP up about 7% over ten years | Goldman Sachs, 2023 |
| Share of employment | Almost 40% of jobs globally; 60% in advanced economies, 26% in low-income countries; AI likely to worsen inequality | IMF, Jan 2024 |
| Where it applies | From 200,000 Copilot conversations: computer and mathematical, office support, sales and information jobs score highest on “AI applicability,” which the authors stress is not the same as displacement | Microsoft Research, July 2025 |
| What people actually ask | “Directive” automation rose from 27% to 39% of Claude.ai conversations between early and late 2025, the first time it exceeded augmentation; 77% of business API use is automation; coding is 36% of consumer use | Anthropic Economic Index, Sept 2025 |
Exposure means a job contains tasks a model can do, not that the job disappears. Usage is also geographically lopsided: Anthropic found each 1% of extra GDP per person is associated with about 0.7% more use, with Israel and Singapore far above their expected share.
Companions, classrooms, deepfakes and fortunes
72% have tried an AI companion
Common Sense Media’s 2025 survey found 72% of U.S. teens had used AI companions, about half regularly; a third had chosen an AI over a person for a serious conversation. Common Sense recommends no one under 18 use them. Pew reported in December 2025 that 64% of teens use chatbots (a figure from the writer’s brief, listed for checking). See risks for the lawsuits.
Half expect it to make us worse at both
In Pew’s September 2025 survey, 50% of Americans said AI will worsen people’s ability to form meaningful relationships (5% said improve) and 53% said it will worsen creative thinking.
The homework business collapsed
Chegg, the online tutoring company, lost 38% of its value in a day in May 2023 after blaming ChatGPT, cut 22% of staff in May 2025 and 45% of the remainder that October, and sued Google over AI Overviews. Data on classroom use was not gathered for this page.
$40 billion in deepfake losses by 2027
Deloitte’s projection, as compiled on Wikipedia, alongside a current estimate near $12 billion and an average loss above $500,000 per criminal deepfake incident; the best-known case is the $25 million Arup transfer in 2024. Treat the totals as estimates.
Consent, replicas and a $1.5 billion settlement
SAG-AFTRA’s video-game strike (July 2024 to July 2025) won rules requiring consent and disclosure before a digital replica is made and pay comparable to live work; 95% of members approved. The AI “actress” Tilly Norwood drew boycott calls when her creators claimed agencies wanted to sign her. Sora 2’s opt-out copyright policy drew protests from Studio Ghibli’s trade group and the Motion Picture Association before the app was shut down. The largest copyright payout, Anthropic’s $1.5 billion for pirated books, is on the training page.
3,428 billionaires, $20.1 trillion
Forbes’ 2026 list, as compiled on Wikipedia, counted a record 3,428 billionaires worth $20.1 trillion, up from $16.1 trillion a year earlier, with Elon Musk first at $839 billion and AI-linked fortunes (Huang, Ellison, Page, Brin, Zuckerberg) crowding the top. Nvidia alone was worth $5.5 trillion in May 2026. The IMF warns AI is likely to widen inequality within and between countries.
More worried than excited, and it is growing
Pew’s September 2025 survey of 5,023 adults found 50% more concerned than excited about AI in daily life, up from 37% in 2021, with 10% more excited; about 60% want more control over how AI is used in their lives. In Pew’s later 2025 reports, 57% rated the risks of AI as high and 44% said they trust the federal government to regulate it; those two figures come from the writer’s brief and are listed for checking. Gallup’s 2026 series found 70% opposed to a local data center, growing caution about business use of AI, workers who use it often twice as likely to fear displacement, and roughly as many employees saying it worsened workplace culture (25%) as improved it (24%).
The United States is an outlier. Worldwide, 59% of people surveyed for Stanford’s 2026 AI Index said AI’s benefits outweigh its drawbacks. Americans are among the most skeptical publics, which shapes the state-by-state fights on the policy page.
Concerned or excited?
Share of U.S. adults, Pew Research Center
The remainder said they feel equally concerned and excited. Fielded June 9–15, 2025.
The costs above are measurable. The next pages cover the ones that are not.
Safety, misuse and the people sounding alarms
Chatbot harms and lawsuits, deepfakes in elections, cyberattacks by AI agents, and the debate over whether the technology can be controlled at all.
Read the risks pageWhat a citizen can do about any of this
Who represents you, which committees matter, what the pending bills say about data centers and ratepayers, and the organizations working on each side.
Go to the action page