What Energy, Capital and Compute Say About When AGI Arrives
Electricity, capital and compute are pouring into AI at a pace with no precedent, yet a model’s release date predicts its ability better than any of them: ten times the compute buys about four months. Around 2030, electricity and then money start to limit the biggest training runs.
Your prediction
Every forecast of Artificial general intelligenceAI that can do most of the work a skilled person can. This article uses the length of task an AI agent can complete as a measurable stand-in. rests on three raw inputs. Energy went first: the world’s data centres used about 415 terawatt-hours of electricity in 2024, about 1.5% of all the electricity the world consumed, and the International Energy Agency expects that to more than double by 2030. Capital followed: Alphabet, Amazon and Meta’s Capital spendingMoney spent on things that last for years, such as data centres, chips and the buildings and power lines around them, rather than on running costs. came to $289 billion in 2025, about 17 times what it was a decade ago, and with Microsoft the four biggest spenders plan at least $735 billion for 2026. Compute grew fastest of all: each year’s biggest AI training run has used about four times the compute of the year before, 1.3 million times more over ten years.
The bet behind all of it is a chain. More capital buys more chips and electricity, more chips and electricity train bigger models, and bigger models get smarter, until one of them can do a skilled person’s job. This article analyses the arrival of AGI from these fundamental inputs, testing each link of the chain against the public data.
The chain holds in total and breaks in detail: a new AI model’s release date predicts its capability better than the compute that trained it. Models get better every year by a fairly steady amount, almost whatever they’re trained with, because methods, data and tools keep improving. Ten times the compute puts a model only about four months ahead of others released at the same time. Around 2030, the electricity and then the money needed for the biggest training runs start to run short. For business, the constraint is whether AI finishes whole jobs reliably, rather than how capable it is on single tasks.
The aim was to see how energy, capital, compute and revenue measure up against each other, and whether the money flows back through the economy at all: a reference point, to be taken with a pinch of salt, for how AI is shaping the world. AI is reshaping it before it touches most people’s lives, just by moving money, chips and electricity around.
Capital and compute climb together, but the numbers can’t show capital setting the pace
Capital spending and compute have both soared over the decade. But the years when spending grew fastest weren’t the years compute grew fastest, not even a year or two later. Year by year, the public numbers can’t show the money setting the pace.
Over the last ten years, the three companies’ capital spending rose about 17 times, from $17 billion in 2015 to $289 billion in 2025, about 33% a year. The Training computeThe total number of FLOPs (floating-point operations, the calculations a chip performs) used to train an AI model. It’s the main thing a training budget buys: more chips, running for longer. used in each year’s largest training run, the single biggest job of training an AI model that year, rose far faster: about four times a year, 1.3 million times over the decade. Across all leading models, Epoch AI’s trend from 2018 to 2024 is about 4.5 times a year. The price of the largest run nearly tripled each year.
The two climb together as yearly totals and move independently as yearly growth.
Capital and compute rise together, but not in step.
One dot per year, 2015 to 2025: Alphabet, Amazon and Meta’s capex against the FLOPs in that year’s largest training run, both on log scales. “Yearly growth” plots each year’s change instead.
When Big Tech spends faster, does training compute grow faster?
Two series that both grow for a decade will always look linked. The fair test is whether their yearly changes move together, and they don’t: the correlation is −0.27, where 0 means no link. Allowing one or two years for capital to turn into compute doesn’t change that. With ten years of data, that rules out a strong link; a weak one could still hide in the noise. The same goes for R&D budgets.
Most of that capital hasn’t reached the books yet
Alphabet, Amazon and Meta spent $289 billion on capex in 2025 and booked $81 billion of depreciation, 28% as much. Most of the bill for the AI build-out arrives over the next five or six years, whatever AI does in the meantime.
Part of the reason spending and compute don’t move together year by year is that a year’s spending isn’t a year’s cost. When a company buys a chip or builds a data centre, the cash leaves at once, but accountants spread the cost over the years the asset is expected to work, a charge called depreciationThe slice of an asset’s purchase price that counts as a cost each year. A $6 billion batch of servers written off over six years costs $1 billion a year on paper, whatever was paid for it and when.. In the three companies’ accounts, a server now lasts five to six years: six at Alphabet, five to six at Amazon, five and a half at Meta. A building lasts anywhere from 7 to 40. So the $81 billion booked in 2025 is mostly the cost of chips and buildings bought in earlier years, and 2025’s own spending shows up as a cost mostly between 2026 and 2031.
Most of the capital spent on AI hasn’t reached the profit and loss account yet.
One dot per year, 2015 to 2025, for Alphabet, Amazon and Meta together: capex (blue) against the depreciation booked that year (red). The hollow dot is their 2026 plans. “Share booked” plots depreciation as a share of the same year’s capex.
How much of what Big Tech spends on AI has reached its costs?
The gap between the two lines is the build-out that hasn’t been paid for yet, in accounting terms. In 2015 Alphabet, Amazon and Meta booked $11 billion of depreciation against $17 billion of capex, about two thirds. In 2025 they booked $81 billion against $289 billion, 28%. Since 2015 they’ve spent just over $1 trillion and booked $420 billion. At the end of 2025 they carried $780 billion of buildings, chips and equipment on their balance sheets, and about $200 billion of it was still under construction or not yet switched on, so it isn’t being depreciated at all.
Depreciation is now one of the fastest-growing costs these companies have. It rose 29% across the three companies in 2025, and 38% at Alphabet alone. Alphabet told investors its depreciation would grow faster still in 2026, and Meta that infrastructure, depreciation included, would drive most of its cost growth. Amazon’s cloud division booked 61% more depreciation in 2025 than in 2024, part of it from shortening some servers’ lives, while its sales grew 20%. The 2026 plans alone, about $505 billion across the three, would add roughly $55 billion a year of depreciation for five or six years from the servers in them, if Alphabet’s split of 60% servers and 40% buildings and networking holds for the other two.Alphabet and Amazon rent their chips out through Google Cloud and AWS, which brought in $59 billion and $129 billion in 2025, and Alphabet reported $240 billion of signed cloud contracts waiting to be delivered. Meta has no cloud business, so its chips have to pay for themselves through advertising.
The assumed life of a server is an estimate, and changes to it have moved reported profit by billions. Alphabet stretched its servers from four years to six in 2023, which cut that year’s depreciation by $3.9 billion. Amazon went from four to five in 2022 and five to six in 2024, saving $3.6 billion and $3.2 billion, then pulled some of its servers and network equipment back to five years in 2025, adding $1.4 billion. Meta stretched some of its servers from at most five years to five and a half in January 2025, which cut its 2025 depreciation by $2.9 billion and lifted its profit by $2.6 billion. Through the years the industry has argued that AI chips wear out faster, all three lengthened their server lives at least once, and only Amazon has pulled any back.
Depreciation, a smoothed-out measure of the compute these companies are consuming, tracks the largest training run no better than spending does: the correlation between yearly growth in depreciation and in the largest run is 0.11, about as close to zero as spending’s −0.27. What the accounts add is timing: when the spending shows up as a cost. The frontier training run is a lump of compute used for a few months, while the money behind it is paid for over six years by a fleet that mostly serves cloud customers and everyday products. The hardware for 2028’s largest run, about $27 billion, would be about 2 to 3 percent of what Alphabet, Amazon and Meta are on course to spend that year.
Ten times the compute buys a few months
Take two models released at the same time, one trained with ten times the compute of the other. The bigger one is only about four months ahead, and possibly not ahead at all.
Compute, unlike company spending, can be compared model by model, because Epoch AIAn independent research group that tracks AI models: how much compute trained them, what they cost and how they score. estimates it for dozens of models. Among models released at the same time, one trained with ten times the compute scores only about 3% higher on Epoch AI’s Epoch Capabilities IndexA single score for how capable an AI model is, from Epoch AI. It combines results from dozens of benchmark tests onto one scale, on which the best model today scores about 30% higher than GPT-4 did in 2023.. The best score rises about 9% a year, so a 3% edge is about four months of ordinary progress. Ten times the compute costs about ten times as much, so a lab that spends ten times more on training buys a lead of about four months. That’s the middle estimate: it could be anywhere from one month to seven, and on the models whose compute figures Epoch AI rates as most reliable, it could be zero.
Adding compute to a forecast makes it worse. When the models released up to a cut-off date are used to forecast the capability scores of the models released after it, forecasts from the release date alone come out more accurate than ones that also use each model’s compute: its effect is small and uncertain, so it added more noise than information. The same holds with a delay: no lag of up to two years shows a reliable link between a jump in compute and a later jump in capability. Where the two jump together, it’s mostly the same model setting both records in the same month.
The release date stands in for everything that improves with time: better training methods, models that reason step by step before answering, better tools around them, better data, and bigger training runs too. Nobody outside the labs can separate these, so the release date stands for all of them together.
The late-2024 “reasoning” leap looks like ordinary yearly progress, split in two
The famous late-2024 jump in AI ability appears wherever you split the models by date, so the public data can’t show that reasoning models caused it.
In September 2024 OpenAI released o1, the first widely used “reasoning” model: one trained to work through a problem step by step before it answers. The popular story is that AI stepped up then. Newer models do score about 15% higher than older ones trained with the same compute, half the entire climb from GPT-4 to today’s best model.
But scores also rise every year at any given amount of compute, by about 8% a year. The models after September 2024 came out about 1.8 years later, on average, than those before it, and 1.8 years of that ordinary rise is about 15%: the whole gap. Any date splits a steadily rising group into an older, lower half and a newer, higher half, and the two halves will look like a jump. The chart below measures that gap for every dividing date from late 2023 to mid-2025, as measured and with the yearly rise removed.
Newer models lead by about 15% whatever date divides them, and the lead vanishes once the yearly rise is removed.
Across the bottom runs the date used to split the models into older and newer, late 2023 to mid-2025. The lines show how much higher the newer group scores at the same training compute: as measured (blue) and after removing the ordinary yearly rise (red).
Did AI get a sudden boost when “reasoning” models arrived in late 2024?
Wherever the dividing date goes, the measured gap stays near 15%, and with the yearly rise removed it stays within 3% of zero.
Reasoning may well have mattered: whatever it added is part of the yearly gain, along with everything else that improved over the same months. Late 2024 is not a special date, since splitting the models at any month gives the same jump. Separating reasoning’s share would need labs to say how much compute goes into the extra training that teaches a model to reason, and almost none do.
AI matches professionals on single tasks and still fails most whole projects
AI now matches professionals on single pieces of work, but it still fails four in five real freelance projects, and the economy has barely felt it.
What a business needs is whole jobs done reliably, which a high test score alone can’t show.
On single pieces of work, AI has caught up. On GDPvalA test from OpenAI built from real work, such as reports, plans and spreadsheets, across 44 occupations. Experienced professionals compare each AI deliverable with one made by a professional, without knowing which is which. A score of 50%, counting wins and ties, means level with the professional., the best models drew level with experienced professionals between September and November 2025, and the best model measured so far wins or ties about 71% of the time. On whole projects AI is still far behind. On the Remote Labor IndexA test from the Center for AI Safety and Scale AI built from real freelance projects. It counts a project as done only if the whole deliverable would satisfy the client., those same late-2025 models completed only 1 to 4% of real freelance projects. Newer models have climbed fast, from about 2.5% in October 2025 to about 21% in September 2026, but four projects in five still fail.
The gap is reliability on long jobs. An AI agentAn AI model set up to carry out a task on its own, step by step, using other software, rather than just answering a question. that completes 10-hour tasks half the time can be relied on, four times in five, only for tasks of about 2 hours, a fifth as long. That gap hasn’t narrowed since 2023. Better tools help, but they haven’t closed it. Add-ons that let AI agents use other software are multiplying: the official public directory of them doubles in size about every 2.4 months, about 30-fold a year, and good add-ons raise the same model’s success rate on real tasks by 9 to 25 percentage points.
The effect on the economy is still small. US labour productivity, the output of an hour’s work, has grown about 1.8% a year since 2019, and AI accounts for about a seventh of that growth. At the pace companies have taken up AI so far, AI’s contribution stays below one percentage point a year through 2050 in nine of ten simulations. There may be an early sign in entry-level jobs, though it isn’t proof yet. In the quarter of jobs where AI is used most, employment of 22- to 25-year-olds has fallen about 10% since 2021–22, relative to older workers and to jobs with no measured AI use.That comparison was chosen after seeing the data. The comparison chosen in advance, against the quarter of jobs where AI is used least, is too uncertain to say either way: anywhere from a 15% fall to a 9% rise.
On the current trend, AI handles a month of work by late 2028 and a year of work by 2030
On the middle estimate, AI agents complete a month of skilled work half the time they try by late 2028, and a year’s work by late 2030. The range runs from 2027 to the late 2030s.
AGI has no agreed definition, so this analysis uses one that can be measured: the length of task, timed by how long it takes a skilled person, that AI agents complete in half of their attempts. METRAn independent research non-profit that times how long tasks take skilled people, then measures which of those tasks AI agents can complete. tracks it. In 2019 it was a few seconds. The trend through METR’s measurements puts it at about 17 hours today, somewhere between 12 and 35. A working month is about 167 hours, and a working year about 2,000. Since 2019 it has doubled about every 7 months. Since 2024 it has doubled about every 4, but that faster pace disappears without OpenAI’s models, so it gets less weight.
At the long-run pace, AI completes a month of skilled work, half the time, by late 2028.
One dot per model: release date against the longest task it completes half the time, timed by how long it takes a skilled person, on a log scale. The red line carries the trend forward from about 17 hours today; the dates mark a working month and a working year. The slider sets the doubling time: about 7 months over 2019 to 2026, about 4 since 2024.
When will an AI agent finish a task that takes a person a month?
The forecast blends three paths: the long-run pace, with half the weight, and the faster recent pace and a path where progress gradually slows, with a quarter each. The fast and slow paths pull in opposite directions, so the middle estimates match the long-run pace: September 2028 for a month of work and October 2030 for a year. What they change is the range. With eight chances in ten, a month of work arrives between mid-2027 and late 2029, and a year’s work between late 2028 and the late 2030s. At the faster recent pace alone, a month comes in 2027 and a year by late 2028.
When the method is tested on the past, forecasting from mid-2024 and again from the start of 2025, the record-setting models that came out afterwards often beat the forecast’s range and never fell below it. If anything, these dates lean late.
The forecast runs on the calendar alone, because adding spending or compute did nothing to improve forecasts of the past. What could pull the dates earlier is AI speeding up AI research, by writing code and running experiments for the people building the next models. How much earlier rests on assumptions nobody can measure yet.Dario Amodei, who runs Anthropic, has predicted “powerful AI” by early 2027. Read as AI handling month-long tasks by March 2027, this forecast gives that about a 1% chance.
The forecast can be checked within a year. The trend says models measured now should manage about 17 hours. If METR’s measurements of late-2026 models come in above about 40 hours, the faster recent pace is winning. Below about 12 hours, progress is slowing.
By about 2030, electricity and then capital cap how big a training run can get
Around 2030 the biggest training runs will need more electricity, and then more capital, than anyone can easily supply. That slows the growth of the biggest runs; how much it slows AI itself can’t be told from this data.
If the largest training run keeps growing 4.5 times a year, the chips and servers to train it, bought up front, cost about $27 billion in 2028. That’s 2% of what the four biggest spenders, Alphabet, Amazon, Meta and Microsoft, are on course to spend that year. By 2030 it’s about $182 billion, 9%, and by 2032 about $1.2 trillion, a third. The bill grows more slowly than the run, because chips keep getting cheaper for the same work, and longer runs get more out of each chip.
Electricity becomes scarce before capital does. Today’s largest run draws roughly 0.2 gigawatts. The same run, still growing, needs 1.4 gigawatts in 2028, about what a large nuclear reactor produces; 10 in 2030, about seven of them; and 75 in 2032, which would be 62% of all the electricity the world’s data centres are expected to use. These thresholds mark where supply starts to strain, and all of them can move with enough effort. Electricity counts as the limit once one run needs 5 gigawatts or 5% of the world’s data-centre electricity, whichever comes first: that happens around 2029 to 2031. Capital counts as the limit once one run’s hardware costs more than a quarter of the four biggest spenders’ capital spending: around 2031 to 2032. Spreading one run across many data centres would ease the electricity limit, since no single site would need all of it. Once the limits bite, the largest runs can grow only as fast as spending and chips’ value for money together allow: about 1.8 times a year, instead of 4.5.
More spending can push the capital limit out, at a price. If the four biggest spenders grew their capital spending by 80% a year, rather than the third a year Alphabet, Amazon and Meta managed over the past decade, the capital limit would move out by two years or more. That would mean about $7.7 trillion in 2030, about ten times the $735 billion planned for 2026, and electricity would still run short before capital does.
The forecast above doesn’t include these limits. The month-long date comes before they bite; the year-long date lands just as they start. How much they would slow progress can’t be told from this data, because bigger runs improve together with everything else that comes with time, and nobody outside the labs can separate them.
What this means if you’re deciding where to invest or what to build
- A bigger training budget buys a lead of months, and rivals close it with time. Ten times the compute puts a model about four months ahead of rivals released at the same time.
- The bill for the build-out arrives from 2026 to 2031. Alphabet, Amazon and Meta booked depreciation in 2025 equal to 28% of their capex that year. What’s spent now comes through as a cost over the five or six years after, whatever AI does. The companies that rent their chips out have revenue to set against it. The ones that don’t are betting on their own products.
- Plan around late 2028 and 2030. On the middle estimate, AI agents complete a month of skilled work half the time by late 2028, and a year’s work by late 2030. From about 2030, electricity and then money slow the growth of the biggest training runs.
- The opportunity is making AI reliable on long jobs. AI matches professionals on single pieces of work, but the best model completes only about one real freelance project in five.
How the study was run
The research ran in four rounds, each answer setting the next question: first what goes in (capital, compute, electricity), then the gap between what AI can do on tests and at work, then what had been missed, then whether the shortcuts taken held up. Four choices make the results worth trusting.
It was run as a team. AI agents with defined jobs and budgets did the work: librarians gathered data, an auditor checked it against sources, a modeller ran the numbers, a forecaster projected them and a sceptic tried to break everyone else’s work. The most capable AI went to the jobs that needed judgment, and cheaper AI to data collection.
What would prove the thesis wrong was decided before looking. Each round’s main predictions, and the results that would sink them, were written down before the analysis of that round’s data began.
A sceptic was on the team. Before a single result existed, it caught four bugs in the statistics code. The worst tripled how often a test would raise a false alarm, from 5% of the time to 15%. All four were fixed before any results came out. In the last round the team tried about 117 alternatives to the plain trend, such as grouping the models by company, by country or by whether anyone can download them, to see if any group followed its own path. None predicted new models’ scores better than the plain trend over time.
Every source is one anyone can check: company filings, the International Energy Agency, chip datasheets, Epoch AI, METR and peer-reviewed papers. Every one of the roughly 1,300 data points links to where it came from. The code, data and working behind every figure are public on GitHub, in the agi_timelines repository.
The release date of the next leading AI model will say more about its capability than its training budget.
The answer
How long do you think it takes for the computing power behind the world’s biggest AI models to double?
5.5 months
Source: Epoch AI, Notable AI Models: the largest training runs grow about 4.5× a year