Ask ten researchers when we build AGI and you’ll get ten answers, most of them hedged twice. Ask when it’s plausible and the answers bunch up between roughly 2030 and 2050. The 2040 mark sits near the middle of that cluster, far enough out to absorb a few serious setbacks, close enough that most people reading this will still be mid-career when the question resolves.
So what is the bet, exactly? And what would have to go right — or wrong — for it to pay off?
What “AGI by 2040” Actually Claims
The phrase hides a bundle of separate predictions: that we’ll build systems able to do most economically useful cognitive work at roughly the level of a competent adult, that they’ll hold together over hours or weeks instead of minutes, and that the cost curve bends far enough for the capability to be widely usable rather than a demo behind an enterprise contract.
Miss any one of those and the headline gets qualified into something much smaller. A model that scores in the 90th percentile on every exam on earth but freezes after forty minutes of unsupervised work is not what most people mean by AGI.
The definitions stay contested, which makes the timelines slippery. Benchmarks like ARC-AGI, designed specifically to resist memorisation, went from near-zero to competitive in about two years. GPQA — graduate-level science questions — crossed expert-human territory in a similar window. Each new benchmark tends to hold for a year or two, then crumple, which is either evidence of rapid general progress or evidence that we keep measuring the wrong thing.
The Compute Curve Is Doing Most of the Work
Training compute for frontier models has grown by roughly four to five times a year for most of the past decade. Nothing else in industrial history scales that fast for that long, and it’s the single strongest argument for a 2040 arrival.
Not every hardware curve keeps its slope, though. Cameras are the cautionary example: we’re arguably reaching peak camera with the Sony A7R VI, where each generation adds polish instead of range, because physics caps how much light a sensor can gather. Compute hasn’t hit that wall — yet. The wall is real, and the interesting question is whether we hit it in 2032 or 2044.
Inference is the quieter half of the story. Prices per token for common tasks have fallen by about an order of magnitude a year, and small models now run comfortably on consumer silicon. That’s why a $450 laptop from an unknown brand can ship with a neural accelerator that would have been a datacenter part a decade ago. Cheap, ubiquitous inference changes the economics of AGI more than another record training run does.
Energy Is the Bottleneck Nobody Budgeted For
A frontier training run in 2025 drew tens of megawatts. The campuses under construction in Texas and Ohio are a different category entirely — hundreds of megawatts, moving toward gigawatts, with gas turbines bolted on because renewables and storage can’t be deployed fast enough.
The grid is the constraint. Interconnection queues in the United States run three to seven years in most regions, and transformers, which used to arrive in months, now take two years or more. Compute can be bought; electricity has to be built.
That tension is already bending corporate climate math. Companies that spent a decade promising to zero out emissions are now signing long-term gas contracts and quietly redefining their targets. It raises an awkward question worth putting to the firms that promised to eliminate carbon emissions while their power bills triple: what does the pledge actually mean now?
Reliability Is Where Timelines Usually Slip
Agents are the part of the AGI story that has to work for the 2040 prediction to mean anything. And agents compound errors.
Do the arithmetic. A system that completes each step correctly 95% of the time has a 36% chance of getting through twenty steps cleanly. At 99% per step, twenty steps land at 82%. Long-horizon autonomy is a brutally unforgiving engineering problem, which is why METR’s measurement work matters more than most benchmark leaderboards. Their data suggests the task length frontier models can complete reliably has been doubling roughly every seven months. Extrapolate that and 2040 looks generous. Assume it slows by half and it still lands somewhere consequential.
Nobody knows which holds. Every previous automation wave hit a plateau where the remaining 10% cost more than the first 90%.
What the Forecasters Disagree About
The 2024 survey of thousands of published AI researchers put the aggregate median at 50% probability of high-level machine intelligence by 2047 — up sharply from 2060 in the 2022 edition. Prediction markets and forecasting communities cluster in the 2030s. Scenario writers have sketched out aggressive paths that arrive before 2030 and slower ones that stretch past 2060.
The spread isn’t incompetence. It reflects genuine uncertainty about whether scaling keeps paying, whether data runs out, and whether the last stretch of reliability turns out to be a cliff rather than a slope.
Governance Moves at Its Own Speed
Policy rarely arrives before the technology it governs, and it often aims at the wrong target. The push to ban robot vacuums is a tidy illustration: regulation driven by perception rather than measured harm, producing rules that satisfy the public without improving anything.
Expect the same pattern with AGI. Bans on capabilities nobody has built, licensing regimes written for last year’s model class, and a lot of activity that looks like governance without much of it. Meanwhile the practical fights will be commercial. Access is already a negotiation — Anthropic telling OpenClaw users to pay up is a small preview of what happens once agentic tools become load-bearing infrastructure inside real businesses.
Signals Worth Watching Between Now and 2040
Forget the press releases. These are the measurable things that would move a forecast:
- Task horizon. When frontier models reliably complete week-long projects without human rescue, the reliability objection mostly dissolves.
- Cost per unit of work. Not cost per token — cost to actually finish a task. That number falling tenfold changes adoption more than any benchmark.
- Gigawatts under construction. Data center capacity with signed power contracts is a nine-figure bet placed by people who’ve seen the internal roadmaps.
- Benchmark half-life. If new evaluations stop collapsing within 18 months, something has genuinely changed.
- Independent audit access. Whether labs submit to outside evaluation with real teeth, or keep self-reporting.
What a 2040 Deadline Asks of You Now
The useful move isn’t to pick a date and wait. It’s to notice that both outcomes — arrival and delay — are survivable if you build for a world of continuous capability improvement rather than a single switch-flip moment.
Practically, that means treating AI tools as infrastructure you learn to operate rather than products you consume passively. It means keeping skills that compound with judgement, taste, and accountability, because those are the parts that stay scarce whether the timeline is 2038 or 2065. And it means reading the energy numbers and the reliability numbers, not just the launch announcements, since those are the two places where a 2040 prediction actually lives or dies.
Fifteen years is a long time in software and a short time in electricity. That asymmetry, more than any benchmark, is what will decide whether the AGI-by-2040 bet pays out.