What is the future of AI?
Nobody knows, and anyone giving you a confident timeline is selling something. What's actually locked in: the physical buildout. The IEA projects data centre electricity roughly doubling from 415 TWh in 2024 to about 945 TWh by 2030, and the five largest hyperscalers are committing over $1 trillion in AI capex across 2025–2026. The capability curve is the genuinely open question.
Why — the first-principles explanation
Separate the question into three layers that move at completely different speeds, and the fog clears a lot.
The physical layer is the most predictable, because concrete and turbines obey schedules. Data centres are being built, grid connections queued, chips fabbed. The IEA's numbers are the closest thing to a fixed point: about 415 TWh in 2024 — roughly 1.5% of world electricity — heading to about 945 TWh by 2030, with data centres accounting for nearly half of all U.S. electricity demand growth through 2030. You can argue about the exact figure; you can't argue that it's being built, because you can go look at it. This layer is largely determined for the next 3–5 years.
The capability layer is where honesty is required. Recent progress came from scaling — more compute, more data, bigger models — and the central open question is whether that keeps paying. There are two credible camps and neither has won. One says scaling plus new tricks (reasoning at inference time, better tool use) keeps compounding toward systems that do most cognitive work. The other says the internet's supply of high-quality text is largely consumed, returns are bending, and we're heading into a plateau of useful-but-not-transformative tools. The evidence to settle this does not exist yet — that's not fence-sitting, it's the actual state of knowledge.
The economic layer is where the tension shows. Adoption is genuinely fast, yet measured returns lag badly: an MIT study of 300+ enterprise AI initiatives found roughly 95% delivered no measurable profit-and-loss impact. Meanwhile the BIS — the central banks' central bank — used its June 2026 annual report to compare this buildout to canal mania, British railway mania, electrification in the 1920s, and the dotcom boom. Its point is precise and worth getting right: those were all real technological revolutions that also attracted more capital than returns could justify. Railways genuinely transformed Britain, and railway investors were still wiped out. Both things happened.
So the defensible shape: the technology is real, the infrastructure is real, and the financing may well be overbuilt. History's pattern for general-purpose technologies is a capital bust followed by broad diffusion — the fiber laid in 1999 carried YouTube in 2006, just not for the people who paid for it.
An example that makes it click
Picture a gold rush town in 1849.
Three things are happening at once and people keep confusing them. First, there really is gold — that part isn't a rumor, you can hold it. Second, someone is building the railroad, the hotels, and the supply depots, and that construction is visible, scheduled, and enormously expensive. Third, a hundred thousand people have bought claims at prices that assume every shovel hits a vein.
All three can be true simultaneously. The gold is real and most claim-holders lose their money and the railroad still stands after the rush ends and carries freight for a century. When someone asks "what's the future of the gold rush?", they're usually asking about the claims. The honest answer is: the gold's real, the railroad's getting built either way, and nobody can tell you which claims pay off.
Key facts
- The IEA reports data centres consumed about 415 TWh in 2024 (~1.5% of global electricity) and projects about 945 TWh by 2030 — roughly Japan's total current consumption; the US accounts for 45% of the global share, China 25%, Europe 15%.
- Data centres will account for nearly half of all US electricity demand growth through 2030, per the IEA.
- The BIS Annual Economic Report (June 2026) states the five largest hyperscalers are set to spend over $1 trillion on AI-related capex across 2025-2026, outpacing earnings and free cash flow, with some issuing debt to finance it.
- The BIS compares the AI buildout to canal mania, 1840s British railway mania, 1920s electrification, and the 1990s dotcom boom — episodes that combined genuine breakthroughs with capital in excess of justifiable returns and ended in investment reversals.
- An MIT study of 300+ enterprise AI initiatives (July 2025) found roughly 95% delivered no measurable P&L impact, on $30-40 billion of enterprise generative AI spending in 2024-25.
- IEA projects data centre CO2 emissions of 300 Mt (Base Case) to 500 Mt (Lift-Off Case) by 2035, remaining below 1.5% of total energy sector emissions.
▶ The 60-second explainer (script)
The future of AI? Nobody knows — and anyone handing you a confident timeline is selling something. But that's not the end of the answer, because the question actually splits into three layers that move at totally different speeds. The physical layer is the most predictable, because concrete and turbines run on schedules. The International Energy Agency puts data centre electricity at about 415 terawatt-hours in 2024 — one and a half percent of world electricity — heading toward roughly 945 by 2030. In the US, data centres account for nearly half of all electricity demand growth this decade. You can argue the exact number. You can't argue it's being built. You can go look at it. The capability layer is where honesty is required. Recent progress came from scaling — more compute, more data, bigger models. Does that keep paying? Two credible camps, neither has won, and the evidence to settle it doesn't exist yet. That's not fence-sitting, that's the actual state of knowledge. Then the economics. An MIT study of over three hundred enterprise AI projects found about ninety-five percent showed no measurable profit impact. And the Bank for International Settlements — the central banks' central bank — compared this buildout to railway mania and the dotcom boom. Read that carefully. Their point is that those were real revolutions that also attracted more capital than returns could justify. Railways transformed Britain and railway investors were wiped out. Both happened. That's the honest shape of it.
What authoritative sources say
People also ask
Will AI reach human-level intelligence, and when?
No one knows. Forecasts from serious researchers range from a few years to many decades to never, and that spread reflects genuine disagreement about whether current scaling keeps working — not a lack of effort. Treat any precise date as marketing.
Is AI progress slowing down?
Contested. Some argue high-quality training text is largely exhausted and returns are bending; others point to gains from reasoning at inference time and tool use. Both camps have real evidence and neither has been settled.
Is the infrastructure buildout more certain than the capability?
Yes, substantially. Data centres, grid connections, and chip fabs are physical projects on multi-year schedules. The IEA's projection of roughly doubling data centre electricity by 2030 is the most solid claim in this whole area.
Does the BIS say AI is a bubble?
It draws a careful parallel rather than a verdict. It notes canal mania, railway mania, electrification, and dotcom all combined genuine technological breakthroughs with capital exceeding justifiable returns, and warns the current boom resembles them — highlighting downside risk, not predicting a date.
What's the most reliable thing to say about AI's future?
That electricity demand, capital spending, and physical buildout are largely locked in for the next several years, while capability trajectory and financial returns are genuinely open. The technology being real and the financing being overbuilt are not contradictory.