Core argument: Nobody can tell you what AI will look like in ten years — the field’s dated predictions have missed too often, experts included, to take on faith. What you can do with confidence is read the forces: where the money goes, what is being deployed versus demonstrated, what regulators require, and what resources are running out. Five forces are shaping AI’s next decade, and none of them depends on any single prediction coming true.
This is the seventh chapter of an ongoing series on world models. Previous chapters: why LLMs hallucinate, world model vs LLM, how world models work, who is building them, what artificial general intelligence really means, and a tracker of every AGI forecast. The series hub holds the full map.
1. The short answer
Ask what AI will look like in ten years and most answers you will find are predictions: dates and demo videos. Predictions in this field fail for a structural reason — they forecast single events, and a single event depends on timing that nobody controls.
Forces are different. A force tells you the direction everything is moving regardless of the news cycle, while the arrival dates stay open. Five forces are currently shaping AI’s decade. Each can be verified in spending, deployment, or regulation data, without trusting anyone’s roadmap:
- Capital at wartime scale — hundreds of billions per year, rising.
- The reliability gap — what companies deploy versus what they demo.
- Regulation arriving — phased, uneven, and starting to bind.
- Energy as the constraint — electricity demand becoming the bottleneck.
- World models and the physical world — AI moving from text into the world.
If you came for dated predictions, the tracker in chapter 6 keeps every one of them on the record.
2. Weather versus climate
There is a reason ten-year technology predictions fail, and it is the same reason you should not plan a picnic around a forecast made six months out.
Short-term weather forecasting works because atmospheric physics is stable and the initial conditions are well measured. Long-range climate forecasting also works, but differently: rather than predicting June 12, 2035, it predicts that the direction of travel — warming — will continue, because the underlying forces are understood and slow to reverse.
Technology commentary constantly confuses these two. “AI will pass the bar exam by 2027” is a weather claim. “Training costs are falling, so capability thresholds will keep moving” is a climate claim. The first kind has failed repeatedly in this field. The second kind is checkable today — in spending reports, deployment surveys, and grid data.
The analogy breaks in one place that matters. Climate bends only to billions of small decisions accumulated over decades. AI’s forces bend to a few hundred balance sheets and a handful of legislatures — a funding squeeze, one large regulatory fine, or buyers refusing to pay for unreliable agents could slow any of them within a year or two. That is why the honest frame treats them as forces with error bars.
3. The five forces, with evidence
Force one: Capital at wartime scale
The four largest hyperscalers (Amazon, Microsoft, Alphabet, and Meta) entered 2026 planning roughly $650 billion of capital expenditure, more than 60 percent above the prior record, and kept raising the figure as the year went on; Amazon alone lifted its target to $220 billion by mid-year. Microsoft and Meta together plan to spend roughly $300 billion, on the order of one percent of United States GDP. A budget on that scale is an infrastructure program. Whatever AI becomes by 2035, the physical substrate — data centers, chips, power — is being built now, and that substrate does not disappear if sentiment cools.
Force two: The reliability gap
Two 2026 surveys describe two different worlds. One enterprise survey found that 97 percent of executives say their companies deployed AI agents in the past year, though only about half of employees report using them. Gartner’s count of organizations that have deployed agents is 17 percent, and Gartner projects that more than 40 percent of agentic AI projects — software that plans and acts instead of merely answering — will be canceled by the end of 2027, citing escalating costs and unclear business value. Both numbers are true, and the space between them is where the next several years of enterprise AI will be decided. The vendors that close it first will take business from whoever has the flashier demos.
Force three: Regulation arriving
The European Union’s AI Act became generally applicable in August 2026, when its transparency obligations and enforcement machinery started to bite — with a year-end grace period for marking systems already on the market; obligations for high-risk systems were pushed to December 2027 by the 2026 Digital Omnibus, with the rest following through 2028. Whatever one thinks of the rules, the era of AI as an unregulated frontier in major markets is ending. The forces above will keep moving under regulation, but at a price: compliance costs, liability structures, and market access will all favor incumbents who can afford the process.
Force four: Energy as the constraint
Data centers already consume roughly 1.5 percent of global electricity (around 4 percent in the United States as of 2024), with demand projected to more than double by 2030; McKinsey puts all US data centers at a potential 11.7 percent of national electricity demand by then. Energy is becoming the binding constraint on AI growth. The decade will be shaped as much by utilities and grid planners as by model releases.
Force five: World models and the physical world
The reason this series exists: AI is being pushed out of text and into the world. Earlier chapters argued that today’s language models lack world models — internal simulations good enough to predict how objects, bodies, and places behave. In the physical world, that gap is the reliability gap: a warehouse robot that mispredicts the world answers by dropping the box. Simulation-trained robots and world models for autonomous driving are where this force is already visible, and it is the one that connects all the others. It is also the least mature, which is why its trajectory matters most for the back half of the decade.
Read the forecasts in chapter 6’s tracker against these five forces and the question “what would have to be true” gets a concrete answer: every dated AGI claim is, in disguise, a claim that the capital keeps flowing, the reliability gap closes, the energy gets built, the rules settle into a price, and the systems earn their way into the physical world.
4. My take
Three readings, in descending order of confidence.
First, near certainty: the capital and energy forces are locked in for years regardless of any single company’s fate. Infrastructure on this scale cannot be un-built quickly, and the reliability and energy problems are physical.
Second, medium confidence: the reliability gap closes unevenly — fast in narrow, high-value domains, slowly everywhere else — and the enterprises that treat deployment as an engineering discipline rather than a press release will be the ones still standing when the cancellation wave Gartner warns about arrives.
Third, lower confidence but worth stating: the language of AI’s future will shift from capability to constraint — energy prices, queues to connect new data centers to the grid, compliance costs, insurance. When those become the headlines, the frontier era is ending and the utility era is beginning. That will be a better decade for builders than the current one: when constraints decide outcomes, disciplined teams beat funded ones.
5. Questions people ask
What will AI look like in 10 years? On current forces: far more computing capacity and much higher electricity demand; AI tools that are narrower in practice than demos suggest; higher compliance costs; and the first serious wave of AI that acts in the physical world. Which force bends first — money, reliability, regulation, energy, or physics — decides the details.
What will AI be able to do by 2030? Already locked in: much bigger compute and data center capacity. Compounding now: reliability in narrow, high-value domains. Still open: whether world models make machines capable in the physical world. Treat any list of dated promises the way this page does: with an error bar.
Should I trust any ten-year AI prediction? Trust the forces, not the dates. Check whether a claim is backed by spending you can verify, deployment data with methodology, or regulations with legal force. Dated capability predictions in this field have missed repeatedly, and the AGI forecast tracker on this site shows the pattern.
Will the AI investment bubble burst? Some of it will. When two companies alone spend on the order of one percent of US GDP in a year while their free cash flow compresses toward zero, a correction follows from the arithmetic alone. But corrections prune; the data centers and grid capacity remain, as does the knowledge. The dot-com comparison helps up to a point: the 2000 crash did not stop the internet — the fiber built then still carries traffic today. It misleads in one way: this wave’s spending is concentrated on just four balance sheets.
How should I prepare for AI’s next decade? Watch two things: how reliable the AI tools you already use are in practice, and what is happening to their costs, which are set by energy and regulation. Careers and businesses that track the forces will be ready sooner than those that track announcements.
Sources
- Bloomberg, “How Much Is Big Tech Spending on AI Computing? A Staggering $650 Billion in 2026”, February 6, 2026 — the combined capex baseline.
- CNBC, “Amazon (AMZN) Q2 earnings report 2026”, July 30, 2026 — Amazon’s capex guidance raised to $220 billion.
- European Commission, “AI Act” — the official application timeline, including the 2026 Digital Omnibus rescheduling of high-risk obligations to December 2027.
- Writer, “Enterprise AI adoption in 2026: Why 79% face challenges”, April 7, 2026 — the survey behind the 97 percent executive figure.
- Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, June 25, 2025; the 17 percent deployment figure is from Gartner’s 2026 CIO and Technology Executive Survey, as reported in its 2026 Hype Cycle for Agentic AI coverage.
- IEA, “Energy and AI”, April 2025; press release — data centre electricity demand projected to more than double to around 945 TWh by 2030.
- Pew Research Center, “What we know about energy use at U.S. data centers amid the AI boom”, October 24, 2025 — the 4 percent US share as of 2024.
- McKinsey, “How data centers and the energy sector can sate AI’s hunger for power”, September 17, 2024 — 11.7 percent of US power demand by 2030.
Next in this series: does video generation understand the world? The Sora debate, tested against the physics benchmarks.
This page belongs to the world models series. Related: the hub · why LLMs hallucinate · world model vs LLM · how world models work · who is building them · what AGI really means · the AGI forecast tracker