Up to 2035, the development of artificial intelligence (AI) can be described by three branches. Branch A is “normal technology” that spreads over decades, branch B continues the current trend and is taken as the baseline, and branch C is acceleration, in which AI automates its own research.
Which branch is coming true is checked not by opinion but by external benchmarks: the price of a token, the horizon of autonomous tasks, the electricity use of data centers, the share of decisions taken by agents. For a company the conclusion is common to all three branches: there are already more machine actions per person than a person can check, so the rules, the keys and the proof of actions stay with people in any scenario.
In branch A capability growth slows, and AI is adopted gradually and mostly complements human work. In branch B compute grows 4–5 times a year, the price of a token at the same quality falls about 10 times a year, and agents take over operational execution. In branch C AI automates its own research, and the leap compresses the decade. Each indicator has a 2026 value, a source and a check date, and the register is recalculated every year. Control of agent actions and independent testing are needed in all three branches, and only the density of the action flow changes.
Three branches of AI development to 2035
| Branch | What happens | Control benchmark by 2028 |
|---|---|---|
| A, “normal technology” | Model capabilities grow more slowly; AI spreads over decades like electricity and the internet; institutions have time to adapt; automation is fragmentary | The task horizon on the scale of METR (Model Evaluation and Threat Research) doubles more slowly than every seven months; the Gartner forecast of a large share of agentic projects being canceled by the end of 2027 comes true; open models carry more than half of the token flow |
| B, the trend continues (baseline) | Training compute grows 4–5 times a year, the price of a token at fixed quality falls about 10 times a year, the autonomous task horizon doubles every 3–7 months; agents take over execution, people keep the rules, the keys and the responsibility | The best model sustains a METR horizon of at least a working week; 15% of daily work decisions are taken autonomously on Gartner’s scale |
| C, acceleration | AI automates its own research cycle; “a century of progress in a decade” | An automated coder appears before 2028; the growth of the gross domestic product (GDP) of the United States exceeds 5% a year; capital investment in AI exceeds 5% of GDP |
Branches and benchmarks follow the book of the institute “Proven Action” (introduction and Appendix G).
Key indicators and their benchmarks
| Indicator | 2026 value | Branches by 2035: A / B / C | Where to check |
|---|---|---|---|
| Token price at fixed quality | Falling about 10 times a year (a16z, 2024; Epoch AI, 2025) | Slowdown / 10 times a year / 100 times a year | Epoch AI, yearly |
| Task horizon of the best model (median) | About 12 hours for the best model (METR, 20 February 2026) | A month / a quarter to a year, a “turnkey project” / not measurable on the current scale | METR, yearly |
| Arrival of an automated coder | Medians of two AI Futures Project authors: November 2027 and January 2030 (update of 16 August 2026) | After 2035 / 2028–2030 / before 2028 | aifuturesmodel.com, quarterly |
| Data center electricity use worldwide | Forecast of the International Energy Agency (IEA): 945 TWh by 2030 and 1,200 TWh by 2035 | ≈ 1,000 / ≈ 1,200 / > 1,700 TWh | IEA World Energy Outlook, yearly |
| Energy-intensive computing capacity in Russia (data centers and mining) | 4.2 gigawatts by April 2026 (System Operator of the Unified Energy System) | 8 / 15 / 15 gigawatts | System Operator, Ministry of Energy, yearly |
| Share of daily work decisions taken by agents | Gartner forecast: 15% by 2028 (0% in 2024) | 10 / 30 / > 50% | Gartner, yearly |
| Employment of young workers aged 22–25 in AI-exposed occupations | −19% (Stanford Digital Economy Lab, 12 August 2026) | 0 / −20 / −40% | Stanford Digital Economy Lab, yearly |
| Agent liability policies that require an execution log | First precedent: the AIUC-1 standard (AIUC is the company that publishes it), ElevenLabs policy of 12 February 2026 | A narrow niche / a corporate market standard / required by law | Terms of the policies of leading insurers, yearly |
The full register of fifteen indicators is on the page of the forecast register to 2035 (in Russian).
What this means for a company
The price of a model falls, and value moves. If a token gets about 10 times cheaper every year (branch B), the ordinary 2026 level of intelligence becomes almost free within a few years. Money and responsibility move to where they cannot be made cheaper: approving rules, holding keys, proving actions and independent testing.
Power generation remains a hard physical boundary: scaling compute runs into available gigawatts, so both global data center consumption and Russian capacity grow according to the plans of power systems, not the wishes of developers. For Russia this is a question of access to compute and of sovereignty.
The structure of professional work changes unevenly: autonomous agents take over routine operational execution—a shift visible first in entry-level hiring—while people remain where rules are signed and legal responsibility is borne.
Control is needed in any branch. Even in the slow branch A, with a small share of autonomous decisions, a large company generates millions of machine actions a day, more than people can check. In branches B and C the gap grows by orders of magnitude. So a deterministic check of actions before execution and independent testing do not depend on which branch comes true (Ashby’s law and Beer’s model).
How to tell which branch is coming true
- Check the benchmarks once a year: Epoch AI for compute and token price, METR for the task horizon, the IEA and the System Operator for energy, Gartner for the share of agent decisions, and the Stanford Digital Economy Lab for employment.
- Look at the whole picture, not one number, because the branch is defined by several indicators together.
- Recalculate plans by branch, not by the forecast of one expert.
- Mark forecasts with a label and a benchmark, following the rule of truth marking.
Frequently asked questions
What are the scenarios for AI development to 2035?
Three branches: A, “normal technology” spreading over decades; B, the trend continues, with compute growing 4–5 times a year and token prices falling about 10 times a year; C, acceleration, in which AI automates its own research.
Which scenario is the baseline?
In the book of the institute “Proven Action” the baseline working branch is B, the trend continues. But the management conclusions are built to stay true in all three branches.
How can one check which scenario is coming true?
By external benchmarks: token price and compute (Epoch AI), the horizon of autonomous tasks (METR), the electricity use of data centers (IEA), the share of decisions taken by agents (Gartner) and the employment of young workers (Stanford Digital Economy Lab), with the entire register reviewed and recalculated annually.
How much will AI cost in a few years?
In branch B the price of a token at fixed quality keeps falling about 10 times a year, and the 2026 level of models becomes almost free. In branch A the fall slows down; in branch C it speeds up.
What will happen to people’s jobs because of AI agents?
Agents take over operational execution; according to the Stanford Digital Economy Lab, in 2026 employment of young workers aged 22–25 in AI-exposed occupations is already 19% lower. By 2035, depending on the branch, the range is from recovery through new roles to a 40% fall. People keep the rules, the keys and the responsibility.
Why is control of agent actions needed in any scenario?
The flow of machine actions per person has already exceeded the limit of attention even in the slow branch. Only the density of the flow changes, and checking every action by hand is impossible in any of the branches.
Sources
- The institute’s book “Proven Action. Power and Control over AI Agents: Strategy 2027–2035”, introduction and Appendix G “Forecast register to 2035” (in Russian): https://isslab.ai/book/
- Forecast register to 2035 (in Russian): https://isslab.ru/kniga/reestr-prognozov/
- Epoch AI, compute trends: https://epoch.ai/data/trends
- METR, task time horizons: https://metr.org/time-horizons/
- AI Futures Project, timelines model: https://aifuturesmodel.com/
- IEA, World Energy Outlook: https://www.iea.org/reports/world-energy-outlook-2025
- Stanford Digital Economy Lab: https://digitaleconomy.stanford.edu/
Checked: 18 October 2026, against the text of the book. Forecasts are forecasts, not measurements; the 2026 values follow the sources named in the book.