Gertrud Goldschmidt (Gego), Untitled, 1966
Image: Gertrud Goldschmidt (Gego), Untitled (1966).

Five scenarios for thinking clearly about AI when the range of plausible outcomes is still unusually wide.

We do not have a consensus on the trajectory of artificial intelligence from 2025 to 2035. Below are five plausible scenarios, each mapped across three strata: compute and hardware, models and software, and applications.

I teach and speak about these dynamics at UCL School of Management and with firms such as 13Books Capital, which invited me to speak at its annual meeting in November 2023.

I would be the first to say that I have been surprised by the pace of recent progress. At times, it has felt as though a decade of capability gains has arrived every six months, with real shifts in the effectiveness, aptitude and efficiency of AI models.

On the available evidence—and without privileged access to frontier labs—I still find it hard to choose between these futures. My current base case is Scenario 3 because physical infrastructure and regulatory inertia look more constrained than demand.

Executive summary

Scenario 1 — Diminishing returns

Frontier gains plateau. Firms win by specialising models and tightening workflows.

Scenario 2 — Dot-com rhyme

Valuations reset. The surviving builders concentrate on adoption with demonstrable returns.

Scenario 3 — Bottlenecked boom

Energy, chips and regulation throttle the frontier. Firms learn to manage constraints as strategic inputs.

Scenario 4 — Compounding automation

There is no singular breakthrough, but steady adoption of agentic systems lifts productivity year after year.

Scenario 5 — Transformative takeoff

Breakthroughs unlock long-horizon AI systems that reshape research, development and knowledge work.

These scenarios are not mutually exclusive. In practice, the next decade may look like “2 + 4” or “3 + 4”: a market correction or infrastructure bottleneck alongside steadily compounding adoption.

A brief glossary

  • Tokens: the units of text, code or other data that a model processes and generates.
  • Retrieval-augmented generation (RAG): a system that retrieves relevant information from an external source before asking a model to answer.
  • Evaluations (evals): repeatable tests used to measure model or system quality, reliability and safety.
  • Agentic systems: AI systems that can plan, use tools and take a sequence of actions towards a goal.
  • Distilled models: smaller models trained to reproduce much of a larger model’s behaviour at lower cost.

Scenario 1 — Plateau without a crash

What it looks like

The capability gains produced by ever-larger models slow visibly. Adding data and compute yields only minor improvements. Value shifts towards small, specialised and on-device models, along with better product design. AI remains essential, but the frontier-model arms race cools.

The three strata

  • Compute and hardware: capital expenditure flattens, GPU spot prices stabilise and inference efficiency becomes the priority.
  • Models and software: returns from scaling taper. Progress shifts towards retrieval, tool use and evaluation.
  • Applications: user-experience and workflow redesign matter more. Domain-specific assistants outperform generalists in bounded settings.

What could drive it

  • Data limits: the supply of high-quality human-generated training material becomes a binding constraint.
  • Algorithmic returns: additional scale produces less useful capability per pound spent.
  • Compliance friction: deployment becomes slower or narrower in regulated settings. The EU AI Act entered into force on 1 August 2024, with most provisions applying from 2 August 2026 after a staged introduction.

Early indicators

  • The rapid historical growth in training compute slows for 12–18 months.
  • Frontier releases deliver only marginal gains on robust, real-world evaluations.
  • Energy forecasts for data centres are revised down or remain flat.

Playbook

Invest in task-specific models, retrieval, interface and operational quality, governed data and model-agnostic infrastructure. If raw intelligence is becoming a commodity, advantage moves to context and execution.

Scenario 2 — Bubble pops, builders persist

What it looks like

Valuations across the AI stack fall sharply. Start-ups fold, consolidation accelerates and speculative platform bets unwind. Technical progress continues. After two or three years, durable businesses built around measurable returns begin to emerge—much as they did after the dot-com correction.

The three strata

  • Compute and hardware: an oversupply follows the capital-expenditure surge, shifting pricing power towards buyers.
  • Models and software: open models and smaller laboratories gain ground as capital becomes scarcer.
  • Applications: workflow products with clear returns survive; products sustained mainly by narrative do not.

What could drive it

  • A macroeconomic or policy shock.
  • A widening gap between promised and realised returns.
  • GPU supply catching up with demand just as investors become less patient.

Early indicators

  • Valuation multiples contract despite continuing model releases.
  • Hyperscalers trim capital expenditure while maintaining their research cadence.
  • Customers insist on shorter payback periods and stronger evidence before buying.

Playbook

Protect runway and unit economics. Avoid long, inflexible GPU commitments. Prioritise painful, expensive problems where improvements can be measured in time, quality or cash.

Scenario 3 — Bottlenecked boom

What it looks like

Power, chips, cooling, water and regulation throttle the pace of progress. Capability arrives in lumpy leaps when new capacity comes online, followed by pauses while infrastructure catches up.

The three strata

  • Compute and hardware: power connections and semiconductor supply cap scaling; lead times become strategically important.
  • Models and software: sparsity, quantisation and distillation matter as much as scale.
  • Applications: firms automate within bounded environments and defer high-risk or compute-heavy uses.

What could drive it

  • Power constraints: the International Energy Agency projects that global data-centre electricity consumption could reach about 945 TWh by 2030—slightly more than Japan consumes today—with AI as the main driver of growth.
  • Geopolitics: export controls and industrial policy complicate access to advanced chips and models.
  • Regulatory staging: rules come into force at different times and with different interpretations across markets.

Early indicators

  • Grid queues, power-purchase delays or cooling constraints push training runs back.
  • Chip restrictions and planning delays begin to shape model-release schedules.
  • Efficiency claims become as prominent as benchmark gains.

Playbook

Treat energy as a strategic input. Explore power-purchase agreements, siting and efficiency improvements. Diversify compute suppliers, and build compliance into systems from the beginning.

Scenario 4 — Steady compounding automation

What it looks like

There is no artificial-general-intelligence moment. Instead, modest annual gains compound. Agentic systems take on bounded work under human oversight. By the early 2030s, productivity in services and knowledge work is materially higher—even though no single release felt transformative at the time.

The three strata

  • Compute and hardware: inference costs continue to decline, allowing AI to spread to edge devices and back-office systems.
  • Models and software: tool use, memory and orchestration become the main differentiators.
  • Applications: copilots and process automation reshape work one workflow at a time.

What could drive it

  • Adoption beats invention: organisations learn to redesign work around capabilities that already exist.
  • Economics: McKinsey estimated that generative AI could eventually contribute US$2.6–4.4 trillion in value each year across the use cases it studied.
  • Labour exposure: the IMF estimated that AI could affect almost 40% of employment globally, with a larger share in advanced economies.

Early indicators

  • Organisations repeatedly report 10–30% improvements in cycle time or error rates in bounded processes.
  • A growing share of ordinary work is mediated by AI, even when customers never see it.
  • Productivity statistics begin to show steady rather than spectacular gains.

Playbook

Create an automation funnel: identify work, assess its value and risk, test it, then scale what works. Plan workforce transitions before productivity gains arrive. Make evaluation, monitoring and named human accountability part of normal operations.

Scenario 5 — Transformative takeoff

What it looks like

Breakthroughs enable systems that outperform experts across most economically valuable cognitive tasks and can complete long-horizon projects. Timelines for research, coding and design collapse. AI-led discovery and autonomous operations move from demonstrations into production.

The three strata

  • Compute and hardware: frontier laboratories sustain scaling without power becoming an immediate bottleneck.
  • Models and software: algorithmic advances in reasoning, planning and tool use produce a step-change in capability.
  • Applications: AI systems run significant parts of research, product development and operations.

What could drive it

  • Sustained growth in compute combined with algorithmic progress and better synthetic or curated data.
  • Infrastructure programmes large enough to support repeated frontier training runs.
  • Policy frameworks that normalise testing, incident reporting and red-teaming without stopping research altogether.

Early indicators

  • Independent evidence that systems can pursue coherent goals over weeks, not minutes.
  • Novel proofs, designs or discoveries that experts verify but did not originate.
  • Automated laboratories making repeatable scientific progress with limited human intervention.

Playbook

Strengthen safety, security and governance before the capability arrives. Design containment, shutdown and recovery mechanisms. Make consequential actions auditable and subject to approval. Re-underwrite organisational moats: expertise, speed and information advantages may all become less durable.

No-regrets moves across all five futures

Despite the uncertainty, several moves remain useful in almost every scenario:

  • Build model-agnostic infrastructure with evaluation, monitoring and incident response.
  • Govern data, provenance, permissions and intellectual-property rights.
  • Maintain an automation funnel alongside workforce-transition plans and clear accountability.
  • Treat energy, compliance and auditability as design constraints.
  • Preserve supplier diversity so that one model, cloud or chip provider cannot become a single point of failure.

The point of these scenarios is not to pick a single winner. It is to expose which assumptions a strategy depends on—and which moves remain sensible if those assumptions fail.