Charts answer what happened. This page answers why it matters, what might happen next, and what leaders should do about it. Every scenario is built the same way, and every part of it is labeled as what it is: verified information, a published forecast, ACY's interpretation, a conditional path, or an open unknown. No forecast here is presented as guaranteed, and no scenario as fact.
How to read this page: five kinds of content, five labels.
Verified historical information
Facts drawn from the primary source named on the item, checked by ACY before publication.
Published forecasts
Projections published by a named third party, such as the IEA. They carry their publisher and their case, and they are not ACY predictions.
ACY Interpretation
AnaConnectsYou's own analysis: who is affected, what the evidence suggests, and what leadership should consider. Labeled exactly that, never blended into the facts.
Conditional future scenarios
Plausible paths built from the evidence, each with the signals that would be consistent with it. A scenario is a framework for watching, not a prediction, and no path is given a probability.
Unknowns and limitations
What would invalidate a scenario, where the evidence is thin, and what cannot be attributed from the data available. Stated, never smoothed over.
Listening version
A narrated version of this page is reserved as a later phase. The scenario structure above is built so each labeled block can be narrated on its own. Narration is not active yet, and nothing on this page plays audio.
Oil, Energy, and the Two Directions of AI
Which effect dominates, for which industry, in which region, over which horizon: AI improving forecasting, logistics, maintenance, and efficiency inside oil, gas, and power operations, potentially lowering some production costs, or AI infrastructure raising electricity and fuel demand, potentially raising energy system costs? This is a framework for investigation, not a forecast.
What changed
Verified historical information
AI has become two things at once in the energy system: a new driver of electricity demand growth, through data centres, and a working tool inside oil, gas, and grid operations. The IEA's Energy and AI report (April 2025) documents both directions in the same analysis, which is why this scenario is framed as interacting forces, never as a single netted price prediction.
Who is affected
ACY Interpretation
Oil and gas operators; electric utilities and grid operators; data centre developers and the construction and equipment supply chains behind them; cloud providers siting new capacity; and any business signing large power or compute contracts.
The evidence
Verified historical information
Demand side: the IEA projects global data centre electricity demand to more than double by 2030, to around 950 TWh, with AI optimised data centres more than quadrupling, and U.S. data centres on course to account for almost half of U.S. electricity demand growth to 2030 (IEA, Energy and AI, April 2025, Base Case).
Efficiency side, same report: the IEA states the oil and gas industry has been an early adopter of AI for exploration, production, maintenance, and safety, and that AI based fault detection can reduce outage durations by 30 to 50%, with up to 175 GW of transmission capacity unlockable without new lines.
The IEA frames both directions in one sentence: AI is driving a surge in electricity demand from data centres while also unlocking significant opportunities to cut costs, enhance competitiveness, and reduce emissions.
Published forecasts
The demand figures in this scenario are the IEA's published Base Case projections (Energy and AI, April 2025), a third party forecast with its own slower and faster cases. They are not ACY forecasts and not measurements.
Assumptions in this reading
ACY Interpretation
Data centre build out continues near the IEA Base Case rather than its slower or faster cases.
Oil and gas operators keep deploying AI applications that are already commercial today, which is what the IEA's efficiency figures describe.
Generation and grid capacity can be added in the regions where data centres concentrate.
Efficiency gains are not fully consumed by rebound demand, where cheaper capability simply produces more use.
What would invalidate this scenario
Unknowns and limitations
The IEA itself emphasises the significant uncertainties that remain, from the macroeconomic outlook to adoption speed and bottlenecks. An observed change in oil production costs cannot be attributed to AI from price data alone: commodity cycles, OPEC+ decisions, and drilling technology unrelated to AI are competing explanations. On the IEA's own account, the net emissions sign is conditional on how widely AI enabled reductions are adopted.
Indicators to watch
ACY Interpretation
IEA Observatory on Energy, AI and Data Centres updates against the IEA Base Case for 2030 (945 TWh in the April 2025 report).
Actual U.S. electricity demand growth and the data centre share of it, region by region.
Documented operator disclosures of AI driven production or maintenance cost changes, labeled as company reported when they appear.
Grid interconnection queues and outage duration statistics where AI fault detection is deployed.
Whether renewables and natural gas in fact lead in meeting new demand, as the IEA projects.
Decisions leaders should consider
ACY Interpretation
Energy buyers and operators: track which effect reaches your own cost base first. Contract flexibility is worth more than a view on prices formed today.
Businesses siting compute or signing power agreements: region, grid capacity, and interconnection queue position are decision variables now, not abstractions.
Treat any single 2030 demand number as one planning scenario among several, and revisit it against the IEA Observatory updates rather than locking plans to it.
Plausible alternative outcomes
Conditional future scenarios
Each path below is a possibility built from the evidence above, not a forecast. No probabilities are assigned, because the evidence does not support any. Select a path to see what it would look like, the signals consistent with it, and its business implications.
The path · Expansion accelerates
Data centre build out continues at or above the IEA Base Case: more announced capacity, more power procurement, more grid investment, and more construction activity, with AI optimised data centres more than quadrupling their electricity demand by 2030 as the IEA projects.
Signals consistent with this path
Announced data centre capacity and construction starts rising against the IEA Base Case.
Power procurement agreements and grid interconnection requests growing in the regions where data centres concentrate.
Documented investment in generation, transmission, and cooling capacity tied to AI demand.
Business implications
Potential opportunities for power infrastructure, equipment suppliers, construction materials, logistics, and specialist services.
Energy buyers face tighter capacity and stronger competition for interconnection in AI heavy regions.
ACY would track actual project approvals, financing, power availability, construction starts, and procurement awards rather than announcements alone.
Falling Inference Costs and the Price of Business Software
If the cost of a fixed level of AI capability keeps falling at the documented rate, what happens to per seat software pricing, to vendor margins built on scarce intelligence, and to the build versus buy decision for ordinary firms?
What changed
Verified historical information
The cost of a fixed level of AI capability has fallen by orders of magnitude in under two years: querying at GPT-3.5 equivalent accuracy fell from $20 per million tokens in November 2022 to $0.07 by October 2024 (Stanford AI Index 2025), while capability per unit of model size rose with it. At the same time, about one in five organizations already report AI operating costs constraining their use (McKinsey, 2026).
Who is affected
ACY Interpretation
Business software vendors and their per seat pricing models; firms approaching software renewals; internal product and IT teams weighing build against buy; AI platform providers competing on price and capability; and any firm whose margin rests on scarce intelligence.
The evidence
Verified historical information
Stanford's AI Index 2025: the cost of querying a model at GPT-3.5 equivalent accuracy fell from $20 per million tokens in November 2022 to $0.07 by October 2024, a more than 280 fold reduction in about 18 months; depending on the task, inference prices fell 9 to 900 times per year.
Capability per unit of size is rising with it: the smallest model scoring above 60% on MMLU fell from 540 billion parameters in 2022 to 3.8 billion in 2024.
Adoption context: the Stanford AI Index 2026 reports 88% of surveyed organizations using AI in at least one function in 2025, up from 78% in 2024, with generative AI use at 79% (McKinsey survey, self reported).
Anthropic's Economic Index found evidence of weak price sensitivity in business API use: the most used tasks tend to cost more, suggesting buyers pay for capability rather than minimising token cost. That is Anthropic's own platform data.
McKinsey's State of AI Global Survey 2026 (1,719 participants, 97 nations, fielded May to June 2026): 32% of respondents report their organizations decided against buying one or more software products or features because they could be built internally with agentic coding tools, and about 20% report AI operating costs, including token costs, constrained their AI use.
Published forecasts
No third party forecast is relied on in this scenario. The evidence is historical price records and survey results; the open question is how far the documented trend continues, which is why the outcomes below are conditional.
Assumptions in this reading
ACY Interpretation
The price decline continues, but at an unknown, task dependent rate; the documented range is too wide to extrapolate a single path.
Frontier capability keeps commanding a premium even as yesterday's capability becomes a commodity.
Switching costs, data position, compliance, and workflow integration, not token price, remain the practical lock in for business software.
What would invalidate this scenario
Unknowns and limitations
Falling unit prices do not mechanically mean falling total spend: usage volume can rise faster than price falls, which the weak price sensitivity finding is consistent with. Vendors may hold headline prices and compete on agents, data, and distribution instead. Open weight models are a competing explanation for price pressure, independent of inference engineering.
Indicators to watch
ACY Interpretation
The next Stanford AI Index update on inference prices at fixed capability levels.
Whether major software vendors move from per seat pricing to consumption or outcome pricing in their own announcements.
Government reference pricing now visible in GSA OneGov, at 40 to 50% token discounts.
API price pages at each major model launch, read as company reported prices.
Decisions leaders should consider
ACY Interpretation
Software buyers: the build option is now credible for narrow features. Price renewals against the cost of building internally with coding agents before signing.
Vendors and service firms: margin built on scarce intelligence compresses as capability gets cheaper. Differentiation moves to data position, workflow integration, and accountability for outcomes.
Every firm: watch your own AI operating costs quarterly. McKinsey finds cost already constraining use at about one in five organizations, even while most plan to spend more.
Plausible alternative outcomes
Conditional future scenarios
Each path below is a possibility built from the evidence above, not a forecast. No probabilities are assigned, because the evidence does not support any. Select a path to see what it would look like, the signals consistent with it, and its business implications.
The path · Prices fall, spend rises
Unit prices keep falling but usage volume rises faster, so total AI operating cost grows. This path is consistent with Anthropic's own platform finding of weak price sensitivity in business API use: buyers pay for capability, not for the cheapest token.
Signals consistent with this path
Successive Stanford AI Index updates showing continued inference price declines at fixed capability.
Company disclosures of rising AI operating costs alongside expanding use.
McKinsey survey shares reporting cost as a constraint staying near or above the current one in five.
Business implications
Budget discipline, not access, becomes the management problem: quarterly review of AI operating costs against measured output.
Vendors can hold headline prices while competing on agents, data position, and distribution instead of token cost.
Entry Level Knowledge Work and the Senior Pipeline
If AI performs or accelerates the documented tasks junior staff traditionally learned on, how do firms redesign entry level hiring, apprenticeship, and training so that a senior pipeline still exists in five to ten years?
What changed
Verified historical information
The measured employment pressure from AI is concentrated where careers start: workers ages 22 to 25 in highly AI exposed occupations stand about 19% below the kept pace counterfactual, with adjustment running mainly through reduced hiring of young workers (Stanford Digital Economy Lab, revised August 2026). Field studies add the mechanism: an AI assistant lifted novice support agents about 34% while barely moving experts, substituting for part of the experience curve itself.
Who is affected
ACY Interpretation
Employers of entry level knowledge workers in software, support, research, and professional services; training and education providers; junior professionals whose first roles are the most exposed; and senior teams whose future pipeline depends on today's junior hiring.
The evidence
Verified historical information
Stanford Digital Economy Lab, revised August 2026: workers ages 22 to 25 in highly AI exposed occupations stand about 19% below the kept pace counterfactual, with adjustment mainly through reduced hiring of young workers rather than increased separations.
The support agent field study found an AI assistant lifted novice agents about 34% while barely moving experts: evidence that AI can substitute for part of the experience curve, which is precisely the training problem.
The consulting field experiment found AI use reduced correctness on a task outside the capability frontier: the skill juniors most need is verification and frontier recognition, not only production.
The World Economic Forum's Future of Jobs Report 2025 records employer expectations of large scale creation and displacement together, a net positive in aggregate: a redesign expectation, not a disappearance forecast.
Published forecasts
The World Economic Forum figures in this evidence are employer expectations recorded in its Future of Jobs Report 2025: a published survey of what employers expect, not a measurement of what has happened.
Assumptions in this reading
ACY Interpretation
The Stanford pattern is at least partly AI related; the authors describe their patterns as descriptive, not causal estimates.
Firms notice the pipeline problem only with a lag, because senior supply feels adequate while junior hiring is cut.
Training can be redesigned around structured rotation, supervised AI use, and verification drills, rather than restored to its earlier form.
What would invalidate this scenario
Unknowns and limitations
The Stanford authors list their own cautions: the gaps shrink when education is accounted for, some differential trends predate widespread generative AI, and payroll sample estimates exceed national survey benchmarks. The direction of the junior share effect is corroborated across sources; its size and cause are not settled.
Indicators to watch
ACY Interpretation
The Stanford Canaries dashboard's monthly updates: whether the early career gap widens, stabilises, or reverses.
Entry level posting volumes against senior postings in exposed occupations.
Employer announcements redesigning junior roles around supervision and verification of AI output.
Wage data: adjustment so far appears in employment rather than base pay; a change there is a regime signal.
Decisions leaders should consider
ACY Interpretation
Employers cutting junior hiring: decide now how senior capability gets built in five to ten years, or plan to buy it later at market prices.
Redesign junior roles around supervised AI use and verification skill rather than removing the roles outright; the pipeline problem is a design problem.
Training providers: the durable product is judgment formation, verification, and supervised practice, not content delivery.
Plausible alternative outcomes
Conditional future scenarios
Each path below is a possibility built from the evidence above, not a forecast. No probabilities are assigned, because the evidence does not support any. Select a path to see what it would look like, the signals consistent with it, and its business implications.
The path · Roles redesigned around supervision
Employers rebuild junior roles around supervised AI use, verification skill, and structured rotation, treating the pipeline problem as a design problem. The durable training product becomes judgment formation and supervised practice rather than content delivery.
Signals consistent with this path
Employer announcements redesigning junior roles around supervision and verification of AI output.
Training providers launching verification and supervised practice offerings with named employer buyers.
Entry level posting volumes stabilising while role descriptions change.
Business implications
Firms that redesign early keep a senior pipeline while competitors buy seniority later at market prices.
Training providers with proof of judgment formation hold the defensible position; commodity content delivery stays under pressure.
If a scenario here bears on a decision in front of you, ACY works it through with you as a scoped strategic assessment. Request a strategic assessment.
§ACY Vocabulary
The words, in plain English.
Research uses short words for big ideas. Tap any term below for its plain definition. The definition states what the term means, and, where a term is often oversold, what it does not mean.
API
A set of rules that lets one computer program request information from another and receive it in a standard format.
ACY uses APIs to pull data straight from the publisher, so a figure can be checked against its source instead of copied by hand. When a page says a connector is available, it means an API route like this one exists and has been tested.