No one can say when energy-hungry AI tasks like video and nonstop agents will outnumber the light, everyday ones.
When will energy-intensive AI tasks become the majority of AI workloads?
This explores when energy-hungry AI work, like video generation, long reasoning chains and always-on agents, might overtake lighter tasks as the bulk of what AI systems do, and what the corpus can and can't say about that timing.
This explores when energy-heavy AI tasks might become most of what AI systems run. The short answer: the corpus has no date for that tipping point. What it does have explains why no one can easily give one. The clearest framing comes from the IEA, which treats AI electricity demand as the net result of three trends that move independently: the energy cost of each task falling by orders of magnitude every year, adoption surging, and a steady shift toward heavier tasks such as video and reasoning How do efficiency gains and new AI capabilities balance each other?. Because these trends pull against each other, a forecast of heavy tasks taking over the 'majority' depends on which one wins in a given year. Efficiency gains alone won't tell you.
It also helps to know what the baseline actually is, because it isn't chatbots. In one well-documented production setting, personalized recommendation models (the systems that choose what's in your feed) used 79% of Facebook's AI inference cycles, and just three model types took 65% What dominates AI compute in production systems today?. So in at least one major production system, the 'majority of AI workloads' is quiet, unglamorous ranking work. The question is really about whether generative and reasoning workloads can outgrow that installed base.
The strongest argument that they will comes from agents. One model of the AI compute stack argues that inference grows along a dimension training never had: every autonomous agent is a continuous consumer of compute, not a one-off query Will inference compute soon exceed training compute demand?. Reasoning makes this worse. Reasoning models stay ahead of non-reasoning ones no matter how much extra compute the non-reasoning models get, because their training makes long chains of tokens productive Can non-reasoning models catch up with more compute?. That gives labs a real reason to keep spending more tokens per task. On this view, the crossover is tied to how fast agents and reasoning get deployed, more than to any calendar date.
The counter-pressure is just as concrete. Compact models trained on real multi-step work can reach the cost-performance sweet spot of much larger ones Does model efficiency matter more than peak capability for real work?. Breaking a task into tiny pieces with voting at each step lets small non-reasoning models finish million-step tasks without errors Can extreme task decomposition enable reliable execution at million-step scale?. 'Energy-intensive capability' doesn't have to mean energy-intensive execution. The same agentic work could be routed through cheap models.
Here's what you may not have expected: energy may not be the constraint that decides this. Epoch AI estimates that a 1-gigawatt data center costs about $8.5B a year to run, with servers at roughly $5B and energy at only about $0.6B What actually drives AI data center costs?. Inside the chips, high-bandwidth memory rose from 52% to 63% of component spending in under two years Is memory becoming the dominant cost in AI chips?. At AI labs, compute takes 54–62% of operating expenses Do compute costs really dominate AI company budgets?. So the pace at which heavy workloads take over may depend more on hardware and memory supply than on electricity. If you want to explore further, the open question is less 'when' and more 'what will run out first: chips, memory or power?'
Sources 9 notes
The IEA shows AI electricity demand results from three competing forces: per-task efficiency dropping by orders of magnitude annually, surging AI adoption, and shift toward energy-intensive tasks like video and reasoning. These trends move independently, so efficiency alone cannot predict total demand.
DNN-based personalized recommendation comprises 79% of Facebook's inference cycles, with just three model classes consuming 65% of total cycles. This reflects production infrastructure shaped by embedding-table lookups and sparse feature handling, not transformer or convnet architectures.
A seven-layer AI compute stack model reveals that inference demand grows along a new axis training never had: every autonomous agent is a continuous inference consumer. This makes inference, not training, the likely dominant compute demand as agents and robots scale.
Reasoning models persistently outperform non-reasoning models regardless of inference budget because training instills a reasoning protocol that makes additional tokens productive. The gap is fundamentally about deployment mechanisms and training structure, not raw capability.
Occamy-1.0, a 35B-parameter model further trained on execution-grounded data and long-horizon trajectories, achieves competitive performance with much larger models while sitting at the low-cost knee of the Pareto frontier, suggesting that training for coordination and follow-through substitutes for raw scale in multi-step work.
Show all 9 sources
MAKER solves million-step tasks with zero errors by decomposing into minimal subtasks, applying voting at each step, and flagging correlated errors. Surprisingly, small non-reasoning models suffice when decomposition is extreme enough, inverting the standard approach to hard problems.
Epoch AI's model finds a 1GW data center costs $8.5B annualized, with servers at 60% ($5B/year) versus energy at just $0.6B/year. Capital recovery of hardware dominates operating expenses.
Epoch AI's component cost model shows HBM's share of total AI chip spending rose 11 percentage points between Q1 2024 and Q4 2025, driven by price increases and supply tightness. HBM spend nearly tripled in dollar terms while logic held flat and packaging and auxiliary components declined.
Epoch AI's analysis of Anthropic, Minimax, and Z.ai shows R&D and inference compute account for 54–62% of total operating expenses, while staffing remains below 25% despite high salaries.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Price of Progress: Price Performance and the Future of AI
- The rising costs of training frontier AI models
- The plunging price of thought
- AI Compute Architecture and Evolution Trends
- Key Questions on Energy and AI
- Is a compute crunch coming?
- Reasoning Models Can Be Effective Without Thinking
- The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs