Tracker

The IEA estimates global data centre consumption at 415 TWh in 2024.

• Not an AI-only measurement.

The IEA update estimates global data centre consumption at 485 TWh in 2025.

• Not an AI-only measurement.

The global IEA estimate increased by 16.9% from 2024 to 2025.

• A comparison across editions of IEA estimates, not meter readings.

The IEA's updated central pathway has approximately 950 TWh of data centre consumption in 2030.

• Scenario, not observation.

The IEA's updated central pathway reaches 1.96 times the 2025 level in 2030, equivalent to approximately 14.4% compound annual growth.

• A mechanical description of the IEA scenario, not an independent forecast.

The IEA's central 2030 estimate moved from approximately 945 TWh in the 2025 report to approximately 950 TWh in the 2026 update.

• Both source figures are rounded, and methods may differ between editions.

The IEA assesses that grid risks could delay approximately one fifth of planned global data centre capacity towards 2030.

• Model-based risk assessment; not a project registry.

The module does not yet have a source-compatible regional series for grid queues, available power or connection status.

• A measurement gap does not mean regional data does not exist; it has not yet been normalised in this module.

Energy has two compatible global annual observations for total data centre electricity, 2024 and 2025. They are comparable, but two points do not establish a robust long-term trend, and the 2030 points remain source scenarios.

• The series covers all data centre electricity and is not an AI-only series.
• Regional grid queues have not yet been measured under the same contract.

Theoretical fleet efficiency increased from 791.7 to 1,566.5 H100e per nominal MW from 2022-Q1 to 2025-Q4, approximately 1.98 times.

• H100e is a theoretical dense 8-bit comparison model.
• Nominal board power is not actual electricity consumption, and the result is not useful work per joule.

From 2022-Q1 to 2025-Q4, estimated global compute holdings increased approximately 333.8 times, nominal accelerator power 168.7 times and theoretical compute per nominal MW 1.98 times.

• The scaling index describes theoretical installed/estimated accelerator capacity, not actual AI work.

The world series' nominal accelerator power in 2025-Q4 is equivalent to 117.0 TWh per year if nameplate power is sustained continuously at PUE 1.0.

• This is an upper nameplate sensitivity, not measured energy use.
• It excludes CPU, memory, networking and cooling, and does not automatically account for PUE above 1.0.

The nameplate envelope of 117.0 TWh/year is 24.1 percent of the IEA's estimate of 485 TWh for all data centre electricity in 2025.

• The denominator includes all data centres; the numerator includes only nominal AI accelerator power.
• The ratio is an order-of-magnitude sensitivity, not an observed share.

Compute per nominal watt is a theoretical hardware proxy. Useful AI work per kWh requires workload, precision, system and wall-power measurements that this module does not yet contain.

• Do not use the curve as a benchmark for model quality or actual energy efficiency.

RExBench v3 reports that the best result, even with human-written hints, is below 44% on realistic research extensions.

• Depends on the setup and hints.

RExBench contains 12 research implementation tasks with automated evaluation.

• Not the full research lifecycle.

MLR-Bench describes frequent fabricated or invalidated experimental results, for example 80%, for the tested coding-agent setup.

• Not a universal error rate.

MLR-Bench covers 201 tasks and a four-stage workflow from ideation to paper writing.

• The stages use different models and evaluation methods.

In MLRC-Bench, the best tested agent closed 9.3% of the gap between the baseline and top human performance.

• A specific agent and seven competition tasks.

MLRC-Bench uses seven dynamic ML competition tasks with objective performance measures.

• A narrow ML domain and a small suite.

Together, the three benchmarks point to implementation and experimental validity as a clear limitation, but they cannot be combined into one automation share.

• Qualitative triangulation, not a meta-analysis or common score.

The module publishes no overall R&D automation percentage because the benchmarks do not measure the same task, agent or success criterion.

• A measurement gap, not evidence of zero automation.

The current benchmarks do not establish a complete, independent validation and reproduction loop after experimentation and interpretation.

• A coverage gap in the selected benchmarks, not evidence that such validation never takes place.

The hypothesis stage has partial benchmark coverage: MLR-Bench includes proposals in its workflow, while RExBench defines twelve research extensions as concrete tasks.

• The sources use different tasks and evaluators and do not establish a common success rate.

MLR-Bench includes experimentation in its workflow, but also reports seriously invalidated results for one tested agent setup.

• Coverage of the experimental stage is not the same as reliable experimental control.

Research Automation has three sourced benchmark families, but zero longitudinal points within the same benchmark version. RExBench, MLR-Bench and MLRC-Bench have different tasks, evaluators and denominators and therefore establish no common trend.

• The benchmark families cannot be combined into one automation share.
• Maturity describes evidence coverage, not general research autonomy.

METR TH1.1 reports 320 minutes [170–729] for Claude Opus 4.5 at a modelled 50% success rate.

• A specific model, suite and agent setup.

METR's February–March 2026 assessment places the public frontier at approximately 12 hours [5–61] at 50% success.

• A cohort estimate with a wide interval and a suite nearing saturation.

The same METR assessment places the public frontier at approximately 1.5 hours [50 minutes–2 hours 40 minutes] at 80% success.

• A cohort estimate, not a ranking of named models.

The public frontier point estimate is eight times longer at 50% than at 80% success, showing strong sensitivity to the reliability requirement.

• The ratio uses rounded point estimates; intervals are wide.

TH1.1 has 228 tasks, of which 31 are estimated to take at least eight hours and five of these have a human baseline.

• Not a representative distribution of labour-market tasks.

Only 13.6% of TH1.1 tasks take eight hours or longer, and 2.2% of the entire suite consists of such tasks with a human baseline.

• Task counts say nothing about representativeness or quality.

METR marks estimates above 16 hours as unreliable with the current task suite.

• Not a statement about actual maximum autonomy.

Agent Horizon shows 4 compatible P50/P80 estimates across 2 public model cohorts under the same TH1.1 contract. They are concurrent model and reliability observations, not a time series; suite coverage and the 16-hour boundary limit generalisability.

• P50 and P80 are reliability levels, not separate points in time.
• The suite is limited to software, ML and cybersecurity tasks.

METR TH1.1 reports 341.735276 expert-minutes [186.581591–768.779526] for GPT-5.4 at 50% success.

• Model-, suite- and setup-specific estimate.

METR TH1.1 reports 53.877851 expert-minutes [23.957027–108.679232] for GPT-5.4 at 80% success.

• Model-, suite- and setup-specific estimate.

SEMI forecast 2.2 million 200 mm-equivalent wafers per month for nodes at 7 nm and below in 2025.

• A forecast, on a 200 mm-equivalent basis, and not AI-only.

The SEMI series shows 850,000 advanced 300 mm wafers per month in 2024 and a forecast pathway through 982,000 in 2025 and 1.16 million in 2026 to 1.4 million in 2028.

• A forecast of front-end capacity; it does not specify yield, chip mix or deliveries.

SEMI states that capacity for 2 nm and below rises from below 200,000 wafers per month in 2025 to above 500,000 in 2028.

• Source-stated bounds, not point estimates; a forecast, not deliveries.

TSMC states that annual capacity at its managed fabs exceeded 17 million 12-inch wafer equivalents in 2025.

• Not advanced-node, AI or packaging capacity.

SIA/BCG estimates that the US share of global fab capacity rises from 10% to 14% by 2032.

• A scenario with policy and investment assumptions; not AI output.

No public, method-compatible series for HBM and advanced packaging capacity has yet been accepted into the module.

• A measurement gap does not mean capacity is absent or that the stage is necessarily the bottleneck.

The module's 200 mm equivalents, 300 mm wafers, 12-inch annual capacity and geographical shares are kept as separate measures.

• No hidden area, yield or node conversion is used.

Public sources show strong expansion of advanced front-end capacity and leading nodes, while the module cannot yet rank HBM or advanced packaging on the same measurement basis.

• This is an evidence map, not a complete supply model or bottleneck ranking.

The module does not yet have a compatible public series for the throughput of lithography and other critical equipment.

• The coverage gap does not mean that equipment is the binding bottleneck.

Chip Supply has one compatible observed point in the advanced 300 mm series. Three other points are source forecasts and can be read as a SEMI source pathway, not as observed deliveries; HBM, packaging and equipment remain separate measurement gaps.

• Three forecast points are shown separately from the single observed point.
• Different wafer, HBM, packaging and equipment measures are explicitly kept separate.