OBSERVATION
The September Compute update adds and revises source-reported data-centre plans. The tracker now shows eight dated frontier-facility plan points, up from seven. Its global ownership series is unchanged and still ends in December 2025.
PUBLISHED AND REVIEWED
More planned compute, not a new measurement of growth
CALIBRATION / NOT A LIVE FEED
- REVIEW STATUS
- reviewed
- PUBLISHED REVIEW DATE
- PREVIOUS REVISION
- briefing/today@1
INTERPRETATION
The accepted Epoch AI archive adds Fairwater Wisconsin and Stargate Shackelford plans and updates the current facility selection. Sines remains a documented minimum of 31,834 H100e at its latest register date, 16 August 2026. H100e compares theoretical accelerator capacity; it does not measure useful work. Plans, ownership estimates and location-based minimums remain separate.
FORECAST IMPLICATION
The overall assessment remains provisional. This briefing updates the accepted Compute evidence; it does not renew every module's evidence or assign a probability to a scenario.
WHY IT MATTERS
The update improves our view of possible future infrastructure. It does not show that the planned capacity has been built, powered or made useful for AI research. It gives no new basis to call rapid takeoff confirmed.
UNCERTAINTY
Facility capacity and opening dates are estimates. The Sines 2027 plan carries a modelled 360,451–811,016 H100e range and a ±6-month date band.
COUNTEREVIDENCE
The update supplies no new measurement of agent reliability, autonomous research or regional grid availability. Those limitations have not been resolved by a larger plan register.
WATCH NEXT
Look for evidence that planned facilities become operational, with reported capacity and power access. Separately, watch for reliable completion of long tasks and independently validated end-to-end AI research. New source checks must not be confused with new observations.
CHANGE SINCE LAST TIME
The September Compute update adds and revises source-reported data-centre plans. The tracker now shows eight dated frontier-facility plan points, up from seven. Its global ownership series is unchanged and still ends in December 2025.
Scenario and milestone history
Scenarios
- What failure looks likeNot assessed
Useful AI systems optimize what is easy to measure while human control becomes less effective. This pathway can end in gradual disempowerment or a cascade of failures; it does not require a single sudden breakthrough.
- AI 2040: Plan ANot assessed
An international agreement changes the incentives for developing powerful AI, combining research transparency, verification and controlled scaling. This is a proposed route to a flourishing future, not a prediction or Vardark policy endorsement.
- Production web: economic exclusionNot assessed
In this scenario, competing AI-run firms and governments expand production while people lose income, access to essentials and practical influence. It does not require rogue AI or recursive self-improvement.
- AI 2027Watching closely
Agent progress and compute scaling fit parts of the pathway. Robust automation of frontier AI research has not been publicly confirmed.
- MannaNot assessed
The Observatory does not yet have a direct employment and institutions module capable of assessing this pathway.
- MannaNot assessed
The current direct trackers do not measure broad distribution, material abundance or personal agency well enough to support an assessment.
- PaperclipNot assessed
The Observatory does not yet have a direct alignment and control module. An absent assessment is not evidence for or against the pathway.
- Slow or uneven developmentReference, not a score
The reference pathway is retained to counter the automatic interpretation of every breakthrough as takeoff. It receives no overall probability score.
Milestones
- Agents solve longer, economically relevant tasksPARTIAL SIGNALS
Measured progress is clear in software tasks, but generalisation and reliability beyond the current test range are uncertain.
- AI performs a substantial part of frontier AI researchWATCHING
The decisive distinction is independent research from hypothesis to validated result, rather than assistance with coding or literature searches alone.
- AI 2027: superhuman coderNOT CONFIRMED
The scenario describes a system that can do the work of the best developers in AI research faster and cheaply enough for many copies to run.
- Autonomous optimisation with sustained resource controlNOT ASSESSED
The Observatory does not yet have a direct alignment and control module capable of assessing this milestone.