2×10²⁸ FLOP training run completed
Tracker summary: Agent-2 trained at 2×10²⁸ FLOP (Apr 2026 to Mar 2027), a 1000× increase over the scenario’s GPT-4 baseline (tracker arithmetic).
At a glance
- Assessment: Not Yet Testable
- Confidence in assessment: 50%
- Predicted timing: March 2027
- Primary source: AI 2027 Compute Forecast supplement (Agent-2 row; not reproduced in the archived PDF); related archived context: Late 2025: The World’s Most Expensive AI, p. 4
What AI 2027 Predicted
Tracker summary: The scenario’s Compute Forecast table shows Agent-2 trained at 2×10²⁸ FLOP, with training running from April 2026 through March 2027. GPT-4 was trained at 2×10²⁵ FLOP, making Agent-2 a 1000× increase over that baseline by tracker arithmetic (2×10²⁸ ÷ 2×10²⁵), not a quoted multiplier. Note: the “Late 2025” section describes datacenter capacity being built, not a completed 10²⁸ training run. The modeled 2×10²⁸ FLOP run completes in March 2027. Archived PDF, p. 3.
Scenario source: Compute Forecast, §5, Agent-2 row; archived version and page-number reference: Compute Forecast supplement (Agent-2 row; not reproduced in the archived PDF); related context: Late 2025: The World’s Most Expensive AI, p. 4. Archived PDF, p. 3.
How We Track This
The threshold is 2×10²⁸ FLOP used in a single completed training run, from the Compute Forecast supplement’s Agent-2 row, with training modeled from April 2026 through March 2027. Confirm only with a comparable estimate or disclosed training account and a documented completion date. The scenario is the target, not evidence that any real run occurred.
We monitor Epoch’s estimates, lab training reports and credible reconstructions. Hardware capacity and shipments provide context but cannot establish consumed training FLOP.
Current Evidence
The reviewed evidence does not establish the original 2×10²⁸ FLOP training run. Training-compute estimates are uncertain; capacity announcements and hardware shipments do not establish compute used in a completed run. The comparison window is April 2026-March 2027. A generic 70%-speed estimate cannot infer a different completion date.
As retrieved on November 7, 2026, the largest nonempty training-compute estimate in Epoch’s model dataset is 9.4×10²⁶ FLOP for GPT-6 Astra, with a publication date of September 3, 2026. That is 4.7% of 2×10²⁸ FLOP, or about 21.3 times below the target (tracker arithmetic). A publication date is not a verified training-completion date, and an estimated dataset maximum does not rule out undisclosed larger runs. The Agent-0 training reaches 10²⁷ FLOP separately compares characterized 2025 runs with its earlier target.
Sources:
- Frontier open models may surpass 1e26 FLOP: Epoch AI
- Tracking Large-Scale AI Models: Epoch AI
- Grading AI 2027’s 2025 Predictions: AI Futures Project
Counterevidence & Limitations
- Labs are increasingly opaque about training compute, making verification difficult
- Algorithmic improvements may reduce the need for raw compute scaling
- Some labs may have completed large runs without publicly disclosing details
- Algorithmic improvements may change what a training budget achieves, but this dossier still tracks the original FLOP quantity.
What Would Change Our Assessment
- Strengthen the assessment: Credible technical evidence describes a run approaching 2×10²⁸ FLOP.
- Confirm: Evidence establishes completed training compute and date under a consistent definition.
- Reassess timing after March 2027: Infrastructure delays matter insofar as they affect the original window.
Update History
| Date | Update |
|---|---|
| 2026-11-07 | Assessment reviewed. Status: Not Yet Testable; confidence: 50%. The Epoch dataset retrieved November 7 estimates its largest listed run, GPT-6 Astra, at 9.4×10²⁶ FLOP, 4.7% of the 2×10²⁸ FLOP target. Its notes give an approximate 5×10²⁶ to 2×10²⁷ interval. Hardware-based estimates are not audited training disclosure or an exhaustive census. The modeled run ends in March 2027; unknown completion dates, utilization and undisclosed runs prevent current resolution. |
| 2026-11-07 | Correction: Aligned the target to the Compute Forecast’s single completed 2×10²⁸ FLOP Agent-2 run over April 2026 through March 2027. Against the scenario’s 2×10²⁵ FLOP GPT-4 baseline, the ratio is 1,000x, correcting 100x. The March 13 “6-12 months too aggressive” inference confused late-2025 capacity with a completed run and used the wrong deadline. The December 2025 upper bound was not established by characterized runs; Epoch estimates Grok 4 at roughly 5×10²⁶ FLOP. Added the separately dated November 7 dataset maximum, GPT-6 Astra at 9.4×10²⁶ FLOP (4.7% of target, about 21.3 times below), and removed the GPU-hosting listicle. Dataset estimates do not exclude undisclosed runs. |
| 2026-09-06 | Current evidence and assessment explanation reviewed. Clarified measurement scope, source interpretation and limitations. |
| 2026-03-13 | 10²⁸ FLOP run now expected mid-to-late 2026 based on infrastructure buildout timelines. Prediction was ~6-12 months too aggressive. |
| 2025-12 | No confirmed 10²⁸ FLOP training run. Frontier estimated at ~10²⁶·⁵–10²⁷, well below the target. |