Effective training compute required for AI systems capable of automating all economic tasks (FLOP, in 2025-equivalent effective FLOP)
- Measures
- The training compute, expressed in effective FLOP normalized to 2025 algorithmic efficiency, of the first AI system (or systems) able to perform essentially all economically valuable tasks currently done by human workers at or above human cost-effectiveness, given adequate deployment compute.
- Population
- Frontier AI training runs
- Conditions
- Effective FLOP = physical FLOP multiplied by the algorithmic-efficiency gain relative to 2025 methods. Excludes physical-world bottlenecks such as robotics deployment and regulation; concerns capability requirement only.
- Unit
- eFLOP
Most likely about 3,000,000,000,000,000,000,000 trillion eFLOP; 90% interval 30,000,000,000,000,000 trillion eFLOP to 10,000,000,000,000,000,000,000,000,000 trillion eFLOP, with a 15% chance that Probability that the quantity is ill-defined: full automation is not reached by scaling a single effective-compute axis (e.g. because data, embodiment, or paradigm changes make 2025-equivalent eFLOP an unworkable conversion).
Stated as a distribution, the precision the evidence supports. Roughly log-uniform to log-normal over twelve orders of magnitude, with a hard floor near 1e27
Assessment
The graph's estimate, all things considered: the figures and arguments it rests on, and how they were combined.
Nobody knows this value to within several orders of magnitude, and every published figure for it comes out of a model rather than a measurement. The one hard constraint is a floor. Frontier training runs through late 2026, at roughly 1e26 to 1e27 FLOP with algorithms at least as good as 2025's, have not produced systems that automate all economic tasks, so the requirement is above about 1e27 effective FLOP.
Above that floor, there are two broad schools. Anchoring models put the median around 1e35 to 1e37 effective FLOP. These include Cotra's biological anchors, Davidson's takeoff model and Epoch's GATE, which uses a range of 1e33 to 1e41 with a median of 1e36.5. All of them assume that going from partial to full automation takes several orders of magnitude more effective compute. The other school points to the rapid growth of agentic task horizons on modest compute increases, and suggests the requirement might be only two to five orders of magnitude above today's runs. The size of that partial-to-full gap is the main empirical crux between the two.
The graph's estimate takes a median near 3e33 effective FLOP, with a 90% interval from about 3e28 to 1e40. This sits below GATE's median, which reflects the strong capability gains of 2024 to 2026. It also reserves about 15% probability that the quantity is ill-posed, because the whole idea rests on capability being a function of a single effective-compute variable. That assumption becomes hard to defend across paradigm shifts and over a dozen orders of magnitude.
The work: which figures, how they were combined, and the arithmetic
Inputs: (1) Floor: the largest runs to 2026 are of order 1e26 to 1e27 FLOP (Grok 4 and GPT-5-class scale), and none automates all tasks, so P(value < 1e27) ≈ 0. (2) Epoch GATE (arXiv 2503.04941, 2025) gives the parameter "AGI training requirements": 1e33 to 1e41 eFLOP, median 1e36.5. This is a judgment-based prior used as model input, not a derived estimate. (3) Cotra (2020) biological anchors: a median of about 1e35 FLOP for transformative AI with 2020 algorithms. Five years of algorithmic progress at about 0.5 OOM per year would convert that to roughly 1e32 to 1e33 in 2025-equivalent terms, although "transformative" is a weaker bar than "all tasks". (4) Shorter-timeline views (time-horizon extrapolation, forecasts of automated coders at about 1e28 to 1e30) apply to software/cognitive work, not physical tasks, so they bear on the lower tail. Combination: the median is placed in log space between the converted anchors (≈33) and GATE (36.5), shaded down for the recent evidence on capabilities, giving ≈33.5 (3e33). The tails are wide because the input priors span 8 or more OOM and are weakly calibrated. 5th percentile at 1e28.5 (a narrow gap plus fast progress). 95th percentile at 1e40 (the GATE upper range). 15% is put outside every value for ill-definition. The GATE paper was opened for its framing; the parameter appendix was not read in full this pass, so the normalization year of GATE's eFLOP is taken as stated in the instance context.
Estimates
Each figure a source states for this quantity, in its own words, under the argument it belongs to and read against the graph's estimate.
Grey band: the graph's 90% interval, with its central value as a line. Dark marks: each source's own figure. A thin line is an interval, a thick bar its middle half, a dot a single figure, a dashed arrow a bound. An open box is where an argument places the value.
Decomposition
The arguments the estimate rests on, each stated as it runs, with what it shows about the value and its premises linked inline. ↗︎ opens a premise; the map shows how they fit together.
Because frontier systems trained at 10^26 to 10^27 FLOP, with 2025-era or better algorithms, fall short of automating all economic tasks, the effective compute needed for full automation must exceed the effective compute of those runs, roughly 1e27 eFLOP in 2025 terms.
- Does it follow?
- certain if the premises holdIf the premises are true, the conclusion must be. Whether they are is shown beside this.constitution →
- Do the premises hold?
- 1 not separately assessed
Granting that current frontier systems fall short of full automation and that their effective compute in 2025 terms is about 1e27, the requirement must exceed that figure. The argument rests on As of 2025–2026, frontier AI systems trained at 10^26–10^27 FLOPs fall short of automating all economic tasks., which nobody seriously disputes. It fixes only a floor and says nothing about how far above the floor the value lies.
Taking capability as a function of a single effective-compute variable, biological anchors (Cotra's estimate of about 1e35 FLOP for transformative AI with 2020 methods) and Davidson-style models place partial automation near current scales, and because going from 20% to 100% automation requires several orders of magnitude more effective compute, full automation lands near 1e35 to 1e37 eFLOP, with very wide uncertainty (Epoch's GATE uses 1e33 to 1e41, median 1e36.5).
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 2 not separately assessed
Even granting its premises, the argument yields only a wide range, because the anchors are themselves judgment-laden priors and the conversion between biological and machine compute is uncertain by many orders of magnitude. Its central placement depends on Increasing AI automation from 20% to 100% of economic tasks requires several orders of magnitude more effective compute., which is a modelling assumption with little calibration. The whole argument also presupposes AI capabilities can be adequately modeled as a function of a single effective-compute variable combining physical compute and algorithmic efficiency..
Because the horizon of tasks AI agents can complete keeps doubling about every seven months while effective compute grows only a few orders of magnitude over the same period, and because the compute gap from partial to full automation is on this view narrower than anchoring models assume, systems able to do month- or year-long work across occupations may arrive at only 1e29 to 1e32 eFLOP.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 doubtful1 not separately assessed
Granting that task horizons keep doubling, the argument supports a lower requirement for software and cognitive work. But horizon length on benchmark tasks is a reading, not a measure of automating all occupations, and physical tasks lie outside it. It also needs Increasing AI automation from 20% to 100% of economic tasks requires several orders of magnitude more effective compute. to be false, which is unsettled. The argument mainly thickens the lower tail.
Provenance
Where this quantity's evidence comes from. The figures sources state are under Estimates above.
1e33 to 1e41 (1e36.5)
AGI training requirements T (eFLOP): Effective FLOP required to train AI systems capable of automating all tasks.
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
Created by extractor · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.