Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 41 of the 42 references it cites were found in a registry or an index; none of the 36 sources its claims cite is checked for support yet.

AI and Compute: How Much Longer Can Computing Power Drive Artificial Intelligence Progress?

Open access
Lohn, Andrew, Musser, Micah · 2022 · Center for Security and Emerging Technology

Abstract

Between 2012 and 2018, the amount of computing power used by record-breaking artificial intelligence models doubled every 3.4 months. Even with money pouring into the AI field, this trendline is unsustainable. Because of cost, hardware availability and engineering difficulties, the next decade of AI can't rely exclusively on applying more and more computing power to drive further progress.

Overview

CoChat Summary

## TL;DR

The paper argues that the decade-long AI trend of rapidly increasing compute use is likely nearing its limit because training costs, chip availability, and parallelization bottlenecks are all constraining further growth. It concludes that future AI progress will depend more on algorithmic and hardware efficiency, talent, and application-specific approaches than on simply scaling compute.

## Key findings

  • The paper states that compute demands for top AI models doubled about every 3.4 months between 2012 and 2018, far faster than the prior Moore’s-law-like pace.
  • It estimates that training GPT-3 required about 3,600 petaFLOPS-days, costing about $1.65 million on Google TPUs at maximum continuous performance; its compute trendline says a comparable leading model by late 2021 would cost about $9.4 million, and by the end of 2021 the trendline predicted a model of just over one million petaFLOPS-days costing over $450 million.
  • The paper says the price of compute has not decreased since 2017 for gigaFLOPS, and cloud GPU prices at AWS and Google Cloud have remained constant since at least 2017 and 2019, respectively; Nvidia’s newer A100 offered only 5% more FLOPS per dollar than the 2017 V100.
  • Based on accelerator shipments, it estimates about 35 million accelerators reach datacenters annually, and says that by December 2025 the trendline would require a single model to use every GPU in every datacenter for three years; even if the estimate were off by an order of magnitude, the breaking point would still arrive by December 2026.

## Methods

  • The paper is a trend-analysis / projection paper that extrapolates historical AI compute growth, then tests the sustainability of that trend against training cost, accelerator supply, and parallelization limits.
  • It uses cloud compute pricing as a baseline, focusing on training only and explicitly excluding data generation/preparation costs; it also compares Google TPU pricing with state-of-the-art GPU pricing.
  • For availability, it estimates annual accelerator supply to datacenters from public shipment and market-share figures; for large-model scaling, it uses published scaling laws and reported training setups to estimate compute needs for models like GPT-3 and hypothetical larger models.

## Limitations

  • The paper states that its supply estimate for accelerators is imprecise and likely a substantial overcount, because it counts accelerators across cloud datacenters without separating inference chips from training chips.
  • It also states that its cost estimates are conservatively low because they assume 100% accelerator utilization and peak advertised performance, while real-world utilization is typically lower; this is an explicit caveat in the paper.
  • Evident from the text: the analysis focuses on training compute only, so it does not capture the full compute burden of AI systems, especially inference, which the paper itself notes can be substantial.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.51593/2021ca009

Cites

41 of 41 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 16 Not valid · 36 Not checked · 88 Own results and interpretations

47 more claims

Create a free account to see all 140 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 58
  • It may state two separate claims as one. 45

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 42 references

Cited by

See all 39

Metadata

Authors
Lohn, A., & Musser, M.
Venue
Center for Security and Emerging Technology
Published
2022-01
OpenAlex
W4205930673
Cited by
39
References
42
Citations by year
2026: 3 · 2025: 4 · 2024: 14 · 2023: 16 · 2022: 2
Type
report
Language
en
Open access
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence Applications
  • Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
Keywords
computational power, Moore's law, scalability limits, AI research funding

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: AI and Compute: How Much Longer Can Computing Power Drive Artificial Intelligence Progress? https://app.cochat.ai/papers/10.51593/2021ca009
The paper
Lohn, A., & Musser, M. (2022). AI and Compute: How Much Longer Can Computing Power Drive Artificial Intelligence Progress? Center for Security and Emerging Technology. https://doi.org/10.51593/2021ca009

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.