AI and Compute: How Much Longer Can Computing Power Drive Artificial Intelligence Progress?
Open accessThe DOI matches its Crossref record; 42 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.
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
Cites
Statistics
Verification
Claims the paper makes
0 Valid · 16 Not valid · 36 Not checked · 88 Own results and interpretations
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Nvidia has used techniques to further reduce precision that have allowed their newest processors to train two to three times faster than in 2017.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“83,000 GPUs represents only 0.2 percent of the 35 million accelerators we estimate go into the cloud every year, and $2 billion is a very high sticker price, though well within the budgetary capacity of a nation-state.”
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“During this period, manufacturers have improved performance by developing chips that can perform less precise computations rather than by simply performing more of them.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
In the paper
“An estimate of about $4.6 million is probably more accurate.”
What the sources say
- Not matchedcites Chuan Li
This doesn’t read as a citation we can look up.
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Not matched
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For now, assuming that developers can achieve near optimal efficiency, the equation estimates that building GPT-4 -which we define as one hundred times bigger than GPT-3 (17.5 trillion parameters) -would take at least 450,000 petaFLOPS-days.”
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
47 more claims
Create a free account to see all 140 claims and every source verdict.
- Probably not a substantial claim. 58
- It may state two separate claims as one. 45
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.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
“As one example, the 530 billion parameter Megatron-Turing model used 4,480 GPUs in total. Eight different copies of the model ran simultaneously on different slices of the data, but each copy of the model was so big that it was stored across 280 GPUs.”
Reported 4480 · recomputed 2240
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Managing Massive Models
“As one example, the 530 billion parameter Megatron-Turing model used 4,480 GPUs in total.”
§ Managing Massive Models
“Eight different copies of the model ran simultaneously on different slices of the data, but each copy of the model was so big that it was stored across 280 GPUs.”
Citations
Cites
ImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
Highly accurate protein structure prediction with AlphaFold
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., et al. (2021) · Nature · doi:10.1038/s41586-021-03819-2 · Cited by 43,764
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
RoBERTa: A Robustly Optimized BERT Pretraining Approach
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019) · arXiv · doi:10.48550/arxiv.1907.11692 · Cited by 16,515
Kaugeranna, R., Kaugeranna, E., & Aion, (. S. 4. 6. (2023) · DROPS (Schloss Dagstuhl – Leibniz Center for Informatics) · doi:10.4230/lipics.giscience.2023.43 · Cited by 13,690
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929
Scaling Laws for Neural Language Models
Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., & Amodei, D. (2020) · arXiv · doi:10.48550/arxiv.2001.08361 · Cited by 9,550
Playing Atari with Deep Reinforcement Learning
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., & Riedmiller, M. (2013) · arXiv · doi:10.48550/arxiv.1312.5602 · Cited by 9,471
A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play
Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D., Graepel, T., et al. (2018) · Science · doi:10.1126/science.aar6404 · Cited by 3,562
On the Opportunities and Risks of Foundation Models
Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. (2021) · arXiv · doi:10.48550/arxiv.2108.07258 · Cited by 2,676
Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D., Graepel, T., et al. (2017) · arXiv · doi:10.48550/arxiv.1712.01815 · Cited by 1,102
Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity
Fedus, W., Zoph, B., & Shazeer, N. (2021) · arXiv · doi:10.48550/arxiv.2101.03961 · Cited by 1,005
Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer
Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., & Dean, J. (2017) · arXiv · doi:10.48550/arxiv.1701.06538 · Cited by 933
Efficient large-scale language model training on GPU clusters using megatron-LM
Narayanan, D., Shoeybi, M., Casper, J., LeGresley, P., Patwary, M., Korthikanti, V., Vainbrand, D., Kashinkunti, P., Bernauer, J., Catanzaro, B., et al. (2021) · Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis · doi:10.1145/3458817.3476209 · Cited by 639
Torres H., J. P. (2025) · Zenodo (CERN European Organization for Nuclear Research) · doi:10.5281/zenodo.17754506 · Cited by 580
The Computational Limits of Deep Learning
Thompson, N. C., Greenewald, K., Lee, K., & Manso, G. F. (2020) · arXiv · doi:10.48550/arxiv.2007.05558 · Cited by 538
Pham, H., Dai, Z., Xie, Q., & Le, Q. V. (2021) · 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr46437.2021.01139 · Cited by 509
The Computational Limits of Deep Learning
Thompson, N., Greenewald, K., Lee, K., & Manso, G. F. (2023) · Ninth Computing within Limits 2023 · doi:10.21428/bf6fb269.1f033948 · Cited by 495
Carbon Emissions and Large Neural Network Training
Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L.-M., Rothchild, D., So, D., Texier, M., & Dean, J. (2021) · arXiv · doi:10.48550/arxiv.2104.10350 · Cited by 486
Improving language models by retrieving from trillions of tokens
Borgeaud, S., Mensch, A., Hoffmann, J., Cai, T., Rutherford, E., Millican, K., Driessche, G. V. D., Lespiau, J.-B., Damoc, B., Clark, A., et al. (2021) · arXiv · doi:10.48550/arxiv.2112.04426 · Cited by 382
Scaling Language Models: Methods, Analysis & Insights from Training Gopher
Rae, J. W., Borgeaud, S., Cai, T., Millican, K., Hoffmann, J., Song, F., Aslanides, J., Henderson, S., Ring, R., Young, S., et al. (2021) · arXiv · doi:10.48550/arxiv.2112.11446 · Cited by 235
Benchmarking TinyML Systems: Challenges and Direction
Banbury, C. R., Reddi, V. J., Lam, M., Fu, W., Fazel, A., Holleman, J., Huang, X., Hurtado, R., Kanter, D., Lokhmotov, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2003.04821 · Cited by 216
Zeng, W., Ren, X., Su, T., Wang, H., Liao, Y., Wang, Z., Jiang, X., Yang, Z., Wang, K., Zhang, X., et al. (2021) · arXiv · doi:10.48550/arxiv.2104.12369 · Cited by 172
Survey of Machine Learning Accelerators
Reuther, A., Michaleas, P., Jones, M., Gadepally, V., Samsi, S., & Kepner, J. (2020) · 2020 IEEE High Performance Extreme Computing Conference (HPEC) · doi:10.1109/hpec43674.2020.9286149 · Cited by 163
Scaling Laws for Autoregressive Generative Modeling
Henighan, T., Kaplan, J., Katz, M., Chen, M., Hesse, C., Jackson, J., Jun, H., Brown, T. B., Dhariwal, P., Gray, S., et al. (2020) · arXiv · doi:10.48550/arxiv.2010.14701 · Cited by 150
Deep Learning's Diminishing Returns: The Cost of Improvement is Becoming Unsustainable
Thompson, N. C., Greenewald, K., Lee, K., & Manso, G. F. (2021) · IEEE Spectrum · doi:10.1109/mspec.2021.9563954 · Cited by 139
Measuring the Algorithmic Efficiency of Neural Networks
Hernandez, D., & Brown, T. B. (2020) · arXiv · doi:10.48550/arxiv.2005.04305 · Cited by 89
Truth, Lies, and Automation: How Language Models Could Change Disinformation
Center for Security and Emerging Technology, Buchanan, B., Lohn, A., Musser, M., & Sedova, K. (2021) · doi:10.51593/2021ca003 · Cited by 74
Self-citationHernandez, D., Kaplan, J., Henighan, T., & McCandlish, S. (2021) · arXiv · doi:10.48550/arxiv.2102.01293 · Cited by 51
AI Chips: What They Are and Why They Matter
Center for Security and Emerging Technology, Khan, S., & Mann, A. (2020) · doi:10.51593/20190014 · Cited by 47
The Carbon Footprint of Bitcoin
Stoll, C., KlaaBen, L., & Gallersdörfer, U. (2019) · doi:10.2139/ssrn.3335781 · Cited by 43
Meta Pseudo LabelsPreprint
Pham, H., Dai, Z., Xie, Q., Luong, M.-T., & Le, Q. V. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2003.10580 · Cited by 28
AI Education in China and the United States: A Comparative Assessment
Peterson, D., Goode, K., & Gehlhaus, D. (2021) · doi:10.51593/20210005 · Cited by 21
Keeping Top AI Talent in the United States
https://cset.georgetown.edu/publication/keeping-top-ai-talent-in-the-united-states/, Zwetsloot, R., Dunham, J. H., et al. (2019) · doi:10.51593/20190007 · Cited by 20
Poison in the Well: Securing the Shared Resources of Machine Learning
Lohn, A. J. (2021) · doi:10.51593/2020ca013 · Cited by 19
Self-citationStrengthening the U.S. AI Workforce
Zwetsloot, R., Heston, R., & Arnold, Z. (2019) · doi:10.51593/20190003 · Cited by 17
U.S. AI Workforce: Policy Recommendations
Gehlhaus, D., Koslosky, L., Goode, K., & Perkins, C. (2021) · doi:10.51593/20200087 · Cited by 16
Immigration Policy and the Global Competition for AI Talent
Huang, T., & Arnold, Z. (2020) · doi:10.51593/20190024 · Cited by 11
Most of America’s “Most Promising” AI Startups Have Immigrant Founders
Huang, T., Arnold, Z., & Zwetsloot, R. (2020) · doi:10.51593/20200065 · Cited by 9
Immigration Policy and the U.S. AI Sector
Arnold, Z., Heston, R., & Zwetsloot, R. (2019) · doi:10.51593/20190009 · Cited by 8
Yagnik, N. S. (2021) · CRC Press eBooks · doi:10.1201/9781003140474-6 · Cited by 1
Cited by
The Computational Limits of Deep Learning
Thompson, N. C., Greenewald, K., Lee, K., & Manso, G. F. (2020) · arXiv · doi:10.48550/arxiv.2007.05558 · Cited by 538
Compute Trends Across Three Eras of Machine Learning
Sevilla, J., Heim, L., Ho, A., Besiroglu, T., Hobbhahn, M., & Villalobos, P. (2022) · 2022 International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn55064.2022.9891914 · Cited by 363
Predictability and Surprise in Large Generative Models
Ganguli, D., Hernandez, D., Lovitt, L., DasSarma, N., Henighan, T., Jones, A., Joseph, N., Kernion, J., Mann, B., Askell, A., et al. (2022) · arXiv · doi:10.48550/arxiv.2202.07785 · Cited by 324
Intelligent Computing: The Latest Advances, Challenges, and Future
Zhu, S., Yu, T., Xu, T., Chen, H., Dustdar, S., Gigan, S., Gunduz, D., Hossain, E., Jin, Y., Lin, F., et al. (2023) · Intelligent Computing · doi:10.34133/icomputing.0006 · Cited by 290
The Computational Limits of Deep Learning
Thompson, N., Greenewald, K., Lee, K., & Manso, G. F. (2023) · Ninth Computing within Limits 2023 · doi:10.21428/bf6fb269.1f033948 · Cited by 495
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
- gold
- doi.org
- 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
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.