Machine-learning compression does not have a single carbon break-even point, according to a new modeling study. The point at which the emissions from training and running the BOA compressor match the carbon cost of storing uncompressed data shifts sharply with the carbon intensity of the electricity used for the computation. The finding means that the environmental case for compressing data with machine learning depends on where the work is done, not just on how much storage it saves.
The work is a proof-of-principle calculation rather than a measurement of a live deployment. Its authors modeled the carbon-equivalent cost of training, compression and decompression, then compared that cost with the embodied carbon of the additional storage required for uncompressed data. In this model, the break-even size was the uncompressed dataset size at which those two modeled costs were equal.
The document is an arXiv version 1 preprint dated 20 August 2026. Its result is presented as a scenario-dependent conclusion tied to the stated electricity and storage assumptions, rather than as a threshold that can be applied to every dataset or computing site.
What the model compared
For the demonstration, the researchers used a 49.92 MB bundled CMS file named CMS_DATA_float32.bin. Each record contained 24 float32 features. They checked whether training and inference carbon costs behaved linearly, then scaled the results from that file to the larger modeled dataset size shown in the paper’s analysis.
BOA was modeled as a two-layer Mamba-v1 system with a hidden dimension of 64, a byte vocabulary of 256 and a sequence length of 10,000 bytes. The training setup used batches of 5, eight FP32 epochs, a learning rate of 5 × 10−4 and random seed 42. These details define the particular configuration behind the reported estimates; they do not describe every possible machine-learning compressor.
Energy use for the processor, graphics card and memory was measured with CodeCarbon in one-second machine mode. The CPU estimate used 8 watts based on thermal design power scaling for a 2.00 GHz Intel Xeon processor, while graphics-card power was obtained through nvidia-ml-py. The measured energy was then converted into carbon-equivalent values using country-average electricity-mix scenarios.
The model also ran compression-decompression round trips with ZSTD and LZMA, giving the BOA results a comparison with standard algorithms under the same setup. The reported trade-off was consistent across that comparison: BOA generally achieved a better compression ratio, but it had lower throughput and higher carbon intensity per unit of processed data.
Storage changes the answer
The storage side of the calculation compared hypothetical hard-disk-drive and tape scenarios in the UK, each with a five-year lifetime. The model included both embodied carbon—the emissions associated with making the storage equipment—and operational carbon from using it. The authors report that tape had a lower carbon-equivalent footprint than HDD storage, but slower access to the data.
The HDD proxy represented an 18 TB Seagate Exos X18 drive with 5.3 watts of idle power, 27 kgCO2e per drive and a proportionality factor of 1.50 gCO2e per gigabyte. The LTO-8 tape proxy used 12 TB cartridges with total lifecycle carbon of 13.72 kgCO2e, including 5.70 kgCO2e of operational carbon and 8.02 kgCO2e associated with non-use. These are model inputs, not measurements of every storage product or facility.
That distinction matters because a storage system is not chosen on carbon alone. The model presents tape as a lower-footprint option with slower access, while the compressor presents a higher-compression option with slower processing. The resulting environmental calculation therefore combines storage capacity, access expectations, processing speed and the electricity mix rather than reducing the question to a single compression ratio.
Why the result remains provisional
The modeled case used full-dataset training followed by one compression round and one decompression round on an existing temporary disk. The authors treated that arrangement as an upper-bound energy scenario because training dominated the energy consumption, and they deferred broader training and deployment patterns to future work. Repeated use, different retraining schedules or other ways of sharing the training cost could produce different modeled balances, but those scenarios were not evaluated in the supplied analysis.
The calculation also counted only operational CPU, GPU and RAM costs for the machine-learning compression. The GPU was assumed to be already available, and its use was treated as negligible compared with other Large Hadron Collider uses. As a result, the study is not a full lifecycle assessment of all computing infrastructure or a carbon audit of a real deployment.
The single input file limits how far the result can be generalized. Although the authors checked linearity before scaling its training and inference results, broader validation was not established. The modeled break-even could change with other datasets, hardware, electricity mixes, storage policies, storage lifetimes or facility-efficiency assumptions.
The storage assumptions have their own qualifications. The scenarios were hypothetical UK examples, and the model set power-usage effectiveness, or PUE, to unity rather than adding a data-centre overhead. The tape operational-carbon allocation was also not matched to the HDD choice because it reflected the availability of existing data. Those choices make the output useful for examining the stated scenario, but not for declaring BOA environmentally preferable to ZSTD or LZMA at scale.
The paper’s practical message is therefore a question for system designers: do the storage savings justify the extra computation under the conditions in which the compressor will actually be used? The reported evidence supports a model-specific, directional answer for the stated BOA configuration, electricity scenarios and storage assumptions. It does not establish a universal break-even size or show a causal environmental benefit or harm in real-world deployments.
Paper data and sources
Original title: Green BOA: Determining the environmental break-even point for ML-based data compression
Authors: Caterina Doglioni, Akshat Gupta, Thomas Elliott et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text