Compression is capacity. Capacity is the amount of useful work a given store, a given link, and a given watt can support. If the same plant carries more bytes of meaning, you have added capacity without pouring more concrete.
When data grows, the reflex is to build: more racks, more fiber, more power. The build is real, and it is also a tax charged four times. You pay to keep the bytes, to move them, to process them, and to power all of it. Each new hall multiplies all four. Compression runs the other direction. Less data moved means less infrastructure required, and the saving compounds across storage, bandwidth, compute, and energy.
That is the job of a data capacity engine. The rest of this note is what has to be true before the sentence is allowed to stand.
The reconstruction is exact
Lossless means the decoded output matches the input byte for byte. A result exists for us only after that check passes. Anything short of it is a different product, with a different buyer and a different way to fail. We keep the two apart. Approximate codecs already exist, and they are frozen the day they ship. This is the other instrument: exact bytes, smaller.
The decompressor is in the score
A compressor can hide its model, its dictionary, or its decoder beside the payload and call the remainder the ratio. Matt Mahoney's Large Text Compression Benchmark does not allow that. Ranking is by total size: the compressed enwik9, the first 10⁹ bytes of a fixed English Wikipedia dump, plus a zip of whatever is required to read it back. The rules put the decompressor in the number on purpose.
We reached #1 on that board with altxs 1.0.0, by official total size, as listed 6 September 2026. The total is 106,924,811 bytes: 93,434,410 of payload and 13,490,401 of decompressor. The next listed total is nncp v3.2 at 107,261,318. The public row is the authority for that claim. A later model at 91.7 MB is not on the board, and it is not the rank. The benchmarks page quotes the row the way a partner will check it. The methodology says what we will not claim around it.
Decode is cheap enough to place
The constraint that kept stronger compressors in the lab was decode, not ratio. A ratio you cannot afford to read back is a paper. It does not make a store, a link, or a watt carry more work, because the plant still has to pay the cost of opening the file. Capacity shows up only when the smaller representation is the one you actually keep and the one you actually read.
The listed altxs 1.0.0 decode is slow on the benchmark machine. The board permits that. A product does not. The binary we ship is portable, keeps most of the ratio the benchmark demonstrates, and makes decode cheap enough to run where the data already lives. Those are two proofs. The benchmark says the floor moved. The binary says you can stand on it. We do not trade one for the other in a sentence.
Encode can still be expensive. That cost is a budget, set from what the asset is worth, spent once. Decode stays cheap because the read happens more often than the write.
Spend follows the value of the asset
One setting for every file is how you leave bytes on the table, or how you spend compute the file will never earn back. We choose the mode from the asset.
One model, across many kinds of data, is the mode most bytes should see. The cost of training is amortized across every domain it covers, and each training run raises the floor for all of them at once.
Specialising to a single asset is for data valuable enough to justify the compute. Compression has a floor, and that floor is specific to the file. A general tool has to be right about everything at once. A specialised pass does not.
For the highest-value data we search for a shorter exact representation, many candidates at once. Every candidate is verified to reconstruct the input bit for bit. Anything that fails the check is discarded. Nothing unverified is ever counted in a published number.
The method stays unpublished. The approach page describes what each mode does and what it costs. What matters for capacity is the shape: encode cost follows the value of the asset, and decode stays cheap.
Where the bytes are worth it
The order of markets follows the bytes, not the elegance of the demo.
Video is the beachhead. It is the largest volume of bytes in motion, paid for in bandwidth, carried by codecs that cannot move after they ship. Every increase in resolution or frame rate is a bill. A portable decoder you can place in the path is how that bill changes. The video note is the commercial shape of that bet.
AI infrastructure is next. Training clusters and the stores around them are constrained by power, interconnect, and floor space. Meta's public account of two 24,576-GPU training clusters is a picture of that constraint: the fabric exists because the GPUs have to exchange huge volumes of data in lockstep. Moving less of it between clusters is capacity you do not have to build. The timely version of that argument, with the IEA electricity figures beside it, is AI infrastructure still pays for the byte.
Cold and archival storage, contribution and mezzanine feeds, and write-once high-value data fit the same curve today. Spend the compute once at encode. Keep reading cheap forever.
The lead has to be allowed to move
A standard is frozen the day it ships. AV1 is frozen. AV2 will be frozen. A compressor that improves after you deploy it — next year's version beating the one you signed for, without a new standard or a hardware swap — compounds. A competitor catching a single release is catching a photograph. Catching a system that keeps moving is a different problem.
Size the hall for the uncompressed byte and you will overbuild, then overbuild again when the next model arrives. Power, land, and fabric will still be bought. Demand is larger than any one compressor. The order of operations is the decision: measure the bytes you can remove, with the decompressor in the score and decode cheap enough to use, before you pour the next hall for the bytes you could have kept off the wire.
The world that is waiting is the same plant, carrying more. Compression is how you get there.