The constraint on the next round of AI is the plant required to store, move, and power the bytes.
The International Energy Agency's Energy and AI report, published 10 April 2025, puts global data-center electricity at about 415 terawatt-hours in 2024, roughly 1.5% of the world's electricity. In the IEA Base Case that figure more than doubles, to around 945 TWh by 2030, a bit more than Japan's electricity use today. Consumption has grown near 12% a year since 2017, more than four times the growth of total electricity use. In the United States, data centers account for nearly half of electricity-demand growth through 2030.
Operators are answering with buildings, megawatts, and fabric. Meta's March 2024 infrastructure note describes two clusters of 24,576 H100 GPUs — one on RDMA over Converged Ethernet, one on NVIDIA Quantum-2 InfiniBand — both interconnecting 400 Gbps endpoints, and both used for Llama 3 training. A June follow-up states the reason the network exists: large-scale training transfers vast amounts of data between GPUs in lockstep, and a slow exchange on a subset of devices slows the whole job. In that write-up, an unoptimized large cluster utilized anywhere from 10% to 90% of the small-cluster baseline until the network, the scheduler, and the collective library were designed together.
That is a capacity story told in hardware. Every additional byte of checkpoint, dataset, activation, and gradient is still paid for four times: storage to keep it, bandwidth to move it, compute to process it, and energy to power all three. Another hall does not change the multiplier. It scales it.
The ratio only counts if the bytes come back
Lossless compression is the other lever, and it only counts if the original comes back byte for byte. Matt Mahoney's Large Text Compression Benchmark ranks programs on enwik9 — the first 10⁹ bytes of a 2006 English Wikipedia dump — by compressed size plus the decompressor. The rules are explicit. A ratio you cannot read is not a result. Research systems have posted dramatic ratios for years and stayed in the lab, because reading the data back cost more than the bytes were worth.
On that public board, as listed 6 September 2026, Overwrld's altxs 1.0.0 is #1 by official total size: 93,434,410 bytes of payload, a 13,490,401-byte decompressor, 106,924,811 bytes total. The next row, nncp v3.2, is 107,261,318. That listing is the research proof. It is a result on a named public corpus, decompressor counted. It is not a figure for a customer's video file or a training run. We do not publish those until a protocol is registered before the measurement. The benchmarks page keeps the row, and the distinction, in one place.
The listed altxs 1.0.0 decode is slow on the benchmark hardware. That is allowed by the board, and it is the wrong machine for a product. The product proof is separate: a portable lossless binary that keeps most of the ratio and makes decode cheap enough to place next to the data. Encode can be expensive. Decode has to be cheap, because the data is read more often than it is written — checkpoints, archives, contribution media, the corpora a cluster will touch again.
What the buildout does not purchase
None of this retires a 400 Gbps fabric or a substation. It changes how much useful work that fabric and that substation can carry. Less data moved means less infrastructure required, and the saving compounds across storage, bandwidth, compute, and energy. The argument for why that is capacity, rather than a feature you turn on after the hall is sized, is in Compression is capacity.
The buildout will keep buying power. The open question is whether it also buys the bytes back.