I wrote code that could cut a data center's footprint by 90%. It is not running in any data center yet.

I build a project called The Human Brain App. One corner of it — the Compression Initiative — runs real compression algorithms in your browser: gzip through the browser's own CompressionStream, JPEG re-encode on a canvas, and foveated pixel masking that keeps full detail only where your eye actually looks. The ratios it prints are measured, not simulated.

That made me ask an honest question: if a real data center adopted this code today, what would actually change? Here is what I found — with the receipts, and with the gaps clearly marked.

The problem is real, and it is not mine

Data centers already burn roughly 1–1.5% of global electricity (IEA). The average facility runs at a PUE of ~1.55 (Uptime Institute), meaning more than a third of the power a site pulls is cooling and overhead, not compute. Up to 70% of paid-for servers sit idle (McKinsey; ~30% are fully comatose). The sector emits on the order of ~100 Mt CO₂ a year (IEA estimate).

These are published industry numbers. I did not measure them. A data center does not need my code to have these problems — it already has them.

What the code actually does today

Four stations are live in the lab right now, and you can watch them work:

• Text — gzip, lossless. Real CompressionStream in your browser. The before/after byte counts are measured.
• Image — JPEG re-encode. A canvas re-encodes at an adaptive quality; you see the size drop and the visual cost in real time.
• Foveated — perceptual. Fewer bits in the periphery, where the retina doesn't resolve full detail. Modeled on a 1/16 periphery resolution factor.
• Folder ingest — before & after. Point it at a real folder of photos on an external drive, compress to a trial target, and confirm it still looks fine.

These four run now. Open the lab and the bytes shrink in front of you.

What it would do if adopted — modeled, not deployed

If a modeled 50 MW hyperscale facility adopted the full stack the lab points at — high-density compression, idle elimination, liquid cooling, district heat-reuse — the planning estimate is roughly:

• PUE 1.8 → 1.12 (vs. ~1.10 best-in-class hyperscale, Google)
• ~50% electricity cut
• ~90% fewer storage racks from compression alone
• ~340 GWh saved per year
• ~134 kt CO₂ avoided per year (at 0.39 kg CO₂/kWh US grid average)

I need to be plain about this: these are modeled estimates with stated assumptions (50 MW IT load, 8,760 h/yr, 90% storage-data reduction, idle eliminated). They are planning estimates, not guarantees, and not a measurement of any facility I operate — because I don't operate one. "Adopted" means a real operator wires this into a real corpus. I have not done that.

The honest gaps

• The code is not running in any data center. I am one builder with a lab, not a company with an SRE team. A pilot is the missing step.
• Foveated savings are modeled, not measured on real display traffic. The 1/16 periphery factor is a stated assumption, not a field result.
• The savings station assumes the compression ratio holds at scale. Real corpora compress differently; a pilot would measure the true ratio on the operator's actual data.
• The green-facility design (PUE 1.12, liquid cooling, heat-reuse) is synthesized from industry-standard benchmarks into a target, not a built and commissioned building.

Why I'm writing this

I think the honest version of "AI will save energy" is the one that shows its work: here is code that already compresses real data, here is the modeled upside if a real operator adopts it, and here is the exact line where my claim stops being a receipt and starts being an assumption.

If you run a data center and want to point the compression code at a real corpus and see what it actually saves — the lab is live, the methods are open, and I would rather you break it than have me oversell it.

The Human Brain App — Compression Initiative is live. The bytes shrink in your browser. Come measure them.

— Steve