All products
/ 05Coming soon

Quandong

Twice the work from the same power.

Quandong is a resource operating system for datacenters. It runs more work on the machines you already have by packing them against how workloads actually behave together, not against each job's worst case — and it does that without ever crossing a memory, HBM, or breaker limit. Here's where it stands today.

2.3–2.6×Efficiency on a modeled AI datacenter, for identical work
0Breaker trips, OOMs, and HBM violations across every seed
−30 to −48%Power-draw variance removed by capping the throughput fleet
~1.8%Work traded to unlock the variance dividend

Measured in simulation, on held-out evaluation windows, averaged across seeds.

The idea

Pack to the quantile of the sum.

Most schedulers admit a job to a machine when that job's own worst case fits. Add up enough worst cases and the machine sits mostly empty, because those peaks almost never land at the same moment.

Quandong admits a job when the predicted quantile of the machine's combined moment-to-moment usage stays under budget. Jobs that pull in different directions — or just vary independently — cover for one another, so a machine can safely hold far more of them than per-job math suggests. The same idea runs across CPU cores, GPU streaming multiprocessors, and the watts behind a breaker.

There's one rule it never bends: it only takes that credit on a resource that genuinely diversifies. Memory doesn't average out, so memory is treated as hard. A limit is a limit.

What we've accomplished

An itemized efficiency ledger.

Quandong runs the same offered work — CPU services, GPU inference, training, batch, and deliberately planted waste — through two complete stacks: standard practice, and quandong. Then it prints exactly where the difference came from. On a modeled, training-heavy AI datacenter, the layers stack up like this:

LayerMechanismTypical gain
CPU consolidationPortfolio-quantile packing in place of request-based reservations~2.8–3.0×
GPU consolidate + harvestInference co-located on shared SMs (HBM held hard), preemptible training backfilling the diurnal troughs~1.6–2.5×
Purge reclaimRecovering capacity lost to zombie and idle reservations~1.1×
Power oversubscriptionRacks packed to the predicted power quantile instead of nameplate~1.05–1.4×
Variance dividendCapping the throughput fleet shrinks power variance, which unlocks more oversubscription+0.2–0.4

The headline is just the provisioned watts of the old stack divided by the provisioned watts of quandong for the same work — around 2.3 to 2.6×. Operating savings, like roughly 1.2× more tokens per joule from GPU power capping and 25–35% lower energy cost from deadline-aware load shifting, are reported next to the headline but never multiplied into it.

The result we didn't expect

Running a little slower to fit a lot more.

Power capping and power oversubscription look like separate levers — one saves energy, the other saves capacity. They aren't separate, and the overlap is the interesting part.

A rack is packed against the predicted quantile of its combined power, and that quantile depends on the variance of the power, not just its average. Capping the throughput fleet clips the top off each machine's power curve — which barely moves the mean but cuts the variance sharply. Less variance means more headroom under the same breaker.

So capping on its own frees no rack capacity, but capping plus oversubscription frees more racks than oversubscription alone. Across eight held-out seeds: an auto-selected cap near 0.80, power-draw variance down 30–48%, and about 23% fewer throughput-fleet racks on top of oversubscription — for roughly 1.8% of work slowed and zero breaker overage. And it's falsifiable: a perfectly flat power fleet has no tail to clip, so the dividend is exactly zero, which a regression test holds in place.

How it stays honest

Every number has to survive a test.

What's inside

The pieces.

core

Packing core

Covariance-aware quantile admission — the portfolio check that decides what actually fits.

gpu

GPU harvest

SM harvesting and inference co-location, with HBM held hard and training filling the troughs.

power

Power

Rack oversubscription, throughput capping, and deadline-aware load shifting.

purge

Purge

Reclaims the capacity quietly lost to zombie and idle reservations.

fabric

Fabric

A model of the datacenter's topology that the rest of the system plans against.

eval

Eval

The held-out evaluation and multi-seed regression harness that keeps every claim honest.

Coming soon.

Quandong isn't public yet. It's the infrastructure we're building to run advanced AI efficiently — and when it's ready, it will show up here.