United States · halls + owned CPU
First rack, or 500 kW leftover. The tenant buys the computer. The hall sells the kW. We introduce both.
Work
Motion 1 · ~30 kW
A tenant already in a US colo, or willing to pick one in 90 days, moves CPU work onto a rack-scale commercial system that brings public cloud capabilities directly into the hall. Pipelines, inference, HPC, VMware exit, general compute. Not a warehouse swap. Not a GPU rip.
Entry is one factory-cabled rack. Up to 4,608 cores / 9,216 threads at ~30 kW. First instances in about an hour after power and uplink.
Motion 2 · 500 kW+
The buyer already chose the machine. It still needs a US home. We pair that fleet with published leftover power of 500 kW and up.
The designated OEM invoices the rack. The facility invoices reserved kW. We do not take the hardware paper.
Why Aislewatt
Everyday colo and data-center customers still rent most of their CPU by the hour. That is the right buy for burst. It is the wrong buy for always-on pipelines, inference prep, VMware, and the rest of the bronze layer that never leaves the building. The better pattern is hardware the customer owns, seated on leftover US kW.
Aislewatt places that pattern. We do not manufacture the rack. We do not own the hall. We match a factory-cabled, US-built computer to a cage that already has power, or a fleet to leftover capacity that is actually published.
Data
We do not commodify, share, or mine what runs on the rack. Consent is express, or it does not happen. The hall does not get a copy. Neither do we.
Independence
Aislewatt is not a subsidiary of a listed company and will not be owned by one. No parent that needs the workload as training data or as a line on someone else's earnings call.
Ownership
The customer procures and holds the rack, the storage, and the equipment in the cage. Colocation is the seat. Ownership is not a slogan for a lease.
We seat rack-scale commercial systems that bring public cloud capabilities directly into on-premises data centers. Designed and assembled in the United States. Compute, networking, storage, power, and the control plane land as one factory-cabled rack. Manufactured and shipped in weeks. Once the hall has three-phase power and an uplink, instances are up in about an hour. Software travels with the rack. The designated OEM does not meter it. The customer holds the keys.
Elastic cloud on the API. On-prem control of hardware, software, and data.
Origin
Designed and assembled domestically. Shipped as one machine.
Lead time
Factory schedule, not a six-month integration program.
Turn-up
Power, uplink, pipelines or VMs. Plug and play, not rack-and-stack.
Software
Control plane, virtualization, networking, storage. No separate license meter.
What the hall needs before it rolls in
| Item | Current generation |
|---|---|
| Cores / threads | Up to 4,608 physical cores, 9,216 threads |
| Compute sleds | Up to 24 per rack · AMD EPYC, 192 cores per sled |
| Memory / storage | Up to 36 TiB DRAM · up to 6.5 PiB NVMe |
| Power | 3-phase, 180–300 V · 21.6 kW redundant, 30 kW non-redundant |
| Thermal | Air cooled, standard aisle · ~123 kBTU/hr |
| Footprint | Standard rack · ~92.7 × 23.7 × 41.8 in · ~2,518 lb full |
Always-on CPU is most of the bill most tenants still pay a public cloud to run. That bronze layer is often 40 to 70% of the platform invoice. Move it onto an owned rack in a hall you already lease and the unit economics change in three places at once.
of the public-cloud platform bill is bronze-layer CPU. That is the work this rack is for.
Power and network to running jobs. Not weeks of BIOS, licenses, and cabling.
Lower unit cost versus a legacy 1U row. Often more versus hyperscale CPU meters.
Compute density. Same cores on a fraction of the tiles you already pay for.
Cooling efficiency. Overhead power from about 25% of draw to under 2%.
Less total power for the same work. One DC bus. That is what moves PUE.
Compute
Own the rack, depreciate it, run it. No hypervisor tax, no surprise software SKUs, no egress for work that never needed to leave the building. Public-cloud APIs stay. The meter does not.
Space
Conventional 1U stacks spend a row on servers, TOR switches, and management. A rack-scale computer folds that into one cabinet. Same work, a fraction of the floor, leftover tiles stay leftover.
Power
One DC power path instead of a PSU in every chassis. Cooling overhead drops from a large share of the draw to a rounding error, which is what actually moves PUE. The hall sells kW you use. The tenant stops paying for wasted conversion and fan.
Method
Public cloud still wins on credit card and APIs. It loses on unit economics once the bill is mostly bronze pipelines, idle cores, and egress you did not mean to buy.
Aislewatt exists for the unglamorous middle: an owned rack in a hall you can name, with a support path that is not a ticket into a hypervisor tax.
CPU that can move. A hall you already sit in, or one you will pick in 90 days. Or 500 kW leftover for a fleet already chosen.
The designated OEM warrants the computer. The hall sells power and space. Aislewatt introduces. We do not impersonate either side.
No empty megawatts on hope. New rows only with offtake. Leftover air-cooled capacity first.
Contact
Aislewatt is a two-motion desk run in the United States. If you have a cage and a CPU bill, or a fleet and no hall, write. About Larsen.
larsen@aislewatt.com · Replies are read by a person. If you want off a list, say so in the thread.