Compute
Compute on surplus Energy
AI training and batch inference — interruptible HPC on a Flexbox at the surplus. Checkpoint in the contract.
Workloads
Interruptible loads on a Flexbox at the surplus.
-
AI
Training and batch.
The load Kelvin is built for is interruptible AI.
Long runs that write a resume point. Queued inference that can wait a minute.
-
HPC
Simulation and rendering.
Same interruptible contract — jobs that can wait.
High-density runs and framesets that queue. Pause inside the window and resume.
-
Cloud
Batch and burst.
Capacity is on request.
Burst that can pause. Queued jobs that can wait.
-
Edge
The module at the source.
Behind the meter at generation.
A Flexbox at the source. Extra grid load stays low.
-
Bitcoin
Compute that flexes.
Paid on stand-by, paid on activation, paid while active and paid in yield. No matter the market price of Bitcoin, we produce for Production Price.
One pause-able load. Same pause-and-resume as training.
Why it holds
The site is the jurisdiction.
Sit the job at the surplus — behind the meter or spare industrial.
-
01
Local
The surplus is local. So is the job. Energy.
Locked-in surplus cannot be copied without that Energy.
-
02
Behind the meter
Spare industrial or behind the meter at generation.
Extra grid load stays low.
-
03
Based in Stockholm.
Kelvin Compute AB. First Sites in Hälsingland. 10 MW. Status Plot.
Postal c/o only.
-
04
Electrons first.
Compute, Flexibility-as-a-Service, and Thermal/Heat.
One connection. Three public products.
-
05
Computational sovereignty.
The job sits at the surplus. Not a remote campus on a queued feeder.
The job does not leave the Energy.
-
06
Weeks against years.
The module sits when the pad is ready.
Not a feeder queue.
The contract
Four load tiers. One dispatchable site.
Checkpointable means the job writes a resume point. When the grid calls, the job pauses inside the contract window and comes back.
01
Buffer
Not a public product — it rides the megawatt. Interruptible HPC. First off, first on.
02
Batch
Async AI and queued inference. Pause inside the contract window. Cost of flexing is zero if the window is respected.
03
Training
Checkpoint and resume — no lost work. Forgone training time, not an SLA breach. Semi-flex on the same Flexbox.
04
Firm inference
User-facing, latency-bound. Never curtailed.
Interruption windows are contracted per tenant. See Flexibility-as-a-Service.
The rack
Flexbox is the rack for that load
The rack at the surplus — not a second campus.
-
Power
≈1 MW / 40 ft
Operating power tracks loaded server draw.
-
Compute
4 racks · 336
Up to 336 servers. Air or liquid in the same 40 ft ISO. Stackable two-high on the same pad (≈2 MW / 672). 20 ft on request.
-
Dispatch
<0.7 s
Built to start, pause and resume in well under a second. Flexibility-as-a-Service is in the contract.
-
Connect
LV or MV
Behind the meter or grid-connected. Heat capture designed in — Thermal.
Full stack at the surplus
-
Workload
The workload.
Interruptible batch: training, batch inference, simulation, rendering.
-
Kelvin
Full-stack.
Infrastructure. Compute, Hardware, Software, Flexbox. Flexibility-as-a-Service, Thermal/Heat. Operations. Kelvin handles everything.
Not always-on colo. See Compute Flexibility.
Cooling
Air. Liquid. Closed-loop.
The Flexbox packages high-density servers and forced-air or liquid cooling in a 40 ft ISO.
-
01
Air
Forced air. Heat flange still ships.
-
02
Liquid
Direct or closed-loop. Recoverable Heat. See Thermal grade.
-
03
Closed-loop
Closed-loop on the unit. Grid in; Heat flange out.
Pad
One module on the pad.
Compute, controller, Heat flange, connection.
01 / 04
01 FB-40
One module. Pre-tested.
Back to the unit · Compute Flexibility · Energy · Thermal · Home