Mantissa Request capacity

103 mCampus KJN-1, Kajaani

From the grid to the gate.

Mantissa builds, owns and runs the GPU clusters behind its inference cloud. Every layer, from the substation to the silicon, is ours to tune, so your models answer faster and cost less to serve.

420MW of contracted power across six regions, 174 MW of it on this campus.

Scroll to travel twelve orders of magnitude

White clay model of the Kajaani campus from above: six data halls, cooling yards and a substation, with one hall open to show its racks.

102 mHall 3

Each hall is its own failure domain.

Separate utility feeds, separate cooling loops and a separate network spine. When something fails it fails small, and traffic moves to a healthy hall in under a second.

99.995% measured availability over the trailing twelve months.
6regions21halls1,440racks in this hall
Clay model of a data hall with its roof removed: rows of racks, hot aisle roofs, cable trays and cooling pipes.

100 mRack C-0412

Water does the cooling, not air.

Each rack draws up to 132 kW and sheds it through a closed liquid loop that runs to a cold plate on every accelerator. In Kajaani the outside air finishes the job for most of the year.

1.12 annualised power usage effectiveness (PUE) across the fleet.
A rack of compute trays with one tray slid out on its rails, coolant pipes above and network cables down the front.

10−1 mCompute node

Eight accelerators, one fabric.

Every node links its eight accelerators through a switched fabric at 1.8 TB/s each, so a 70B model shards across them without waiting on the network.

31,400 output tokens per second per node, 70B model at FP8.
Top view of an open compute tray: eight accelerator modules, seven under cold plates and one open, with coolant tubes and board traces.

10−2 mAccelerator package

Memory sits a millimetre from compute.

Eight stacks of high bandwidth memory share an interposer with the die. Decoding speed depends on how fast weights reach the cores, and here they arrive at 8 TB/s.

8TB/s memory bandwidth per accelerator, with 192 GB on the package.
An accelerator package: the die in the centre with four memory stacks on each side and capacitors around the edge.

10−3 mSilicon, down to 10−6 m

Latency is decided down here.

Our scheduler keeps hot weights and attention caches beside the cores that need them, and sends each request to the hall nearest your users.

support-assistant-prod70B instruct, FP8, eu-north-1 (Kajaani) Healthy
TTFT p50140ms
TTFT p99410ms
Tokens/s18,240
Time to first token, last 15 minutes
0 msp50 140p99 410600 ms
The surface of the die as a clay relief: compute clusters in a grid, a cache band across the middle and memory interfaces at the edges.

10−9 mSignal

Every request ends at a gate.

A single token is trillions of transistor switches. Because we run every layer above them, we can quote capacity by the megawatt and price inference by the token.

The Mantissa mark, three concentric squares drawn in blue, formed from the data paths above a transistor.

Pricing

Priced by the hour or by the token.

Run your own models on dedicated accelerators, or call ours through one API. Same halls, same fabric, same people on call.

Prices
ProductBilledPrice
On-demand acceleratorsPer accelerator hour, billed by the second$2.39
Reserved clusters, 12 monthsPer accelerator hour, from 64 acceleratorsfrom$1.74
Dedicated endpointsPer endpoint hour, eight accelerators, autoscalingfrom$17.90
Serverless, 70B classPer million tokens, input and output$0.20 / $0.60
Serverless, 8B classPer million tokens, input and output$0.03 / $0.06
NVMe block storagePer TB month, three replicas in one region$62

Prices are in US dollars before tax. Committed use discounts start at 512 accelerators.

Enterprise

Talk to an infrastructure architect.

Tell us the model, the traffic and the regions. Within one business day you get a capacity plan and a fixed price for the first twelve months.

A site engineer crouching beside a row of server racks in a brightly lit aisle.
Site lead on call, every regionKajaani, 03:12

Request a capacity review

We reply from a named architect, not a queue.

Single-tenant hallsFrom 4 MW, with your own cage, keys and network.
ISO 27001 and SOC 2 Type IIAudited every year, reports under NDA.
Data stays in regionEU residency in Kajaani and Stavanger.