c1 us-westconnecting

Fig. 1Cluster C1, particle view

Racks you can hold by the hour

Neocloud for AI teams. Real H100 racks in us-west, live telemetry, hours you can hold.

Request access Read the docs

Live now

cluster: C1
region: us-west
gpu: 8x H100 SXM
util: --
jobs: --
uptime 30 d: --

preview

One rack, one feed. Compute, scheduler, telemetry, documentation. The public cloud is built this way. RACK is, too. Hours you can hold.

Cluster C1initialized

Request access

util

0 %

jobs

gpus

gpu0--
gpu1--
gpu2--
gpu3--
gpu4--
gpu5--
gpu6--
gpu7--

Preview data, 1 Hz

Fig. 2 Job lifecycle

Run training, inference and batch workloads with the telemetry, scheduling and hours you can hold of a cloud

Cluster C1 ▸

GPU

gpu0

gpu1

gpu2

8x H100 SXM 80 GB, one NVLink fabric
3.2 Tb/s InfiniBand to shared storage
Owned rack, power and network

Scheduler ▸

Job	GPU	State
job-01k4	gpu2	RUNNING
job-01k3	gpu5	DONE
job-01k2	gpu0	DONE

Reservedheld

reservedon demandidle

Reserved hours held for your team
On demand billed per minute
Batch fills idle capacity, preempted first

Telemetry ▸

C1_UTIL_AVG--

1 Hz per GPU: util, memory, temp, power
24 h history on every page
Same feed site, console and reports

Documentation ▸

▣ read.racksystems.io17 pages

▸ Quickstartguide
Run a jobguide
Schedulerarchitecture
SLAarchitecture
APIfor-labs

Quickstart curl and Python, real endpoints
SLA written before general availability
Reports capacity, monthly

For years, AI teams had two options for GPU capacity. Both required compromise.

#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#

Hyperscaler on demand

Speed, but at a cost

                                  
                                  
                                  
       `+--------+                
     `.                           
    +--------+                    
                                  
      ~> STATUS          ┌─────┐  
      ERROR        ──────┘     └  
      CAP-429-WAIT                
       `+-------`+-------`.       
     `.       `. |     `. |       
    +--------+--------+   |       
    |        |   |    |   |       
    |        |   |    |   |       
    |        |  `+    |  `+       
    |        |`.      |`.         
    +--------+--------+--`.       
     `.       `. |     `. |       
    +--------+--------+   |       
    |        |   |    |   |       
    |        |   |    |   |       
    |        |  `+    |  `+       
    |        |`.      |`.         
    +--------+--------+--`.       
     `.       `. |     `. |       
    +--------+--------+   |       
    |        |   |    |   |       
    |        |   |    |   |       
    |        |  `+    |  `+

Waitlists for H100 capacity, with no date
Opaque telemetry you read a bill, not a rack
Spot revoked with two minutes notice
Egress priced to keep you

#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#

Owning a rack

Control, but at a cost

                                  
                                  
         `+------------`.         
       `.            `. |         
     `.            `.   |         
    +-------------+     |         
    |             |     |         
    |             |    `+         
    | RACK        |  `.           
    |             |`.             
    +-------------+----`.         
       `.            `. |         
     `.            `.   |         
    +-------------+    `+         
    |             |  `.           
    | NETWORK     |`.             
    +-------------+----`.         
       `.            `. |         
     `.            `.   |         
    +-------------+    `+         
    |             |  `.           
    | POWER       |`.             
    +-------------+----`.         
       `.            `. |         
     `.            `.   |         
    +-------------+    `~> READY  
    |             |  `.           
    | COLO        |`.   18 mo     
    +-------------+

18 months from order to first job
Colo, power, network three contracts, three vendors
Idle hours paid for whether you train or not
No exit when the workload moves

For the first time, the speed of a cloud meets hours you can hold.

Real racks

A cluster is a rack we own: eight GPUs, one fabric, one power budget. Soak tested for 72 hours before it is listed.

RACK C1/SLED 0..7

GPU8x NVIDIA H100 SXM 80 GB

FABRICNVLink 4, 900 GB/s per GPU

NETWORK3.2 Tb/s InfiniBand to storage

STORAGE30 TB NVMe local, 400 TB shared

Predictable prices

Three plans, one number each. No egress, no surprise line items.

ON DEMAND2.49 USD / GPU h

RESERVED1.89 USD / GPU h

BATCH1.19 USD / GPU h

API driven from the first hour

One endpoint to submit a job, one to read its telemetry. The same calls the console makes.

~> rack init key ok c1 ok ~> rack jobs sub

job: cluster: c1 gpus: 4 image: lab/train hours: 12

Two business days to an answer

Access is by invitation during preview. Tell us what you run, we reply within two business days.

REQUESTreceived

REVIEW2 business days

INVITEconsole + API key

FIRST JOBsame afternoon

Pricing. Per GPU hour, in USD.

On demand ▸

Billed per minute with a one minute minimum. No commitment, capacity as it is free on C1.

US-WEST

8x H100

PER MINUTE

~:;=  .,,~~::;=  .,,~~::;==  .,,~~::;=
;;=  .,,~~::;==  .,,~~::;==  .,,~~::;
::=  .,,~~::;=  .,,~~::;=  .,,~~::;==
~~::;==  .,,~~::;=  .,,~~::;==  .,,~~
,,~~::;=  .,,~~::;==  .,,~~::;=  .,,~

Meter2.49 USD / GPU H

Job

train-1

Job

eval-1

Reserved ▸

One month or more. Hours are held for your team and nobody else runs on them.

MONTH_1720h
MONTH_2720h
MONTH_3720h
SPOT0h

▶idle▶held

Reserved1.89 USD / GPU H

Batch ▸

Preemptible. Runs when reserved capacity is idle, and is the first to give way.

0 | | o 0 | | o 0 | | o 0
| 0 | | | 0 | | | 0 | | |
| | 0 o | | 0 o | | 0 o |
o | | 0 | o | | 0 | o | |
0 o | | 0 o | | 0 o | | 0

Job

batch-embed

◆job.rack

plan     = "batch"
preempt  = true
price    = 1.19
fills    = "idle reserved"

Compared to list prices at the four largest providers, updated monthly.

Regions. us-west online, eu-central next.

Fig. 3Regions and routes

Drag to rotate

Regions

us-west
Hillsboro, Oregon
online
eu-central
Frankfurt
next

Median round trip to us-west

San Francisco14 ms
New York68 ms
Frankfurt146 ms

Capacity by region

region	clusters	gpus	util
us-west	1	8	--
eu-central	0	0	planned

preview

Request access. Write the request, we reply within two business days.

◆ request.rackdraft

cluster   = "c1"
region    = "us-west"
email     = ""
team      = ""
size      = ""
workload  = ""
hours     = ""
plan      = ""

Estimate / month--

Pick a workload and hours

01Review we read the request and check capacity on C1
02Invite console access and an API key
03First job a reserved window, same afternoon