What this is How it works Billing Blog
Early Access — Limited B200 Capacity

Run your code on a
ready-to-use B200 cluster.

Submit Slurm jobs for large-scale training, distillation, and fine-tuning — without building or operating your own GPU cluster.

Priority allocation & founding-customer pricing for early signups.

Also shipped: a VRAM sizing calculator →
B200 × 8per node, multi-node
Slurmnative sbatch
Per-jobbilled while RUNNING
Daystypical onboarding
01 · What this is (and isn't)

Not a GPU VM. Not a managed AI task.

A Slurm-managed B200 cluster for serious training jobs. If you've used GPU clouds before, this may look familiar. The difference is simple: you submit jobs to a ready-to-use B200 cluster, instead of renting and operating GPU servers yourself.

Category
What you do
Example
GPU VMRented instance, billed by uptime.
SSH into a VM. Install your stack. Manage it like a server. Stop it when you're done — and remember to actually stop it.
On-demand GPU instances
Pod / K8sLong-running container, always-on.
Deploy a container. Keep it warm for inference. Pay for the container, not the work — even when idle.
GPU pods / serverless endpoints
Managed AIPre-defined tasks, no code.
Pick a task (fine-tune Llama X with data Y). Configure parameters. Run. You can't bring custom training code or modify the pipeline.
Hyperscaler AI services
DIY HPC clusterYou buy and run it.
Procure hardware. Rack, network, cool, monitor. Hire HPC engineers. Wait months. Then start training.
On-prem / colocation
Dedicated clusterReserved by contract.
Sign a 6–36 month contract for an exclusive cluster. Pay full price even when nothing is running.
Reserved / committed capacity
Compute ClusterShared HPC cluster, per-job billing.THIS
Submit your code as a Slurm job. The cluster is already built, already wired, already running. You're charged only while GPUs are allocated to your job.
Slurm-as-a-Service
In one line
We give you a ready-to-use B200 cluster. You bring code. You submit jobs. You pay for what runs — nothing more.
02 · Why now

Your training is ready.
The cluster isn't.

Hyperscaler B200 capacity? Waitlisted through mid-2026. Neo-cloud reservations? Multi-month lead times and 6-figure commitments. Public HPC? Lottery, paperwork, queues.

And running a real multi-node B200 cluster yourself? That's InfiniBand cabling, NCCL tuning, Slurm config, and headcount you don't have.

The bottleneck isn't your model. It's the infrastructure between you and the GPUs.

B200 ACCESS · MAY 2026
Hyperscalers
enterprise contract required
Waitlist
Reserved neo-cloud
multi-month commit, $$$ upfront
2–6 months
Public HPC
application + review cycle
Lottery
Build your own
hardware + ops + headcount
Months
Compute Cluster
submit a job, get results
Days
03 · How it works

Submit a job. Get results.

Standard HPC workflow. No new SDK. Bring the sbatch scripts you've already written.

01
Top up your credit balance
Prepay for GPU-time using points. Use them when you need them — they don't expire on a schedule.
02
SSH in to the login node
Standard login environment. Push your dataset, code, and checkpoints to the shared filesystem.
03
Submit with sbatch
Your job enters the queue. Slurm allocates the GPUs, CPUs, memory, and walltime you requested across one or more B200 nodes.
04
Job goes RUNNING → billing starts
The moment GPUs are bound to your job, points start drawing down — by the GPU-allocation hour. Logs and checkpoints stream to the shared FS.
05
Job ends → billing stops
GPUs are released back to the pool. Pull your results. Submit the next job. No idle charges. No "did I leave it on?"
~/jobs/llm-70b/train.sbatch
#!/bin/bash
#SBATCH --job-name=llm-pretrain-70b
#SBATCH --partition=b200x8
#SBATCH --nodes=8
#SBATCH --gpus-per-node=8
#SBATCH --ntasks-per-node=8
#SBATCH --cpus-per-task=12
#SBATCH --time=72:00:00
#SBATCH --output=logs/%j.out
 
module load cuda/12.4 nccl
source ~/venv/bin/activate
 
srun python -m torch.distributed.run \
  --nproc_per_node=8 \
  --nnodes=$SLURM_JOB_NUM_NODES \
  train.py --config configs/70b.yaml
 
$ sbatch train.sbatch
Submitted batch job 8472
04 · Billing model

Pay only while RUNNING.

No idle charges. No "always-on" containers. No 12-month commits before your first epoch.

Top up prepaid points. Points are drawn down while GPUs are allocated to your job — meaning only when your Slurm job state is RUNNING.

STATE 1
PENDING
No charge
STATE 2
RUNNING ●
GPU-hour rate
STATE 3
COMPLETED
No charge
STATE 4
CANCELLED
Stops when GPUs release
→ Charged by GPU-allocation hours. Usage is based on the number of allocated GPUs and the actual time your job spends RUNNING.
→ No GPU idle charges while jobs wait in the queue or after they finish. The cluster being "on" doesn't cost you anything.
→ Cancel anytime with scancel. Billing stops at the moment GPUs are released.
→ Storage billed separately at a low flat rate per GB-month. Datasets and checkpoints persist between jobs.

Exact point pricing and minimum top-up are shared after early-access signup — tailored to your run profile.

05 · Why Compute Cluster

Built for serious training runs.

Not a single VM. Not a managed task. A real cluster, on-demand, by the job.

01 / READY
Pre-built, pre-tuned, pre-wired.
Skip the InfiniBand cabling, the NCCL tuning, the Slurm config. The cluster is already running. You submit a job — that's it.
No infrastructure setup
02 / PER-JOB
Pay only while RUNNING.
No always-on containers. No idle VM bills. Charged only when GPUs are allocated to your Slurm job — drawn from prepaid points by the GPU-hour.
No GPU idle charges
03 / SCALE
Multi-node B200, by design.
HGX B200 nodes (8 × B200 per node) lashed together with high-speed interconnect. Pre-train 70B+ models or distill across dozens of GPUs with the bandwidth they actually need.
Per node: B200 × 8 · 180GB HBM3e each
04 / FAMILIAR
Standard HPC. Bring your sbatch.
Slurm-native. PyTorch, DeepSpeed, Megatron-LM, NeMo, JAX — whatever you already use. No proprietary SDK to learn, no vendor pipeline to bend to.
Bring your existing Slurm workflow
06 · Spec at a glance

The hardware, the workflow.

The numbers that matter.

GPU
NVIDIA B200
Blackwell, 180GB HBM3e, 8 TB/s mem BW
Node
HGX × 8
8 × B200 per node, NVLink-connected
Interconnect
High-speed
Multi-node, InfiniBand fabric
Scheduler
Slurm
Standard sbatch / srun / squeue
Storage
Shared FS
High-throughput parallel filesystem
Lead time
Days
From signup to first sbatch
Billing
Per-job
Points drawn while RUNNING
07 · For whom

Teams that train, not just call APIs.

If you're moving real weights, this is for you.

A.
LLM startups
building foundation models
Pre-training, continued pre-training, or distillation of 30B–200B+ models. Need multi-node bandwidth, predictable capacity, and pay-as-you-train economics — not multi-year reserved commits.
B.
Enterprise AI labs
doing serious R&D
Finance, healthcare, defense, gov-adjacent. Sensitive corpora can't leave jurisdiction. Need a real cluster with NDA-friendly contracts — but don't want to procure and run datacenter infrastructure in-house.
C.
ML research groups
past the API era
University labs, national-lab affiliates, independent research orgs. Done waiting for HPC allocations. Run the experiment now, pay for the hours you use, write the paper before the field moves on.
08 · FAQ

Questions, answered.

If yours isn't here, ask when you sign up.

Is this just a GPU VM with extra steps? +
No. With a GPU VM, you rent the server and pay for uptime whether you're computing or not. Here, you submit a job and pay only while GPUs are allocated to that job. The cluster (and the multi-node interconnect that makes large-scale training possible) is shared and already built — you don't manage it.
What's the actual lead time from signup to first job? +
For early-access customers: typically same-day to a few days, depending on the size of your account onboarding. We send login credentials, partition info, and onboarding docs once your account is set up.
How exactly does billing work? +
Top up points upfront. When a Slurm job enters the RUNNING state — meaning GPUs are bound to it — points draw down by the GPU-hour. Job in PENDING: no charge. Job COMPLETED or CANCELLED: no charge. Storage is a separate low flat rate per GB-month.
Can I run multi-node jobs spanning dozens of B200s? +
Yes. That's the design point. Nodes are connected over a high-speed interconnect suitable for collective operations at scale. Realistic for 70B-class pre-training, distillation across tens of GPUs, and large-batch fine-tuning.
What frameworks are supported? +
Anything standard HPC. PyTorch, DeepSpeed, Megatron-LM, NeMo, JAX, custom CUDA — bring your stack. We provide the scheduler (Slurm), the hardware, the network, and a shared filesystem. The user-space environment is yours.
Can I do interactive development too, or only batch? +
Both. The login node is for editing, building, light testing. For interactive GPU sessions (Jupyter, debugging, profiling), use srun --pty to grab a slice for a short walltime — same billing model: charged only while the interactive job is RUNNING.
Where is the data physically stored? +
Commercial datacenters in Japan. Data does not leave the jurisdiction unless you explicitly export it. NDAs and custom data-handling agreements are available for sensitive workloads.
Pricing · Early Access

Indicative pricing.

Early-access, pay-per-job pricing for a ready-to-use B200 cluster. The figure below is indicative and will be confirmed at launch — not yet bookable. Join the waitlist for confirmed pricing and availability.

GPU Config Billing Price / GPU-hr Region Availability
NVIDIA B200 SXM 1x GPU · 180GB HBM3e · 28 vCPU · 512GB RAM · NVLink On-demand / per-job $7.42 /GPU-hr · indicative, ex-tax Japan (Tokyo) Waitlist

Per-job billing (per-minute, charged while RUNNING) · minimum initial top-up ¥100,000 (~$620, first load only, usable as credit) · storage billed separately, ~$0.12/GB-month · ex-tax, JPY-based, FX-dependent · subject to change at launch

Early access

Save your spot.

We're building on-demand GPU clusters (B200-class) for fine-tuning, distillation, and distributed training. Leave your work email and we'll let you know when early access opens. Early signups go first — and we'll share any early-bird offers as they're confirmed. No commitment.

First in line Early-bird offers No commitment

DISCLAIMER

● Compute Cluster is a pre-launch reservation program. Capacity, point pricing, and exact launch timing are subject to confirmation.

● "High-speed interconnect" refers to InfiniBand fabric used for multi-node GPU-to-GPU communication.

● Lead times shown are typical for early-access customers and may vary with reservation size.

● No service is live at the time of this page. Signup reserves your spot for early access at launch.