Arise: AI compute, scaling to 1 GW

Compute

NVIDIA DGX B300.
Live now.

NVIDIA’s Blackwell Ultra system. Rent a whole one, or start from a single GPU.

Dedicated nodes

A whole DGX B300, yours alone.

Eight GPUs on one NVLink domain. No neighbors, no preemption.

Request capacity

On-demand GPUs

Blackwell Ultra, by the GPU.

Start from a single GPU and add more when you need them.

Ask about availability

2.1TBGPU memory per system

144PFLOPSFP4 Tensor Core, sparse. 108 dense.

14.4TB/sNVLink aggregate bandwidth

800Gb/sNetworking per GPU, on ConnectX-8

Full specifications

GPUs: 8× NVIDIA Blackwell Ultra
FP4 Tensor Core: 144 PFLOPS sparse, 108 dense
FP8 Tensor Core: 72 PFLOPS sparse, 36 dense
CPUs: 2× Intel Xeon 6776P, 128 cores
GPU memory: 2.1 TB of HBM3e
Memory bandwidth: 62 TB/s in total
System memory: 2 TB DDR5, configurable to 4 TB
NVLink: Fifth generation, 18 links per GPU
NVLink Switch: 2×, every GPU connected to both
Bandwidth: 1.8 TB/s per GPU, 14.4 TB/s aggregate
Cluster: 8× NVIDIA ConnectX-8 SuperNIC, one per GPU, up to 800 Gb/s InfiniBand or Ethernet
Storage and management: 2× NVIDIA BlueField-3 DPU, dual-port, up to 400 Gb/s
Host and BMC: 1 GbE management ports
Data: 8× 3.84 TB E1.S NVMe, 30 TB
Boot: 2× 1.92 TB M.2 NVMe, mirrored
Power: About 14 kW per system
Cooling: Air-cooled
Form factor: 10U, 442 × 483 × 904 mm
Tenancy: Dedicated nodes: one tenant, no preemption
Region: us-west-1

Per system, as NVIDIA specifies it.

Inside a DGX B300

Eight Blackwell Ultra GPUs.

One GPU tray: two rows of four, two NVLink switches between the rows, and a ConnectX-8 for every GPU.

2.1 TB of GPU memory.

HBM3e across eight GPUs: about 1.5 times a DGX B200.

One NVLink domain.

Every GPU links to both switches at 1.8 TB/s, 14.4 TB/s in aggregate. GPU-to-GPU traffic stays on NVLink.

Models

The whole model.
One node.

Pick a model and see how it fits on one DGX B300.

One DGX B3008 GPUs · 262 GB each1 bar = 12.5 GB

Weights 1.56 TBKimi K3 · Moonshot AI · across 8 GPUs

About 0.5 TB leftKV cache and runtime

Total parameters2.8T

Active per token104B

PrecisionMXFP4

Checkpoint1,561 GB

2.8 trillion parameters, whole, on one node. Too big for a DGX B200.

Checkpoint sizes as published on Hugging Face.

Platform

The platform around the GPUs.

Dev machines, batch jobs and monitoring, from the CLI, the console, or your agent.

Monitoring

Every GPU. Every second.

Utilization, memory, power and temperature for every GPU.

Utilization91%

GPU memory1.89 / 2.1 TB

Power12.6 kW

Temperature64°C

GPU 0 GPUs 1 to 7 serve-1 · kimi-k3 Sample data

Dev machines

A whole node that feels local.

Eight GPUs in VS Code, Cursor, or Jupyter.

Batch jobs

Submit it. Sleep on it.

Queue jobs on your reservation. Each one starts as soon as its GPUs are free.

gemma-4-finetune4 GPUs
gpt-oss-evals2 GPUs
v4-pro-batchQueued, 4 GPUs

train-1 Sample data

Fabric

Eight GPUs.
One NVLink domain.

1.8 TB/s per GPU over NVLink inside the node, and up to 800 Gb/s per GPU between nodes.

Agent setup

Skip the setup.
Keep the sign-off.

Paste your Arise link into your agent and say what you need. It plans the capacity, you confirm, it builds.

Arise link issued with your account Preview

Agent session Demo

Your agent, one link.Copy the link above, then paste it here.

In this demo

YouPaste the link, ask for Kimi K3 on a dedicated node, and sign in.
AgentChecks availability and plans the capacity.
YouConfirm the reservation.
AgentReserves the machine and deploys the model.

Message your agent. Paste the Arise link here.

Your reservation

DGX B300 · us-west-1

Waiting to start

Reserved
Container
Weights
Engine
Healthy

Elapsed0.0 s

Confirmations0

Your GPUs0 / 8

Copy the link above to start the demo. Link ready. Paste it to the agent. Time-lapse of an example run. One sign-in and one confirmation. Demo ended at the confirmation.

Console

The full console.
Whenever you want it.

Everything your agent sets up shows up here too.

Arise consoleyour-team / overview Sample data

Workspace your-team

Region us-west-1

Overview

your-team · us-west-1

GPUs in use18/ 24 reserved

Endpoints1serving

Jobs2running · 1 queued

GPU-hours this week1,284

Nodes

3 nodes · dgx-b300 · 8 GPUs each

NodeRunningGPU utilizationGPU memoryHealth
serve-1Created by agentDedicated node

Endpoint kimi-k3One engine on 8 GPUs · serving for 3d 2h

8 of 8 busy · 91%

GPU memory 1.89 / 2.1 TB

Healthy
dev-1Dev machine

VS Code · 2 sessionsStarted 6h 12m ago from the CLI

4 of 8 busy · 46%

GPU memory 0.71 / 2.1 TB

Healthy
train-1Dedicated node

gemma-4-finetune, gpt-oss-evals2 jobs running · 1 queued

6 of 8 busy · 72%

GPU memory 0.98 / 2.1 TB

Healthy

GPU utilization by node

Activity

Agent, CLI and console

2 min agoAgentQueued job v4-pro-batch, 4 GPUs
22 min agoCLISubmitted gpt-oss-evals
5h 41m agoCLIStarted gemma-4-finetune
6h 12m agoCLIStarted dev-1
3d 2h agoAgentReserved serve-1, deployed kimi-k3Confirmed by you

Endpoints

1 serving · private to your-team

Endpoint	Model	Runs on	Status	Auth	Created by
kimi-k3	Kimi K3, one engine	serve-1, 8 GPUs	Serving	Key required	Agent

kimi-k3Serving for 3d 2h

Requests96/ min

In flight22

GPUs healthy8/ 8

Errors (1h)0.02%

Call it

curl

curl $ARISE_ENDPOINT/v1/chat/completions \
  -H "Authorization: Bearer $ARISE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Access: Private to your-team
Auth: API key required. Keys are under API keys.
Model name: kimi-k3

GPU memory

One engine across serve-1

GPU 0236 / 262 GBHealthy
GPU 1236 / 262 GBHealthy
GPU 2236 / 262 GBHealthy
GPU 3236 / 262 GBHealthy
GPU 4236 / 262 GBHealthy
GPU 5236 / 262 GBHealthy
GPU 6236 / 262 GBHealthy
GPU 7236 / 262 GBHealthy

Jobs

2 running · 1 queued · 1 completed

Job	Status	GPUs	Runs on	Progress	Runtime	Submitted by
gemma-4-finetune	Running	4	train-1	62% · step 18,600 / 30,000	5h 41m	CLI
gpt-oss-evals	Running	2	train-1	31%	22m	CLI
v4-pro-batch	Queued	4	train-1	Waiting for 4 GPUs · queued 2 min ago	Not started	Agent
gemma-4-evals	Completed	2	train-1	100%	47m	Console

Logsgemma-4-finetune

13:52:42  step 17900/30000  loss 1.4291  lr 7.01e-06  1.10 s/step
13:54:32  step 18000/30000  loss 1.4263  lr 6.91e-06  1.10 s/step
13:54:34  checkpoint saved  /vol/gemma-4-finetune/step-18000
13:56:22  step 18100/30000  loss 1.4270  lr 6.81e-06  1.10 s/step
13:58:12  step 18200/30000  loss 1.4222  lr 6.71e-06  1.10 s/step
14:00:02  step 18300/30000  loss 1.4236  lr 6.61e-06  1.10 s/step
14:01:52  step 18400/30000  loss 1.4189  lr 6.51e-06  1.10 s/step
14:03:42  step 18500/30000  loss 1.4174  lr 6.42e-06  1.10 s/step
14:03:44  checkpoint saved  /vol/gemma-4-finetune/step-18500
14:05:32  step 18600/30000  loss 1.4151  lr 6.32e-06  1.10 s/step

GPUs on train-1

6 of 8 in use

gemma-4-finetune4 GPUs
gpt-oss-evals2 GPUs
Free2 GPUs

Next in queue: v4-pro-batch needs 4 GPUs and starts when they are free.

Open Endpoints to see the API call, or Jobs to see the queue.

Security and compliance

Built for your security review.
Bring your questionnaire.

Our security program is built around these standards. Ask for the security package: policies, architecture and controls, under NDA.

SOC 2 Type II
Controls for security, availability and confidentiality.
ISO/IEC 27001
Information security, run as a managed system.
ISO/IEC 27017 and 27018
Security in the cloud, and the personal data held there.
ISO/IEC 42001
Responsible management of AI systems.
GDPR
A data processing agreement and EU transfer terms.
HIPAA
A business associate agreement for health data.

Audits and certifications: PCI DSSISO/IEC 27701ISO 22301HITRUST CSFBSI C5TISAXCyber EssentialsFedRAMPCMMC 2.0
Laws and frameworks: NIST CSF 2.0CIS Controls v8.1NIST AI RMFEU AI ActNIS 2DORAUS state privacy lawsEssential EightNIST SP 800-53NIST SP 800-171CJIS Security Policy

Request the security package

Roadmap

DGX B300 is the start.
1 GW is the goal.

Customers on DGX B300 get first call on everything we bring online next.

Live nowNVIDIA DGX B300
Eight Blackwell Ultra GPUs per system.
NextNVIDIA GB300 NVL72
72 Blackwell Ultra GPUs per liquid-cooled rack.
NextNVIDIA Vera Rubin
The platform after Blackwell.
Late 20271 GW
Power for hundreds of thousands of GPUs.

Priority allocation

Existing customers see new capacity before it is listed anywhere else.

One platform as you grow

The same console, CLI and volumes as you go from one node to many.

A direct line

You talk to the people who run the hardware.

Reserve your DGX B300.
Available today.

Request capacity contact@ariselabs.ai

Scaling to 1 GW.

NVIDIA DGX B300.
Live now.

A whole DGX B300, yours alone.

Blackwell Ultra, by the GPU.

Compute

Memory

Interconnect

Networking

Storage

Power and physical

On Arise

Eight Blackwell Ultra GPUs.

2.1 TB of GPU memory.

One NVLink domain.

The whole model.
One node.

The platform around the GPUs.

Every GPU. Every second.

A whole node that feels local.

Submit it. Sleep on it.

Eight GPUs.
One NVLink domain.

Skip the setup.
Keep the sign-off.

Your reservation

The full console.
Whenever you want it.

Overview

Nodes

GPU utilization by node

Activity

Endpoints

Call it

GPU memory

Jobs

Logsgemma-4-finetune

GPUs on train-1

Built for your security review.
Bring your questionnaire.

DGX B300 is the start.
1 GW is the goal.

Reserve your DGX B300.
Available today.

NVIDIA DGX B300.Live now.

A whole DGX B300, yours alone.

Blackwell Ultra, by the GPU.

Compute

Memory

Interconnect

Networking

Storage

Power and physical

On Arise

Eight Blackwell Ultra GPUs.

2.1 TB of GPU memory.

One NVLink domain.

The whole model.One node.

The platform around the GPUs.

Every GPU. Every second.

A whole node that feels local.

Submit it. Sleep on it.

Eight GPUs.One NVLink domain.

Skip the setup.Keep the sign-off.

Your reservation

The full console.Whenever you want it.

Overview

Nodes

GPU utilization by node

Activity

Endpoints

Call it

GPU memory

Jobs

Logsgemma-4-finetune

GPUs on train-1

Built for your security review.Bring your questionnaire.

DGX B300 is the start.1 GW is the goal.

Reserve your DGX B300.Available today.

NVIDIA DGX B300.
Live now.

The whole model.
One node.

Eight GPUs.
One NVLink domain.

Skip the setup.
Keep the sign-off.

The full console.
Whenever you want it.

Built for your security review.
Bring your questionnaire.

DGX B300 is the start.
1 GW is the goal.

Reserve your DGX B300.
Available today.