Nova 8 Training

Training or inference?

Two jobs. Two servers. One loop.

Cooled by Nexalus

Training is learning: the model reads huge amounts of data and adjusts billions of weights, so memory and bandwidth matter most. Inference is answering: the model serves users, so tokens per second and per watt matter most.

Serving models to users and apps

Nova 8 Inference

8 × RTX PRO 6000 Blackwell

Up to 55,000 tokens / s measured
8.6 tokens per watt
768 GB GDDR7
Chatbots, copilots, RAG, agents, batch inference

Nova 8 Inference

Training, fine-tuning, the largest models

8 × H200 NVL

1,128 GB HBM3e, 4.8 TB/s per GPU
NVLink-bridged GPUs
Fine-tuning on private data, domain models
Serving very large models and long contexts

Join early access

	Nova 8 Inference	Nova 8 Training
GPU	NVIDIA RTX PRO 6000 Blackwell	NVIDIA H200 NVL
Memory per GPU	96 GB GDDR7	141 GB HBM3e
Memory per server	768 GB	1,128 GB
Bandwidth per GPU	1.8 TB/s	4.8 TB/s
GPU interconnect	PCIe Gen5	NVLink bridges, up to 900 GB/s
Best at	Tokens per watt and per box	Training, fine-tuning, very large models
Status	Taking pre-orders	In development · early access

Nova 8 Training figures are targets based on NVIDIA's published H200 NVL specifications.4

What you can do with it

Your models, trained on your data, in your building.

Cooled by Nexalus

Fine-tune on private data

Adapt open models to your documents, code and customers without sending data to a cloud.

Train domain models

Continued pre-training for legal, medical, financial or engineering language.

Serve the largest models

1.1 TB of HBM3e holds very large open models and long contexts on one server.

Embeddings and RAG at scale

Re-index millions of documents quickly so retrieval stays fresh.

Why liquid for training

Training runs for days. The cooling has to hold.

Cooled by Nexalus

A training job keeps every GPU at full power for days or weeks. Air-cooled servers throttle as the hall warms up; a liquid loop holds the chips at a steady temperature, so the last epoch runs as fast as the first.

Same Nexalus loop proven on Nova 8 Inference: one hour at full load, no card above 80 °C
Redundant pumps and fans: a single failure doesn't stop the job
Steadier temperatures mean less thermal cycling wear on the hardware
Up to 60 °C water out for heat reuse5

Target specification

Nova 8 Training

Cooled by Nexalus

GPUs	8 × NVIDIA H200 NVL, 141 GB HBM3e each, up to 600 W
GPU memory	1,128 GB total · 4.8 TB/s per GPU
GPU interconnect	NVLink bridges (2- or 4-way), up to 900 GB/s GPU-to-GPU · PCIe Gen5 to host
CPU & memory	AMD EPYC · large DDR5 capacity for data staging
Networking	High-speed NICs for multi-node training and storage (configurable)
Cooling	Nexalus direct liquid cooling on every GPU and the CPU · redundant pumps and fans · optional rear quick-connects for heat export
Form factor	Rack-mount, standard 19-inch
Status	In development · early-access list open

Target specification for a product in development. Specification, availability and appearance will change.4

Nova 8 Training.
Memory first.

Two jobs. Two servers. One loop.

Nova 8 Inference

Nova 8 Training

Your models, trained on your data, in your building.

Fine-tune on private data

Train domain models

Serve the largest models

Embeddings and RAG at scale

Training runs for days. The cooling has to hold.

Nova 8 Training

Be first in line for Nova 8 Training.