Technology Generations

Current on every generation since Hopper

Each generation changes the facility problem before it changes the compute problem. Power per rack, cooling mode and network topology all move, and they move before the hardware arrives. Planning for that shift is the work.

H100Hopper
From 2023
DELIVERED
H200Hopper
From 2024
DELIVERED
B300Blackwell Ultra
From 2025
DELIVERING
RTX PRO 6000Blackwell · Server Edition
From 2025
DELIVERING
“Delivered” and “Delivering” refer to systems specified, procured, integrated and commissioned for enterprise clients. Rack-scale systems are supported where required; the mainline is the two air-cooled classes below. Specifications and configurations are provided under NDA during the requirement phase of an engagement.
Node Classes

Two classes, delivered whole

Most enterprise requirements resolve to one of two node classes. Both are air-cooled, so both are deployable in conventional contained-air halls with no chilled water to the rack.

KNOWLEDGE CORE
Training class
Built for large-model training, fine-tuning at scale and retrieval. Specified, integrated, burned in and commissioned as complete systems, not assembled from parts on site.
SERVING TIER
Inference and visualisation class
Built for inference serving, rendering, visualisation and virtual desktops. Sized to concurrency rather than to model size, and deployable in conventional contained-air halls.
Deployment Shape

One array, two tiers

Most first programmes resolve to the same shape: a knowledge core that trains, fine-tunes and retrieves, and a serving tier that answers in real time. The first-phase array pairs them, entering production as one programme and scaling by tier afterwards.

How it scales
Concurrency grows on the serving tier; model size and retrieval depth grow on the core. The two tiers scale independently, so capacity is added where the workload actually moved, not where the original plan said it would. Both stay inside the contained-air band: scaling is an order, not a facility project.
A planning shape, not a quotation. Per-client sizing follows the requirement phase; see Service 01.
Network & Storage

Four networks, not one

Mixing compute collectives with storage traffic on a shared fabric is the most common cause of a cluster that benchmarks well and trains badly. The reference design keeps them apart.

Compute fabric
Rail-optimised and non-blocking, carrying the collective operations during training. Never shared with anything else.
Storage fabric
Dataset read and checkpoint write, kept off the compute fabric so a bursty checkpoint cannot stall a training step.
In-band management
Orchestration, telemetry and image push, separated from compute traffic.
Out-of-band
A flat, isolated network for consoles and power control: the one that still works when everything else does not.
Three storage tiers
Scratch and checkpoint
The highest-throughput tier. Checkpoint write is bursty and blocks training.
Active dataset
Sustained read, scaled to the aggregate accelerator count.
Cold and archive
Capacity-optimised, and never in the training path.
Final selection follows the workload profile, the chosen platform’s supported fabrics and the facility’s cross-connect provision.