PHOTONSTATE
FREE ASSESSMENT →
EN/ไทย

HOME/PRODUCTS/AI Rack

A private AI rack. Delivered, owned, in Thailand.

AMD Instinct™ GPU servers, Edgecore 800G fabric, open software — sold, installed and supported on your premises by Photon State. You own it.

REQUEST A PILOT →RACK CONFIGURATIONS
  1. 14× AGS-class server · 8U
  2. 2LEAF PAIR · 800G
  3. 3NVMe STORAGE
  4. 4DATA BOUNDARY — SOVEREIGN

THE RACK, ITEMISED

SOLD · INSTALLED · SUPPORTED
AGS86008× AMD Instinct™ MI325X · 8U ↗The compute — up to 5 per 42U rack (8U each)
AIS800-64O64× 800G OSFP · Tomahawk 5 · 51.2 Tbps ↗The fabric — rail-optimised, RoCEv2, SONiC
ECS4650-54T48× 1G · 6× 25G uplinks ↗Out-of-band control that survives the fabric
NVMe storage classNVMe parallel filesystem class — sized at design ↗Feeds the cluster at training speed
OSFP800 optics & DAC800G OSFP SR8 / DAC — pre-qualified line ↗Every link from Edgecore's qualified list

Built on AMD Instinct™ accelerators, delivered as a rack.

AMD Instinct™ MI325X accelerator module — image courtesy AMDAMD INSTINCT™MI325XSHIPPING · AGS8600AMD Instinct™ MI355X accelerator module — image courtesy AMD (representative)AMD INSTINCT™MI355XNEXT · CONFIRM AT BOMAMD Instinct™ MI300X accelerator module — image courtesy AMDAMD INSTINCT™MI300XAVAILABLE · ON REQUEST

CONFIGURATIONS

NO PRICES HERE — THE BOM CONFIRMS EVERY NUMBER
PILOT · 1 SERVER8 GPUs
  • 2 TB HBM3E (8 × 256 GB, AMD-published)
  • 400G leaf pair
  • Air / RDHx
  • Part of one rack
  • Lead time: Band S
OPEN IN DESIGN STUDIO →
TEAM · 4 SERVERS32 GPUs
  • 8 TB HBM3E (32 × 256 GB, AMD-published)
  • 400/800G leaf-spine
  • RDHx / DLC
  • One rack
  • Lead time: Band M
OPEN IN DESIGN STUDIO →
CLUSTER · 16+ SERVERS128 GPUs
  • 32 TB HBM3E (128 × 256 GB, AMD-published)
  • 800G spine, rail-optimised
  • DLC / immersion class
  • Multi-rack pod
  • Lead time: Band L
OPEN IN DESIGN STUDIO →

WHAT RUNS ON IT

ROCm™, PyTorch, vLLM and Kubernetes / Slurm options on the servers — SONiC on the fabric. An open stack end to end, no proprietary lock-in. (Product names are trademarks of their respective owners.)

ROCm™PyTorchvLLMKubernetes / SlurmSONiC

DATA STAYS IN THAILAND

✦ Installed on your premises — on-premise✦ Air-gap capable — isolation verified at acceptance✦ You own the hardware, the data and the models

PLATFORM PROOF

TOP500 · SAKURAONE

SAKURA Internet built its 800-GPU SAKURAONE AI cluster on Edgecore AIS800-64O 800GbE switches running SONiC, arranged as a rail-optimized leaf-spine fabric with RoCEv2 in place of InfiniBand. It debuted at #49 on the June 2025 Top500 — per SAKURA's own researchers, the only system in the top 100 on a fully open networking stack — and stands at #69 on the June 2026 list.

EDGECORE DEPLOYMENT · SOURCE ↗
IMMERSION TRIAL

Edgecore contributed its ORv3 immersion-capable switch to a single-phase immersion cooling proof of concept at Stellium Data Centers in Newcastle, built with Submer tanks and Circle B server integration under OCP principles. The deployment is a PoC — the first ORv3 immersion showcase in Europe — not a stated production rollout.

EDGECORE DEPLOYMENT · SOURCE ↗

DELIVERY & SUPPORT

  1. 01Site survey
  2. 02Power & cooling check
  3. 03Install
  4. 04Burn-in
  5. 05Handover + as-built docs
  6. 06SLA tiers

QUESTIONS BUYERS ASK

Owned vs cloud — what changes?
You buy the rack: hardware, data and models are yours. No hourly meter, no data leaving your fence.
What power does it need?
Per configuration: ~10 kW class per server (ESTIMATE — the AGS8600 page publishes a 4+2×3300 W PSU config, not system draw). Binding numbers live in the engineered design.
How is it cooled?
Air/RDHx at low loads; DLC or immersion class past ~40 kW/rack. The design studio computes the class; the BOM confirms it.
Warranty and spares?
Manufacturer warranty passed through per SKU; SLA tiers include on-site spares options.
Who supports it after handover?
Photon State, first line, from Bangkok — with the Edgecore and AMD escalation paths written into the support contract.
REQUEST A PILOT →