Home/Infrastructure
TH-BKK-1 · Bangkok
Everything that sits under the invoice.
Published in full, because the difference between an AI cloud and a reseller is whether they can tell you what is in the rack and who has the keys to it.
5 nodes
HGX B300 chassis
40 × B300
Blackwell Ultra GPUs
10.5 TB
HBM3e fleet-wide
720 PFLOPS
FP4 inference
360 PFLOPS
FP8 training
800 Gb/s
Per InfiniBand port
1.12 TB
System RAM, fleet-wide
153 TB
Local NVMe, fleet-wide
Compute
Blackwell Ultra, in the eight-GPU configuration.
Every node is an NVIDIA HGX B300 baseboard: eight B300 GPUs on a fifth-generation NVLink switch, 2.1 TB of HBM3e in one coherent domain, and 144 PFLOPS of FP4 for inference.
The generation matters most for two things. Attention-layer throughput roughly doubles against Hopper, which is where reasoning and long-context workloads spend their time. And the memory per GPU is finally large enough that the biggest open-weight models stop needing to be cut in half across servers.
We do not oversubscribe. Forty GPUs are installed and forty GPUs are sellable — there is no burst pool that turns into someone else's throttling.
| Node specification | NVIDIA HGX B300, 8-GPU |
|---|---|
| GPU | 8 × NVIDIA B300 (Blackwell Ultra) |
| GPU memory | 2.1 TB HBM3e aggregate |
| NVLink | 14.4 TB/s all-to-all, 5th generation |
| Inference | 144 PFLOPS FP4 |
| Training | 72 PFLOPS FP8 |
| CPU | 2 × 112-core x86, 224 cores total |
| System RAM | 2 TB DDR5 |
| Local NVMe | 30.7 TB Gen5 |
| East-west | 8 × 800 Gb/s InfiniBand (ConnectX-8) |
| North-south | 2 × 100 Gb/s, BGP-capable |
| Management | Dedicated OOB, IPMI/Redfish, hardware console |
Network
Two networks, deliberately separate.
The east-west fabric exists so forty GPUs can behave as one machine during a training run — non-blocking InfiniBand with collective offload, never touching the internet path.
The north-south side is built for one thing: being close to Thai users. Peered at BKNIX with direct sessions to the major domestic ISPs, which is the whole reason a Bangkok region beats a Singapore one for anything interactive.
| East-west fabric | NVIDIA Quantum-X800 InfiniBand, non-blocking across all five nodes |
|---|---|
| Per-node NIC | 8 × ConnectX-8 SuperNIC, up to 800 Gb/s per port |
| In-network compute | SHARP v4 collective offload for multi-node training |
| North-south | 2 × 100 Gb/s per node, BGP and BYOIP supported |
| Peering | BKNIX, plus direct sessions with major Thai ISPs |
| Domestic latency | 2–6 ms round trip from Thai broadband and mobile |
| Regional latency | ~30 ms Singapore · ~45 ms Hong Kong · ~65 ms Tokyo |
| DDoS | Always-on volumetric scrubbing at the edge |
| Private access | IPsec, WireGuard or dedicated cross-connect |
Storage
Nothing leaves the region, including backups.
Object storage is often where residency claims quietly fail — a durability feature replicates a bucket to another country and nobody notices until an audit. Our object tier has exactly one region, so cross-border replication is not a setting that can be left on by mistake.
The full residency position
| Local scratch | 30.7 TB Gen5 NVMe per node, wiped on release |
|---|---|
| Parallel filesystem | NVMe-backed, ~400 GB/s aggregate read, POSIX and RDMA |
| Object storage | S3-compatible, erasure coded, Bangkok only, no cross-region replication |
| Snapshots | Point-in-time, retained to your policy |
| Encryption | AES-256 at rest, customer-managed keys available |
| Erasure | NIST 800-88 purge with certificate on contract exit |
Facility
130 kW racks are not an ordinary data centre.
An HGX B300 node draws roughly 14 kW. Most colocation in Thailand was built for 5–8 kW racks and cannot take one of these, let alone five. Ours sits in direct-to-chip liquid cooling with rear-door heat exchange, in a Tier III facility with dual utility feeds.
This is also the constraint on growth. Adding nodes is a power and cooling conversation before it is a purchasing one, which is why we publish availability rather than promising capacity we have not energised.
| Region | TH-BKK-1 · Bangkok metropolitan area |
|---|---|
| Tier | Tier III design, concurrently maintainable |
| Power feed | Dual utility feed, N+1 UPS, N+1 generator |
| Rack density | Up to 130 kW per rack, liquid-ready |
| Cooling | Direct-to-chip liquid with rear-door heat exchange |
| Fire | VESDA detection, inert gas suppression |
| Physical access | Mantrap, biometric, escorted, CCTV retained 90 days |
| Certifications | ISO/IEC 27001 and ISO 22301 at facility level |
Talk to us
Want to see it?
Prospective customers on a term contract are welcome in the facility, subject to the operator's access process. Bring your infrastructure lead and your security officer.
Direct line sales@thaiaicloud.co.th · Replies in one business day, Thai or English.