BAYWAW AI Container
A 1.2 MW AI data center, delivered in one container.
A modular AI data center with 624 NVIDIA B300 GPUs, single-phase immersion cooling and a non-blocking network, pre‑integrated in one 40ft High Cube. Built and tested in the UAE, operational on your site in as little as four months.
Key figures
- NVIDIA B300 GPUs
- 624
- Hypertec servers
- 78
- Immersion tanks
- 6
- Peak FP4 per GPU
- 13.5PFLOPS
- Total GPU memory
- 179TB
- Power at PUE 1.05
- 1.2MW
FP4 performance is the manufacturer’s peak rating for a single NVIDIA B300 GPU (“up to”), not a total for the container.
01Compute
NVIDIA B300 compute in a single module.
An NVIDIA B300 data center in one container: 624 GPUs of the Blackwell Ultra generation, installed in 78 Hypertec servers of eight GPUs each and fully immersed in dielectric liquid.
Each GPU carries 288 GB of HBM3e memory, about 179 TB across the container, so large models can stay resident in GPU memory. Every server joins the GPU fabric through eight 400G network cards, one per GPU.
- Container624 GPUs6 tanks · 78 servers
- Tank104 GPUs13 servers per tank
- Server8 × NVIDIA B300Hypertec · 8 × 400G east-west · 2 × 400G north-south
- GPU288 GB HBM3eUp to 13.5 PFLOPS FP4, peak, per GPU
Built to run
Training
Pre-train and continue training large models on a rail-optimized, non-blocking fabric built for collective operations across hundreds of GPUs.
Fine-tuning
Adapt foundation models to your own data on infrastructure you own and operate, in the jurisdiction you choose.
Large-scale inference
Serve models to many concurrent users, with the memory headroom to keep large models and long contexts on the GPU.
02Inside the module
Six immersion tanks around one service aisle.
An immersion-cooled GPU container laid out like a compact data hall: three tanks along each long wall, a central service aisle on raised grating and a dedicated coolant distribution zone.
- Tanks
- 6 × GRC NX, arranged 3 + 3 along the long walls
- Per tank
- 13 servers · 104 GPUs
- Service aisle
- Central, on raised grating, with access to every tank
- Coolant
- Redundant loops and CDUs, heat rejection on the roof
Anatomy of an immersion tank
In single-phase immersion, the dielectric liquid stays liquid through the whole cycle. Heat moves from the components into the liquid, then out to the CDU.
Coolant distribution manifold
Cooled dielectric liquid arrives from the CDU and is distributed evenly to every server in the tank.
Server blades
13 Hypertec servers are immersed in each tank, mounted on liquid-cooled plates in direct contact with the fluid.
Dielectric liquid bath
The non-conductive liquid absorbs heat from every component at once, with no chilled air to move through the room.
Return manifold
Warmed liquid is collected and returned to the CDU, which passes the heat to the roof-mounted heat-rejection units.
Redundant by design
Redundant coolant loops and CDUs keep the liquid moving when one path is taken out of service for maintenance, so cooling never rests on a single loop.
03Network
A non-blocking GPU fabric, validated before it ships.
GPU-to-GPU traffic runs on a dedicated, rail-optimized back-end fabric. A separate front-end network connects the container to the internet, your WAN and other sites, with firewalls, PoE access and out-of-band management built in.
Edge routing
2 × Cisco Catalyst 8500‑20X6C (BGP, HA)
BGP routing with high availability between the container, the internet, your WAN and other sites.
Next-generation firewalls
2 × Cisco Secure Firewall 4245 (NGFW)
A pair of firewalls inspects and controls all traffic entering and leaving the module.
Front-end spine
6 × Cisco Nexus 9364 (64 × 800G)
North-south traffic (data ingest, storage and user access) reaches every server over 2 × 400G links.
GPU back-end fabric
8 leaf + 4 spine per tank · 48 leaf / 24 spine in total
Rail-optimized and non-blocking: each server’s eight network cards connect at 8 × 400G, one rail per GPU, so collective operations run at full bandwidth.
PoE access
6 × Cisco Catalyst 9300X‑48HX (UPOE, mGig)
One switch per tank for cameras, wireless access points and sensors, isolated in its own VRF and dual-homed.
Out-of-band management
2 × Cisco Catalyst 9200L‑48T (HA pair)
A dedicated management network for BMC, switch and PDU access, independent of production traffic.
Validated end‑to‑end before shipment
The complete network is tested in the UAE before the container ships, so it arrives configured and verified instead of becoming a commissioning project on your site.
- AI-driven diagnostics
- Topology mapping
- Configuration compliance checks
- Automated tests
04Container
A standard ISO container, built for the field.
The module is housed in an ISO 40ft High Cube container. It ships by road and sea like any other container and is built to withstand transport and harsh environments.
- Container type
- ISO 40ft High Cube
- Tare weight
- ~4,000 kg
- Max gross weight
- 30,480 kg
- Internal volume
- 76.4 m³
Road and sea
Ships as a standard ISO container through existing freight and port logistics.
Built for harsh environments
Engineered to withstand transport and demanding site conditions.
Built in the UAE
Assembled, integrated and tested in the UAE, then shipped ready to run.
05Specifications
AI Container technical specifications
All figures for one BAYWAW AI Container. Request the datasheet to share the complete document with your engineering team.
Request the datasheet| GPU | NVIDIA B300 (Blackwell Ultra) |
|---|---|
| GPU count | 624 |
| Servers | 78 × Hypertec, 8 GPUs each |
| Memory per GPU | 288 GB HBM3e |
| Total GPU memory | ~179 TB |
| Peak compute | Up to 13.5 PFLOPS FP4 per GPU |
| Method | Single-phase immersion |
|---|---|
| Tanks | 6 × GRC NX, 3 + 3 layout |
| Per tank | 13 servers · 104 GPUs |
| Coolant | Dielectric liquid |
| Redundancy | Redundant coolant loops and CDUs |
| Heat rejection | Roof-mounted units |
| Facility power | 1.2 MW |
|---|---|
| PUE | 1.05 |
| Power modules | 2: one active at 100%, one on standby |
| Redundancy | N+2 |
| Failover | < 10 ms |
| Energy at full load | ~10.5 GWh per year |
| Energy sources | Energy-agnostic, optional solar farm + BESS |
| GPU fabric | Rail-optimized, non-blocking |
|---|---|
| East-west per server | 8 × 400G (8 NICs) |
| North-south per server | 2 × 400G |
| Back-end switches | 48 leaf · 24 spine (8 + 4 per tank) |
| Front-end spine | 6 × Cisco Nexus 9364 (64 × 800G) |
| Edge routing | 2 × Cisco Catalyst 8500‑20X6C (BGP, HA) |
| Firewalls | 2 × Cisco Secure Firewall 4245 (NGFW) |
| PoE access | 6 × Cisco Catalyst 9300X‑48HX (UPOE, mGig) |
| Out-of-band management | 2 × Cisco Catalyst 9200L‑48T (HA pair) |
| Type | ISO 40ft High Cube |
|---|---|
| External (L × W × H) | 12,192 × 2,438 × 2,896 mm |
| External, imperial | 40 ft × 8 ft × 9 ft 6 in |
| Internal (L × W × H) | 12,032 × 2,352 × 2,698 mm |
| Internal, imperial | 39 ft 5 in × 7 ft 8 in × 8 ft 10 in |
| Tare weight | ~4,000 kg |
| Max gross weight | 30,480 kg |
| Internal volume | 76.4 m³ |
| Built, integrated and tested | United Arab Emirates |
|---|---|
| Order to operational | In as little as 4 months |
| Handover | Plug & play, configured and validated |
| Transport | Road and sea, standard ISO |
| Support | 24/7/365, follow‑the‑sun |
| Initial response | < 15 minutes |
| Operations | Managed end-to-end by BAYWAW engineers |
06Plug & play
Delivered ready to run, operated for you.
Every BAYWAW AI Container leaves the UAE pre‑integrated and tested. Our engineers commission it on your site and manage it end-to-end, so your team can focus on AI workloads.
As little as
4months
from order to operational
Pre-integrated, tested and validated
Compute, cooling, power and network are integrated and validated together in the UAE before shipment.
Delivered ready to run
The module arrives as one complete unit. Once set on its foundation and connected to power and network, it is ready for commissioning.
Configured and optimized
All systems are configured and optimized for your workloads before handover.
Managed end-to-end
BAYWAW engineers monitor and maintain the container 24/7/365, with an initial response in under 15 minutes.
You focus on AI
Your team works on models and applications. Hardware lifecycle, security patching and warranty management stay with us.
07FAQ
Questions from technical buyers
Short answers to the questions infrastructure teams ask first. For anything deeper, book a technical briefing with our engineers.
Which GPUs does the BAYWAW AI Container use?
It runs 624 NVIDIA B300 GPUs from the Blackwell Ultra generation, installed in 78 Hypertec servers of eight GPUs each. Every GPU carries 288 GB of HBM3e memory, about 179 TB across the container, and is rated by the manufacturer at up to 13.5 PFLOPS of FP4 compute per GPU.
How does immersion cooling work, and how are the servers maintained?
The servers are submerged in a non-conductive dielectric liquid that absorbs heat directly from the components. Redundant coolant loops and CDUs carry that heat to roof-mounted heat-rejection units, for a PUE of 1.05. Servers are reached from the central service aisle and lifted out of the bath for maintenance, which BAYWAW engineers perform as part of the managed service.
What does our site need to provide?
Essentially a power supply sized for 1.2 MW, a prepared, level foundation for a 40ft container and network connectivity. The container is energy-agnostic: it can run on grid power or on an optional solar farm with battery storage. Our engineers review your site during the proposal phase.
How does it connect to our existing network?
Two Cisco Catalyst 8500 edge routers provide BGP routing with high availability to the internet, your WAN and other sites, behind a pair of Cisco Secure Firewall 4245 next-generation firewalls. The configuration is agreed with your network team and validated end-to-end before shipment.
How long does delivery take?
In as little as four months from order to operational. Each container is built, integrated and tested in the UAE, then shipped by road or sea as a standard ISO container. The exact schedule depends on configuration, site readiness and destination, and is confirmed in your proposal.
Can we scale beyond one container?
Yes. The 1.2 MW module is a building block: additional containers can be deployed side by side or across sites as demand grows, and the edge routing supports multi-site connectivity. We size the number of containers with you from your workloads and the power available.
Next step
Bring 624 NVIDIA B300 GPUs to your site.
Request a proposal sized to your workload and site, or receive the full datasheet for your engineering team.