Kimi K3
Agentic-class- Open-weight
- To be self-hosted on our own GPU infrastructure
- Published prices
Planned · production inference
Navid · Industrial AI CloudUnder development · Dammam
Sovereign AI, manufactured in Saudi Arabia. Built to serve the world.
Navid is building a 15 MW AI Factory in Dammam that will turn Saudi energy into intelligence for the Kingdom, the GCC and customers worldwide. Phase one is under construction; capacity is being allocated ahead of it.
- 15 MWTotal facility power
- 6,000 m²Dammam 2nd Industrial City
- 100%Liquid-cooled by design
- 7,488 GPUsBlackwell Ultra · design capacity
Inside the hall, as designed
GPUNVIDIA Blackwell Ultra · 288GB HBM3e per GPU
NetworkNVIDIA Quantum-2 InfiniBand
Node8x GPU per Node · NVLink Switch System
Design Capacity≈104 GB300 NVL72 racks · 7,488 GPUs
CoolingLiquid Cooled · 120–200kW per rack
Inside the hall, as designedNVIDIA DGX B300 Reference ArchitectureSelect a marker for specifications
Scroll to enter
Our Thesis
Energy becomes intelligence here.
Every industrial economy now runs on two inputs: energy and intelligence. Saudi Arabia has the first in abundance. Navid is building the machine that will convert it into the second — anchored in the Kingdom, engineered to serve the region and the world.
That machine is an AI Factory. We are building it in Dammam, and we will open it to everyone who needs what it makes.
The Facility
Engineered for the physics of frontier compute.
15 MW across 6,000 m² of industrial land in MODON's Dammam 2nd Industrial City. Liquid-cooled from day one, engineered in partnership with ProteCooling — Mulhim & Partners Engineering Company.
Power train, cooling loop and rack density are designed as one system around the accelerators they carry. That is what will let frontier silicon run at its ceiling, every hour of the year.
Key specifications
NVIDIA DGX B300 Reference Architecture · NEO Cloud AI Factory · NVIDIA Preferred Partner
| IT Load Capacity | 12.5 MW |
|---|---|
| Total Facility Power | 15 MW | 2.5 MW Cooling & Overhead |
| Power Infrastructure | 33kV Dual Feed | N+1 Redundancy |
| Cooling System | Direct Liquid Cooling (DLC), NVIDIA Reference Design |
| PUE Target | ≤ 1.20 |
| Rack Density | 120 kW per Rack (GB300 NVL72) | 200 kW per Rack (Vera Rubin NVL144) |
| Total POD Capacity | ≈104 × GB300 NVL72 = 7,488 Blackwell Ultra GPUs at 12.5 MW |
| Network | NVIDIA Quantum-2 InfiniBand |
| Storage | High Performance Parallel File System |
| Security | Physical + Cybersecurity by Design |
Design specifications for phase one, currently under construction.
The Cooling Loop
One closed loop, from cold plate to desert air.
Every watt the accelerators draw leaves the building as heat. This is the path it is designed to take: captured at the die, handed across at the CDU, rejected to Dammam's ambient air — and back again, without evaporating water.
IT Load
Captured at the source
Direct liquid cooling cold plates sit on the GPUs and CPUs, taking the full 12.5 MW IT load as heat before it ever reaches the air.
Secondary Loop · TCS
Hot coolant to the CDUs
Warm coolant leaves the racks at roughly 45 °C and runs to the Coolant Distribution Units.
CDU Heat Exchange
Pumps · filtration · flow control
The exchanger isolates the rack loop from the facility loop, so nothing that circulates through a GPU ever touches building water.
Primary Loop · FWS
Facility water out, and back
On the facility side, warm water carries the load out to heat rejection — and returns cool to the exchanger.
Heat Rejection
Dry coolers to ambient air
Dry coolers dissipate the full thermal load to ambient air, optimised for the Dammam desert climate. Dry, not evaporative.
The loop closes. Cool facility water returns to the exchanger, and coolant goes back to the racks at roughly 35 °C. Nothing is drawn in and nothing is thrown away.
- N+1 RedundancyZero single point of failure across pumps and heat rejection.
- Closed-Loop Liquid CoolingThe same coolant circulates; the rack side never meets the facility side.
- WUE ~ 0 L/kWhWater Usage Effectiveness of about zero litres per kilowatt-hour.
Cooling power budget — 2.5 MW
Running the loop is what the other 2.5 MW buys. It is spent across four places:
- CDU pumps
- Dry cooler fans
- CRAH units
- Controls
The budget also covers a 5% air-cooled ancillary load.
And that is the whole of PUE
- 15 MWTotal facility power
- 12.5 MWIT load
- 1.20PUE target ≤ 1.20
The Compute
Two rack platforms of NVIDIA silicon, each matched to its work.
| Platform | Per rack | Deployable at 12.5 MW IT load |
|---|---|---|
| NVIDIA GB300 NVL72 | 72 Blackwell Ultra GPUs and 36 Grace CPUs, at 120 kW per rack. | ≈104 racks — 7,488 Blackwell Ultra GPUs. Rack-scale and NVLink-connected: designed for domain-adaptive training and high-throughput inference for the largest open-weight models. |
| NVIDIA Vera Rubin NVL144 | 144 Rubin GPUs and 36 Vera CPUs, at 200 kW per rack. | ≈62 racks — 8,928 Rubin GPUs. NVIDIA's next generation, from 2026 — the future-ready path as the facility scales. |
Node specification — NVIDIA HGX B300. Eight B300 SXM GPUs, 288 GB HBM3e per GPU, 2.3 TB HBM3e per system, 5th-Gen NVLink.
Heterogeneous by design. When capacity comes online, every workload lands on the silicon built for it — and you pay for exactly that.
Power & Platforms
Fifteen megawatts in. Twelve and a half to the accelerators.
A facility is only as honest as its arithmetic. Here is where the power is budgeted, and what it is allowed to carry.
The power split — total facility power
15 MW
- 12.5 MW IT load — at a PUE target of 1.20
- 2.5 MW cooling & overhead — everything that keeps the loop turning
- 15 MWTotal facility power
- 1.20PUE
- 12.5 MWIT load
Two rack platforms for the same 12.5 MW
The IT load is a budget, not a rack count. Spend it at 120 kW a rack or at 200 kW a rack and you get two different floors.
-
NVIDIA GB300 NVL72
Rack platform- GPUs per rack
- 72 × Blackwell Ultra
- CPUs per rack
- 36 × Grace
- Rack density
- 120 kW
~104 racks → 7,488 GPUs Deployable capacity at 12.5 MW IT load: roughly 104 GB300 NVL72 racks, or 7,488 Blackwell Ultra GPUs.
-
NVIDIA Vera Rubin NVL144
Alternative · future-ready, 2026- GPUs per rack
- 144 × Rubin
- CPUs per rack
- 36 × Vera
- Rack density
- 200 kW
~62 racks → 8,928 GPUs The same power budget spent denser: roughly 62 Vera Rubin NVL144 racks, or 8,928 Rubin GPUs.
Node specification — NVIDIA HGX B300
- 8 × B300 SXM GPUs per node
- 288 GB HBM3e per GPU
- 2.3 TB HBM3e per system
- 5th-Gen NVLink
- 33kV Dual Utility FeedTwo independent paths onto the grid.
- N+1 RedundancySpare capacity in the power train, by design.
- Direct Liquid CoolingCold plates on the silicon, not air on the room.
The Economics
The floor is designed to split inference in two.
Prefill is compute-bound. Decode is memory-bound. We will run them on separate GPU pools, each scaled to its own curve, so every request meets hardware matched to the work it is doing.
NVIDIA reports that disaggregated serving can raise throughput by up to 15× without sacrificing latency. We designed the floor for it from the first rack.
It arrives as the one number that decides economics at scale: tokens per megawatt. That is what will let us serve frontier models at the labs' published rates while owning the metal underneath.
Disaggregated serving
as designed · one request, two pools
Request
prompt in
Prefill pool
compute-bound
Reads the whole prompt at once. Scaled on the compute curve.
Decode pool
memory-bound
Emits one token at a time. Scaled on the memory curve.
Tokens
stream out
up to 15×
throughput from disaggregated serving, as NVIDIA reports — without sacrificing latency
tokens / MW
the one number that decides economics at scale
The Token Factory
The factory that makes Navid a token manufacturer.
The floor is being built to self-host open-weight frontier models on our own GPU infrastructure and serve them as production inference — including agentic-class models such as Kimi K3 and GLM-5.2.
Navid serves models today on the Yehia playground. Dammam is what moves that onto our own metal, at industrial scale.
GLM-5.2
Agentic-class- Open-weight
- To be self-hosted on our own GPU infrastructure
- Published prices
Planned · production inference
Frontier models, published prices, sovereign ground — when phase one comes online it will open to teams in the Kingdom, across the GCC and internationally, on infrastructure you are entitled to inspect.
Arabic-First
Models that think in Arabic.
Fourteen centuries of written heritage, carried forward into models that reason natively in the language.
Our deepest investment is Arabic-native AI. The factory is the sovereign compute layer being built for training, adapting and serving Arabic models in agentic use across government, industry and enterprise. The Kingdom leads the region in Arabic model development. We are building the factory floor beneath it.
نماذج تُفكّر بالعربية
Models that think in Arabic.Fourteen centuriesCarried forward
Sovereign by Design
Three commitments, written into the architecture.
01
Jurisdiction
Anchored in Saudi Arabia and built to the Kingdom's regulatory standard — the strictest bar our customers ask us to clear.
02
Residency
Your prompts, documents and model weights will stay where you place them, with in-Kingdom residency for workloads that require it.
03
Isolation
Multi-tenant NeoCloud with tenant-level separation — or dedicated capacity where regulation requires it.
NVIDIA NEO Cloud Stack
One platform, from facility to application.
- AI Applications Training | Inference | HPC | Digital Twins
- NEO Cloud Services Scheduling | Orchestration | Monitoring | IAM
- AI Infrastructure DGX B300 | Networking | Storage | Security
- Cloud Foundation Compute | Network | Storage | Facility
- Sovereign Cloud In the Kingdom of Saudi Arabia
Sustainable by Design
Efficiency engineered in, not bolted on.
High Efficiency
PUE target ≤ 1.20
Liquid Cooling
Reduced Water Usage
Renewable Ready
Designed for Clean Energy Integration
Carbon Conscious
Supporting KSA Vision 2030
Reach
Rising in Dammam. Built to serve three continents.
Capacity will be open wherever the demand is: national entities and industrial operators in the Kingdom, enterprises across the GCC, and international teams that need frontier inference close to their users.
Dammam sits between Europe, Africa and Asia — one of the shortest network paths to a large share of the world's inference demand, on a grid with the power to feed it.
Built for Industrial AI
Where AI will run the plant.
It is being built for the operators of refineries, ports, utilities, factories and national platforms — where a model is a control loop with consequences.
Computer vision
Predictive maintenance
Digital twins
Autonomous systems
Agentic workflows
What Comes Next
15 MW is phase one.
The site, the cooling architecture and the fabric are designed to scale well beyond it. Phase one is under construction, demand is already arriving from the Kingdom, the GCC and abroad, and early capacity is being allocated now.
TAKING AI FORWARD
Experience Products in AI
Empower your organization with Navid's AI-driven solutions, merging human expertise with advanced technology for unparalleled growth and security.
Copyright © 2026 Navid.