Executive summary: Colossus 2 is xAI's expanded AI training supercluster located in Memphis, Tennessee. Scaling from the initial 100,000 Nvidia H100 deployment to over 200,000 GPUs incorporating next-generation Nvidia Blackwell B200 silicon, Colossus 2 represents one of the largest concentrated compute clusters on Earth, powered by multi-hundred-megawatt utility feeds, direct-to-chip liquid cooling manifolds, and single-hop optical InfiniBand fabrics.
Building frontier artificial intelligence models requires supercomputing infrastructure that challenges the physical limits of municipal power grids, thermal dissipation, and high-frequency network fabrics. When xAI stood up the original Colossus cluster in Memphis in under 122 days, it established a rapid datacenter construction record.
Now, with Colossus 2, xAI is expanding the facility to over 200,000 high-performance accelerators dedicated to training Grok 5. Below is an engineering teardown of the electrical infrastructure, liquid cooling mechanics, and network topology powering xAI's flagship compute engine.
Compute Fabric: Blackwell B200 and H100 Heterogeneous Clusters
Training multimodal foundation models requires massive parameter parallelization. Colossus 2 pairs the initial 100k H100 nodes with incoming Nvidia Blackwell B200 NVL72 rack architectures:
| Specification | Colossus 1 Baseline | Colossus 2 Expansion Phase |
|---|---|---|
| GPU Accelerator Count | 100,000 Nvidia H100 Tensor Core GPUs | 200,000+ GPUs (H100 + Blackwell B200) |
| Interconnect Architecture | 800 Gbps Quantum-2 InfiniBand | 3.2 Tbps Quantum-X800 Optical Fabric |
| Peak Electrical Draw | ~150 Megawatts | 300+ Megawatts (Scaling to 1 GW) |
| Cooling Methodology | Hybrid Air & Closed-Loop Chilled Water | 100% Direct-to-Chip Liquid Cooling |
| Primary Mission | Grok 4.6 & Grok 4.7 Training | Grok 5 Pre-training & Multimodal Video |
Power Infrastructure: Scaling Beyond the Regional Grid
A datacenter demanding several hundred megawatts cannot plug into standard commercial electrical utility lines without causing municipal grid instability.
Utility Interconnections and Megawatt Substations
To support Colossus 2, xAI collaborated with local utility providers and the Tennessee Valley Authority (TVA) to construct dedicated high-voltage sub-stations. The facility incorporates:
- High-Voltage Step-Down Transformers: Step down 161 kV transmission lines directly to 480V/415V three-phase busbars servicing server rows.
- Tesla Megapack Grid Buffering: Industrial-scale battery energy storage systems (BESS) smooth transient power spikes caused by synchronized training checkpoint writes and sudden GPU load shedding.
- Mobile Turbine Generation: Temporary natural gas turbine arrays bridge grid interconnection delays, ensuring training runs proceed without waiting for long-lead utility switchgear deployments.
Thermal Engineering: Direct-to-Chip Closed-Loop Liquid Cooling
Dissipating over 300 megawatts of thermal energy from concentrated server racks is impossible with traditional forced-air HVAC cooling towers alone. The thermal density of an NVL72 rack exceeds 120 kilowatts per cabinet.
graph TD
Pump[Primary Coolant Distribution Unit CDU] --> ColdPlates[Direct-to-Chip Micro-Channel Cold Plates]
ColdPlates --> GPU1[Blackwell B200 Compute Dies]
ColdPlates --> GPU2[ConnectX-8 Network Interfaces]
GPU1 & GPU2 --> WarmFluid[Warm Coolant Return 45°C]
WarmFluid --> HeatExchanger[External Adiabatic Cooling Towers]
HeatExchanger --> PumpDirect-to-Chip Micro-Channel Mechanics:
- Cold Plate Micro-Channels: Dielectric fluid flows over micro-skived copper channels situated directly atop the GPU and HBM3e memory packages, maintaining die junction temperatures below 75°C under continuous 1,000W thermal design power (TDP).
- Coolant Distribution Units (CDUs): Multi-pump CDUs regulate flow rates, temperature deltas, and pressure curves throughout the server rows.
- Leak-Detection and Dry-Break Couplers: Every rack features automated shutoff valves that isolate individual compute sleds in the event of micro-pressure drops, preventing coolant exposure to live high-voltage components.
Network Fabric: Non-Blocking Optical InfiniBand
In distributed large-language model training, GPU compute efficiency depends on all-reduce communication synchronization. If one accelerator stalls waiting for gradient synchronization, all other 200,000 GPUs sit idle.
To prevent communication bottlenecks:
- Single-Tier Non-Blocking Optical Fabric: Incorporating Quantum-X800 optical switches, Colossus 2 eliminates multi-tier network oversubscription.
- RDMA (Remote Direct Memory Access): Enables direct memory transfers between GPUs across disparate racks without passing through host CPU operating system kernel stacks.
- Adaptive Routing: Hardware-level packet distribution routes around micro-burst network congestion in real time, sustaining over 92% Model Flops Utilization (MFU).
Frequently Asked Questions
Where is xAI Colossus located?
Colossus is located in Memphis, Tennessee, occupying a repurposed industrial manufacturing complex re-engineered specifically for high-density compute.
What model is being trained on Colossus 2?
Colossus 2 is dedicated primarily to pre-training and reinforcement learning fine-tuning for Grok 5, xAI's next-generation multimodal foundation model.
How does Colossus 2 compare to other superclusters?
With over 200,000 interconnected accelerators, Colossus 2 ranks among the top concentrated private AI training facilities globally alongside frontier clusters operated by Meta, Google, and Microsoft.
Conclusion
Colossus 2 represents a historic milestone in supercomputing infrastructure. By solving the physics hurdles of extreme power delivery, high-density direct-to-chip cooling, and low-latency optical fabrics, xAI has constructed the computational engine necessary to power Grok 5 and push the frontier of artificial intelligence forward.
