We are witnessing a fundamental shift in the scale of digital infrastructure. The era of megawatt data centers is over; we have entered the age of the 1-Gigawatt (1GW) AI cluster.
Scaling Beyond the Horizon
Thinking Machines Lab, in partnership with NVIDIA, has announced plans to build the world's first 1-gigawatt computing facility. Powered entirely by the Vera Rubin platform, this facility will consume as much power as a small city, dedicated entirely to a single unified AI model.
This isn't just about adding more chips. It requires a total rethink of power delivery, cooling, and networking. To handle the 1,000+ megawatts of load, the facility will feature on-site Small Modular Reactors (SMRs) and a direct high-voltage DC (HVDC) power bus to the server racks.
The 1.6T Networking Revolution
Data movement is the primary bottleneck at this scale. Thinking Machines is utilizing 1.6T (1,600 Gbps) Ethernet and InfiniBand interconnects. Using Expanded Beam Optical (EBO) technology from 3M, these interconnects allow for near-zero loss over longer fiber runs, enabling a cluster that spans multiple buildings to act as a single low-latency machine.
The NVLink 6 Cognitive Interconnects on the Rubin chips will handle intra-rack communication, while the 1.6T fabric manages the massive East-West traffic between racks.
1GW Cluster Architecture
- - Power Capacity: 1,000,000,000 Watts (1GW).
- - Networking: 1.6 Tbps Optical Fabric.
- - Cooling: Direct-to-Chip Liquid Cooling (DLC).
- - Compute: 100,000+ NVIDIA GB300 NVL72 nodes.
Why Do We Need This?
frontier models are now being trained on multi-modal, real-world data—including high-resolution video and real-time sensor feeds from billions of IoT devices. To process this amount of data and achieve the next level of Agentic General Intelligence (AGI), the compute density must increase by orders of magnitude.
The 1GW cluster is the only way to provide the FLOPs necessary for a model to "understand" the physical world with the same fidelity as a human, but at a trillion-token scale.