Enterprises and researchers now have a clearer path to deploying large-scale AI workloads locally, without sacrificing performance or efficiency. Giga Computing has introduced a range of new solutions that integrate NVIDIA’s latest Rubin GPUs and Vera CPUs, pushing the boundaries of what can be achieved on a single desktop—or across an entire data center.
The shift toward on-premises AI is no longer just about raw power; it’s about control. With systems now capable of handling models with up to 1 trillion parameters on a desk, and rack-scale configurations delivering exascale performance, the focus has turned to liquid cooling, unified memory, and software that streamlines deployment. Giga Computing’s latest portfolio demonstrates how these components come together, from compact deskside units to high-density clusters.
Desktop AI at PetaFLOPS Scale
The most immediate impact will be felt by researchers and small teams working on frontier AI. Two new deskside supercomputers—GIGABYTE AI TOP ATOM and W775-V10-L01—offer a glimpse of what’s possible when NVIDIA’s Grace Blackwell Superchip is paired with efficient thermal management.
- AI TOP ATOM delivers up to 1 petaFLOPS of performance, powered by the GB10 Grace Blackwell Superchip and 128 GB of unified memory. It’s designed for AI training and inference tasks that previously required cloud resources.
- W775-V10-L01 scales even further, with a GB300 Grace Blackwell Ultra Desktop Superchip providing 748 GB of coherent memory and up to 20 petaFLOPS of AI FP4 performance. This system can handle models with 1 trillion parameters while supporting NVIDIA RTX PRO 6000 Blackwell GPUs for additional acceleration.
These deskside units are part of a broader effort to reduce dependency on cloud infrastructure. NVIDIA’s NemoClaw, an open-source stack integrated into Giga Computing’s solutions, simplifies the deployment of always-on AI agents—such as personal research assistants or workflow automation tools—directly on local hardware. The result is faster iteration, stronger data privacy, and lower operational costs for smaller teams.
Rack-Scale Efficiency
For larger-scale deployments, Giga Computing’s rack-mounted systems unify NVIDIA Vera CPUs and Rubin GPUs in configurations that maximize performance while minimizing power consumption. The NVL72 platform, for example, combines 36 Vera CPUs and 72 Rubin GPUs in a single rack, interconnected with NVLink 6 for exascale throughput.
- Liquid cooling extends beyond the GPUs to include NVIDIA ConnectX-9 SuperNICs and storage, ensuring stable operation under heavy workloads.
- A modular design allows for incremental upgrades, balancing initial cost with future scalability.
The GIGAPOD ecosystem takes this a step further by clustering multiple GPU servers into a turnkey data center solution. It integrates Giga Computing’s POD Manager (GPM) software, which centralizes monitoring, orchestration, and deployment—critical for managing complex AI workloads at scale.
Cooling and Connectivity
Performance gains are only as valuable as the infrastructure that supports them. Giga Computing has partnered with nVent to introduce advanced liquid-cooling technologies, such as the CX121 CDU, which is modular enough to adapt to different data center layouts.
- GIGABYTE G4L4-SD3 is a 4U liquid-cooled inference server that pairs Intel Xeon CPUs with NVIDIA HGX B300 systems, optimized for high-density AI workloads.
- G2L4-SD4, built around the NVIDIA Rubin NVL8 platform, pushes compute density further in a 2U form factor, incorporating an OCP busbar design for improved power efficiency.
Air-cooled alternatives are also available. The GIGABYTE XL44-SX2, for instance, is the first server to feature NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs with 800 Gb/s SuperNIC bandwidth and PCIe Gen 6 connectivity—a combination that balances performance with operational simplicity.
AI Factory in Taiwan
The most ambitious project is the Giga Computing AI Factory Accelerator (GAIFA) in Taiwan, a centralized hub designed to accelerate the development and deployment of AI factories. By integrating NVIDIA’s Blackwell Ultra GPUs, Quantum-X800 InfiniBand networking, and Spectrum-X Ethernet, GAIFA aims to reduce testing and validation timelines while maintaining high performance.
This facility will serve as a proving ground for full-stack AI solutions, combining Giga Computing’s hardware expertise with NVIDIA’s software stack. The goal is to deliver faster, more seamless implementations for global customers—whether they’re deploying agentic AI models or large-scale HPC workloads.
A Balanced Approach
No system is without tradeoffs. While the new solutions from Giga Computing and NVIDIA push the envelope in performance, their adoption will depend on factors like power consumption, thermal management, and software maturity. Liquid-cooled designs, for example, offer superior efficiency but require more infrastructure than air-cooled alternatives.
For creators and researchers, the key takeaway is clear: the tools to bring AI workloads in-house are now more powerful—and more accessible—than ever. Whether it’s a desktop unit capable of 1 petaFLOPS or a rack-scale cluster delivering exascale performance, the focus has shifted from raw power alone to efficiency, control, and scalability.
