Secure AI: Supermicro’s HGX B300 & GB300 NVL72 with NVIDIA Confidential Computing

In the race for AI innovation, security is no longer optional. While traditional measures protect data at rest and in transit, NVIDIA Confidential Computing addresses the final frontier: data in use. By pairing this technology with Supermicro’s HGX B300 and GB300 NVL72 systems, enterprises can deploy high-performance AI that meets the strictest compliance and data sovereignty standards.

The Power of NVIDIA Confidential Computing

NVIDIA Confidential Computing utilizes Trusted Execution Environments (TEEs) integrated directly into the NVIDIA Blackwell architecture. This creates hardware-rooted, verifiable enclaves where code and data remain encrypted during computation. Protection: Shields data from unauthorized access, such as malicious insiders or malicious code attacks.

• Versatility: Supports Confidential VMs and Confidential Containers, allowing integration into cloud-native environments without major code changes.
• Integrity: Safeguards model weights, prompts, and outputs, enabling secure large-scale training and inference.

Enterprises can confidently deploy AI inference and fine-tuning workloads on-premises, in public clouds, or sovereign infrastructure. It enables AI factories to securely host AI workloads while meeting strict compliance, data sovereignty, and regulatory requirements.

Supermicro’s Performance with NVIDIA HGX B300 Systems

Supermicro has expanded its Blackwell portfolio to deliver extreme density and efficiency for AI factories.

• NVIDIA HGX B300: Available in 8U (air-cooled) or 4U (liquid-cooled) models, it delivers up to 7.5x the performance of previous Hopper systems. It features 8 NVIDIA Blackwell Ultra GPUs with 288GB of HBM3e memory each, interconnected via 1.8TB/s NVLink.
• NVIDIA GB300 NVL72: A liquid-cooled rack-scale solution integrating 72 Blackwell Ultra GPUs and 36 Grace CPUs. It delivers 1.1 exaFLOPS of dense FP4 compute—acting as a single, massive GPU designed for models with over 1 trillion parameters.

Key features

Includes support for Confidential VMs and Confidential Containers, enabling seamless integration into virtualized and bare-metal environments without major code changes. It empowers AI factories to host large-scale training and inference while preserving model integrity and preventing exposure of prompts, weights, or outputs. This technology accelerates the adoption of trustworthy AI, combining unparalleled GPU performance with verifiable security to unlock insights from previously untouchable sensitive data.

Enabling Secure AI at Scale

Supermicro systems fully leverage the industry’s first TEE-I/O capable GPUs. This ensures data remains encrypted even during high-speed communications across NVLink. By combining processor-level security (Intel TDX or AMD SEV) with NVIDIA’s GPU TEEs, Supermicro provides a "defense-in-depth" foundation.

Key Benefits for Enterprises:

• Regulatory Compliance: Meets stringent GDPR and HIPAA requirements for sensitive data in healthcare and finance.
• Sovereign AI: Enables the creation of secure, localized AI infrastructure.
• Efficiency: The GB300 NVL72 includes power-smoothing technology to reduce peak grid demand by up to 30%.

Conclusion

The collaboration between Supermicro and NVIDIA represents the future of trustworthy and confidential AI. By combining record-breaking performance with verifiable hardware security, the HGX B300 and GB300 NVL72 allow organizations to unlock insights from their most sensitive data without compromising integrity.