The Hardware: NVIDIA GB300 NVL72 by HPE

Vultr will deploy the NVIDIA GB300 NVL72 system, built by HPE, as the core of its next-generation AI infrastructure. This rack-scale architecture pairs NVIDIA Blackwell Ultra GPUs with NVIDIA Grace CPUs in a 72-GPU NVLink domain . The key specs are eye-catching:

• AI Performance: Up to 1,440 PFLOPS of FP4 AI performance
• Memory: 279 GB of HBM3e memory per GPU (37 TB total across the rack), with 8 TB/s memory bandwidth per GPU
• Architecture: Optimized for large AI model training, fine-tuning, and real-time inference for models over 1 trillion parameters
• Cooling: Fully liquid-cooled design integrated into HPE's AI factory architecture

The GB300 NVL72 is described as the "new inference/training workhorse" with an explicit cost-per-token narrative, reflecting the industry's pivot toward operationalizing AI models at scale for real-time applications . According to NVIDIA's product page, it delivers 1.5x more dense FP4 Tensor Core FLOPS and 2x higher attention performance compared to NVIDIA Blackwell GPUs .

High-Bandwidth Networking: The Backbone of AI Inference

High-bandwidth networking is a foundational pillar of the deployment. The NVIDIA GB300 NVL72 systems are interconnected using NVIDIA Spectrum-X Ethernet networking, a lossless, high-throughput Ethernet fabric designed to eliminate bottlenecks in multi-node AI workloads .

The deployment features 400GbE and 800GbE interconnects, optical transceivers, NVIDIA Spectrum-X Ethernet switches, and SuperNICs . Without this fabric, scaling inference to enterprise-grade throughput would create crippling latency and bandwidth bottlenecks. The NVIDIA Enterprise Reference Architecture for the NVL72 AI Factory confirms that this dual-plane networking architecture is designed to power enterprise data centers for AI training and inference at massive scale, enabling real-time applications and trillion-parameter models .

Strategic Shift: From Training to Inference

While the infrastructure supports both training and inference , the announcement emphasizes a growing enterprise focus on AI inference (production deployment) alongside training. Several signals point to this shift:

• The GB300 NVL72 is purpose-built for "test-time scaling inference and AI reasoning tasks"
• It is explicitly positioned with a cost-per-token narrative, measuring value in inference output rather than raw training throughput
• HPE's broader messaging at the event focused on bringing AI into production with security, governance, scale, and sovereignty

The implication is clear: enterprises have moved past the experimentation phase and are now seeking infrastructure optimized for deploying AI models at scale in production environments.

Other Announcements at HPE Discover Las Vegas 2026

Beyond the Vultr partnership, HPE highlighted several related initiatives:

• Agentic AI push: HPE announced a broader initiative to bring agentic AI into production with NVIDIA, emphasizing security, governance, scale, and sovereignty
• Customer deployments: S k y Co., Ltd. deployed HPE Private Cloud AI as part of the NVIDIA AI Computing by HPE portfolio
• Nubax in Argentina: The company launched a consumption-based AI-as-a-service model, delivering 62.5% better supply chain forecasting, 35% shorter wait times, and 40% cost savings, powered by NVIDIA AI Computing by HPE

What This Means for Enterprise AI

Vultr's choice of HPE and NVIDIA signals a major inflection point. As the largest privately held hyperscaler, Vultr is betting that enterprise customers need infrastructure that can handle both training and real-time inference at cloud scale. By combining NVIDIA's rack-scale GPU computing with HPE's factory architecture, liquid cooling, and services, Vultr is positioning itself to serve the next wave of enterprise AI workloads—from model training to production inference on trillion-parameter models.

The Hardware: NVIDIA GB300 NVL72 by HPE

Vultr will deploy the NVIDIA GB300 NVL72 system, built by HPE, as the core of its next-generation AI infrastructure. This rack-scale architecture pairs NVIDIA Blackwell Ultra GPUs with NVIDIA Grace CPUs in a 72-GPU NVLink domain . The key specs are eye-catching:

• AI Performance: Up to 1,440 PFLOPS of FP4 AI performance
• Memory: 279 GB of HBM3e memory per GPU (37 TB total across the rack), with 8 TB/s memory bandwidth per GPU
• Architecture: Optimized for large AI model training, fine-tuning, and real-time inference for models over 1 trillion parameters
• Cooling: Fully liquid-cooled design integrated into HPE's AI factory architecture

The GB300 NVL72 is described as the "new inference/training workhorse" with an explicit cost-per-token narrative, reflecting the industry's pivot toward operationalizing AI models at scale for real-time applications . According to NVIDIA's product page, it delivers 1.5x more dense FP4 Tensor Core FLOPS and 2x higher attention performance compared to NVIDIA Blackwell GPUs .

High-Bandwidth Networking: The Backbone of AI Inference

High-bandwidth networking is a foundational pillar of the deployment. The NVIDIA GB300 NVL72 systems are interconnected using NVIDIA Spectrum-X Ethernet networking, a lossless, high-throughput Ethernet fabric designed to eliminate bottlenecks in multi-node AI workloads .

The deployment features 400GbE and 800GbE interconnects, optical transceivers, NVIDIA Spectrum-X Ethernet switches, and SuperNICs . Without this fabric, scaling inference to enterprise-grade throughput would create crippling latency and bandwidth bottlenecks. The NVIDIA Enterprise Reference Architecture for the NVL72 AI Factory confirms that this dual-plane networking architecture is designed to power enterprise data centers for AI training and inference at massive scale, enabling real-time applications and trillion-parameter models .

Strategic Shift: From Training to Inference

While the infrastructure supports both training and inference , the announcement emphasizes a growing enterprise focus on AI inference (production deployment) alongside training. Several signals point to this shift:

• The GB300 NVL72 is purpose-built for "test-time scaling inference and AI reasoning tasks"
• It is explicitly positioned with a cost-per-token narrative, measuring value in inference output rather than raw training throughput
• HPE's broader messaging at the event focused on bringing AI into production with security, governance, scale, and sovereignty

The implication is clear: enterprises have moved past the experimentation phase and are now seeking infrastructure optimized for deploying AI models at scale in production environments.

Other Announcements at HPE Discover Las Vegas 2026

Beyond the Vultr partnership, HPE highlighted several related initiatives:

• Agentic AI push: HPE announced a broader initiative to bring agentic AI into production with NVIDIA, emphasizing security, governance, scale, and sovereignty
• Customer deployments: S k y Co., Ltd. deployed HPE Private Cloud AI as part of the NVIDIA AI Computing by HPE portfolio
• Nubax in Argentina: The company launched a consumption-based AI-as-a-service model, delivering 62.5% better supply chain forecasting, 35% shorter wait times, and 40% cost savings, powered by NVIDIA AI Computing by HPE

What This Means for Enterprise AI

Vultr's choice of HPE and NVIDIA signals a major inflection point. As the largest privately held hyperscaler, Vultr is betting that enterprise customers need infrastructure that can handle both training and real-time inference at cloud scale. By combining NVIDIA's rack-scale GPU computing with HPE's factory architecture, liquid cooling, and services, Vultr is positioning itself to serve the next wave of enterprise AI workloads—from model training to production inference on trillion-parameter models.