NVIDIA GB300 Debuts: 1.5x AI Performance Boost

On March 18, at the annual GTC 2025 event, NVIDIA officially unveiled its latest AI platform, Blackwell Ultra, alongside the highly anticipated GB300 series chips. These new chips offer a 1.5x performance boost over the previous GB200 series, expanding AI capabilities in inference, autonomous AI agents, and real-world AI applications.

With this launch, NVIDIA continues to push the boundaries of AI computing. Industry analysts expect the GB300 series to be manufactured by TSMC, while leading technology firms—including Foxconn, Quanta, Inventec, Wistron, ASUS, and Pegatron—have announced immediate plans to integrate these advanced AI chips into next-generation AI server solutions. Shipments are expected to commence in the second half of 2025, capitalizing on the growing AI market.

NVIDIA CEO Jensen Huang emphasized that AI technology is advancing at an unprecedented pace, requiring significantly greater computational power for inference and autonomous AI operations. The Blackwell Ultra platform is engineered to meet these demands, optimizing pretraining, fine-tuning, and AI inference with superior efficiency.

For autonomous AI agents, Blackwell Ultra enables AI systems to analyze complex, multi-step problems, allowing them to reason, plan, and take actions rather than merely executing predefined instructions.

In real-world AI applications, Blackwell Ultra enhances synthetic media generation in real-time, making it ideal for large-scale training in robotics and autonomous driving.

NVIDIA's new AI infrastructure solutions include the GB300 NVL72 and HGX B300 NVL16, both built on the Blackwell architecture. The GB300 NVL72 system integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs based on Arm's Neoverse architecture, delivering 1.5x the AI performance of the GB200 NVL72. Compared to previous-generation Hopper-based products, Blackwell significantly expands AI-driven revenue opportunities, with a 50x increase in AI factory applications.

Additionally, NVIDIA introduced the DGX SuperPOD AI supercomputing system, featuring GB300 NVL72 rack architecture for a seamless AI factory solution. The DGX GB300 system utilizes rack-scale liquid cooling, while the air-cooled DGX B300 system is based on the B300 NVL16 platform.

Compared to the Hopper-based generation, the HGX B300 NVL16 achieves 11x faster inference speed, 7x higher computing power, and 4x greater memory capacity for large-scale AI workloads. This next-generation performance is designed to accelerate complex AI inference tasks with groundbreaking efficiency.

Starting in the second half of 2025, industry leaders such as Aivres, ASRock, ASUS, Foxconn, Gigabyte, Inventec, Pegatron, Wiwynn, Cisco, Dell, HPE, Lenovo, and Supermicro will roll out a diverse range of Blackwell Ultra-powered AI servers.

Cloud service providers, including Amazon AWS, Google Cloud, Microsoft Azure, and Oracle Cloud, will also be among the first to integrate Blackwell Ultra-powered AI services into their platforms.

ASUS, Gigabyte, and MSI have unveiled their latest AI server solutions. ASUS and Gigabyte introduced GB300-based AI servers, while MSI launched 4U and 2U AI server platforms based on NVIDIA's MGX architecture. At GTC 2025, ASUS presented the ASUS AI POD powered by GB300 NVL72, securing key orders for large-scale AI deployments.