NVIDIA GB300: New 72-GPU AI Servers to Ship in 2025 | Gpu hardware for gaming | Cpu-z | Cpu hardware list | Turtles AI

NVIDIA GB300: New 72-GPU AI Servers to Ship in 2025
Foxconn, Other Taiwanese Makers Speed ​​Up Assembly of Blackwell Ultra Systems as Supply Chain Pressure Mounts and Alternatives to NVIDIA Chips Emerge
Editorial Team1 July 2025

 

NVIDIA GB300 (Blackwell Ultra) AI servers with 72 GPUs to start shipping in H2 2025. Foxconn leads orders alongside Quanta, Wistron, Wiwynn and Inventec. With growing demand, pressure on prices and supplies is emerging, with alternatives in development.

Key points:

  • GB300 H2 2025 shipments: launch expected in H2, with Quanta ready in September.
  • Foxconn main partner: holds the largest number of orders, with early production on the new iPhone.
  • Advanced technology: NVL72 rack with 72 Blackwell Ultra GPUs + 36 Grace CPUs, liquid cooling, 1.6Tbps interconnects.
  • Market pressures: high costs and component constraints fuel interest in alternative solutions (TPU, AWS/Alphabet chips, Broadcom…).

NVIDIA’s latest GB300 servers, based on the Blackwell Ultra platform, are expected in the second half of 2025, thanks to intensive preparation at its Taiwanese assembly sites. Foxconn, the undisputed leader in production, has won the largest share of orders: the decision to anticipate this line compared to the iPhone chain underscores the strategic value attributed to the new generation of servers. Despite the similarity with the GB200, Foxconn admits that there may be obstacles in the production scale phase, especially in thermal control and GPU integration.

Alongside Foxconn, Quanta Computer stands out: it has already started shipments of the GB200 in the second quarter and is completing tests for the GB300, aiming for a mass release in September. Despite the superior power of the new systems, large AI customers may continue to order the previous models, given the urgency of their training workloads. Similarly, Wiwynn and Inventec are expected to start shipping in August-September, with volumes growing in the second half of the year and ambitions to increase their market share.

Technologically, the NVIDIA GB300 NVL72 CIS combines 72 Blackwell Ultra GPUs with 36 Grace ARM CPUs, supported by up to 800Gb/s Ethernet/InfiniBand interconnects and 1.6Tbps optical systems with energy optimizations. The racks adopt liquid cooling and technical innovations, such as liquid trays, cold plates, CX8 interfaces and enhanced 12-layer HBM3E modules (288GB). In addition, the NVIDIA ecosystem includes integration into the DGX Cloud and SuperPOD platform, offering advanced AI instances for distributed computing and reasoning model inference.

Global demand for AI servers is putting pressure on supply chains: high costs, thermal complexities and constraints in the procurement of chips and optical components. The recent improvement in the GB200 lines, which addressed overheating and leaky liquid circuits, supports increasing production, with the goal of meeting a quarterly target of about 240,000 systems globally. Similarly, the arrival of the GB300s is prompting a search for alternative solutions—from Google TPUs to offerings from AWS/Alphabet, Broadcom, Marvell—even though CEO Jensen Huang remains confident in the superiority of the NVIDIA portfolio.

Finally, Foxconn and NVIDIA are reportedly experimenting with humanoid robots in a new Houston plant, coming online early next year, to automate tasks such as component placement and cable assembly, a sign of the increasing automation in the production of highly specialized servers.

The debut of the GB300 servers marks a critical step in the evolution of AI infrastructure, highlighting the intensification of technological competition and the strategic importance of manufacturing capabilities in the new phase of advanced computing.