Oriented on making people closer

NVIDIA has commenced limited production of its new generation of high-performance chips, the Grace Blackwell GB300. These superchips are engineered to handle the most resource-intensive artificial intelligence tasks. According to DigiTimes, initial shipments are anticipated in September 2025, with production volumes subsequently increasing.

The GB300 chip represents an integrated solution, comprising a powerful Grace CPU (based on the Arm Neoverse V2 architecture) and two Blackwell Ultra Graphics Processing Units. This combination enables exceptional performance in AI workloads. Early reports from manufacturers indicate no significant issues with the GB300, and its release is expected to proceed without substantial difficulties.

Simultaneously, high demand persists for the preceding GB200 accelerators, despite identified cooling challenges. Due to their exceptionally high computational power and dense server configurations, GB200 units necessitated complex liquid cooling systems. During operation, instances of leaks were reported, often linked to quick-disconnect couplings that did not consistently withstand varying pressure conditions and pipeline configurations. This led to additional time and financial expenditures for maintenance.

In contrast to the GB200, which required comprehensive platform modifications when transitioning from the previous Hopper architecture, the new GB300 is designed to leverage existing infrastructure. This is expected to circumvent leakage issues and ensure a smoother ramp-up to mass production, scheduled for Q3 2025 with subsequent volume increases in Q4.

NVIDIA has already announced the advent of the next generation of accelerators—the Rubin family, which will succeed Blackwell Ultra in the second half of 2026. These chips will offer even greater performance and necessitate the implementation of an entirely new Kyber rack architecture with even higher packaging density (up to 600 kW per rack), rendering liquid cooling systems mandatory.

Furthermore, the resumption of production for previously curtailed H20 accelerators, which the U.S. has once again authorized for shipment to China, may encounter difficulties. According to Reuters, citing The Information, TSMC, a key manufacturer, had already reallocated H20 production lines to other products, and a full restoration of their output could take up to nine months.

Photo: NVIDIA