According to sources cited by DigiTimes, unnamed industry analysts have significantly reduced their 2025 sales predictions for Nvidia’s GB200-based NVL72 devices, slashing them by more than half. This adjustment diverges sharply from most other forecasts for AI hardware shipments this year. It’s important to note, however, that this revised forecast specifically pertains to systems built with Nvidia’s Arm processors. In contrast, machines using x86 architecture equipped with B200 GPUs for AI and HPC applications are expected to see robust sales.

Extensive Shipments of Blackwell AI GPUs Expected

The shipment estimate for 2025 concerning GB200 NVL72 cabinets has been revised down from between 50,000 and 80,000 units to between 25,000 and 35,000 units. Despite this substantial reduction, starting in the second quarter, shipments are anticipated to increase continuously throughout the year. Even with a 56% reduction, the shipment of 35,000 cabinets, each containing 72 GPUs, still translates to a significant total of 2.52 million GPUs. It’s noteworthy that these figures only apply to Grace Blackwell machines that incorporate Nvidia’s Armv9-based processors. Systems that integrate Blackwell accelerators with x86 processors will also be available.

The precise reasons behind the analysts’ decision to halve their shipment forecasts for Arm-based GB200 NVL72 cabinets remain unclear.

It could be that initial market research overestimated the demand for Arm-based systems for AI training and inference. Alternatively, some companies might prefer sticking with the more familiar x86 architecture to avoid having to adapt their software to the Armv9 instruction set architecture.

Another possibility is that some clients are opting out of using NVL72 cabinets due to their intensive power and cooling requirements, choosing instead to go for less dense configurations.

There are also speculations that some businesses might be delaying their purchases of GB200 systems in anticipation of Nvidia launching the 50% more powerful GB300 and B300 GPUs, which are expected about six months after the initial shipment of the 200-series Blackwell products. Although the introduction of these newer models will also require time, it’s unlikely that this anticipated update has severely impacted sales of the original Blackwell configurations.

Thriving AI Hardware Sector

These revised forecasts by analysts seem somewhat contradictory to other developments in the industry. TSMC, the manufacturer of Nvidia’s Blackwell AI GPUs and many other leading AI processors, is projecting nearly a 35% increase in revenue for Q1 year-over-year, primarily driven by the rising demand for AI hardware.

Leading AI server producers Foxconn and Quanta have verified that their GB200 shipment plans are progressing as expected, with substantial ongoing order volumes. Foxconn has not only met its delivery targets but has also been instrumental in the development and validation processes of the GB200 NVL72 cabinets. Moreover, to boost its production capabilities for AI servers, Foxconn has invested $128 million through its subsidiary, Ingrasys Technology USA, to acquire new facilities in California, which complements its existing locations in Mexico and Texas.

Similarly, Wistron, a major provider of baseboards for Nvidia’s GPUs, remains optimistic about the AI server market, predicting a double-digit growth in shipments for the first quarter of 2025 and a triple-digit increase over the whole year. Toward the end of 2024, the company began receiving NVL72 server cabinet shipments, challenging many of its competitors.