NVIDIA has been riding high in the artificial intelligence sector, but it seems to have hit some roadblocks recently. The latest Blackwell Ultra series product, the GB300, debuted at GTC 2025 with high industry expectations for its performance improvements. However, feedback from the supply chain and customers paints a different picture. This next-generation AI server core hardware appears to struggle to replicate the success of its predecessors, due to both technical complexities and a market preference for mature solutions.
The GB300 is an upgraded version of the Blackwell series, featuring the B300 GPU with a single-card power consumption of up to 1400W. Its FP4 computing performance is approximately 50% higher than the GB200, and its memory capacity has increased from 192GB to 288GB, utilizing a 12-layer HBM3E stack design. Additionally, network performance has been upgraded from ConnectX-7 to ConnectX-8, with optical module bandwidth jumping from 800Gbps to 1.6Tbps. These enhancements cater to the growing demands of AI inference and training, particularly in cutting-edge applications like agentic AI and physical AI. But as a Chinese saying goes: “Take too big a step, and you might trip.” For now, NVIDIA’s technological leap has not directly translated into market acceptance.
By comparison, the previous-generation GB200 has also underperformed somewhat. Throughout 2024, NVIDIA shipped only about 15,000 GB200 AI servers—far below the success of the earlier Hopper series. The reasons include low yield rates in the initial production phase due to TSMC’s advanced packaging technology (like CoWoS-L), though this bottleneck has since been alleviated. However, subsequent deployment challenges followed. Installing and debugging the GB200 NVL72 server rack is exceptionally complex, with a single system taking 5 to 7 days to deploy. It also frequently experiences instability and system crashes during operation. Notably, its NVLink copper cable design requires each of the 5,000 cables to be uniquely fitted, making installation extremely difficult. Even more challenging is the configuration process, which heavily relies on NVIDIA engineers’ expertise, leaving customers nearly unable to operate independently. This dependency has frustrated cloud service providers (CSPs), especially when clusters fail and they must wait for NVIDIA’s technical support.
Supply chain pressures are equally significant. The GB300’s mass production was originally slated for the second half of 2025, but the latest updates suggest that customer test samples won’t be delivered until the end of 2025 at the earliest, with full production possibly delayed to 2026. This delay stems from the GB300’s increased design complexity—such as its full liquid-cooling solution, which sharply increases demand for cooling components, and its high power consumption, which places greater demands on power systems. Compared to the GB200 NVL72 rack’s 140kW power draw, the GB300’s energy consumption is expected to be even higher, driving up production costs and posing new challenges to data centers’ power and cooling infrastructure. While liquid cooling effectively manages heat loads, its adoption remains limited, and many customers’ data centers are not yet equipped for the transition.
In response to these issues, major CSPs like Microsoft, Google, and Amazon have adopted a cautious stance toward the GB300. These companies were early adopters of the GB200, but their experiences fell short of expectations. For instance, Microsoft found that yield issues with the GB200 NVL72 required repeated hardware tuning, disrupting data center computing deployment plans. Disappointed, CSPs have begun shifting to NVIDIA’s more mature solutions, with the HGX series emerging as a popular choice. Known for its stability, the Hopper-based HGX H100 boasts a single-rack power consumption of 60kW to 80kW, compatibility with traditional air-cooling designs, and far greater deployment flexibility than the GB200 or GB300. Additionally, the HGX B200, equipped with eight B200 GPUs connected via NVLink, supports network speeds up to 400Gbps and has been widely adopted in x86-based AI platforms, earning favor among hyperscalers and smaller cloud providers alike.
Shifting market demand is also evident in order data. Supply chain sources indicate that GB300 pre-orders have fallen short of expectations, while HGX system orders are steadily rising. Analysts suggest that the GB300’s high pricing may be another hurdle. For example, a single GB200 NVL72 system costs around $3 million, and the GB300’s liquid-cooling design and performance upgrades are expected to push costs even higher. For budget-conscious customers, this premium may be hard to justify, especially with limited short-term returns on investment.
NVIDIA’s competitive landscape is quietly shifting as well. Though it still dominates the AI GPU market, rivals like AMD and Intel are gaining ground. AMD’s Instinct MI300 series is steadily capturing market share with its cost-effectiveness and open-source ecosystem, while Intel’s Gaudi 3 excels in specific inference tasks. While these alternatives can’t fully challenge NVIDIA’s position, they offer customers more options.
To address these challenges, NVIDIA is considering strategic adjustments, such as launching a single-GPU version of the Blackwell chip to shorten delivery timelines. Whether these efforts can restore customer confidence remains to be seen. Currently, NVIDIA stands at a delicate balance between technological innovation and market acceptance. How it streamlines its supply chain and meets customer expectations for stability and cost will determine its next steps. We all know Jensen Huang is a driven individual, sometimes pushing himself to the brink—and for now, signs of that strain are starting to show.