Nvidia’s key manufacturing partners are accelerating production of its flagship Blackwell AI server racks. This follows the resolution of significant technical issues that previously delayed shipments, the Financial Times reports. Suppliers including Foxconn, Inventec, Dell, and Wistron are now overcoming these challenges, a crucial step for Nvidia to meet surging global AI hardware demand.
The breakthrough is vital for the AI industry, as Blackwell systems are designed to power next-generation large language models. Previously, Nvidia and its partners grappled with multiple problems. These included overheating GPUs, leaks in liquid cooling systems, software bugs, and inter-chip connectivity issues, an engineer at one partner firm confirmed to the Financial Times. These setbacks disrupted major clients like Microsoft and Google in 2024.
Investor confidence appears buoyed by the positive developments. Reuters reported a 2.5% rise in Nvidia’s pre-market stock on May 27. CEO Jensen Huang expressed strong confidence in the supply chain, stating in the last earnings call that Blackwell is “the fastest product ramp in our company’s history, unprecedented in its speed and scale.”
Production Hurdles Cleared
The complexity of the Blackwell GB200 racks, each integrating 36 Grace central processing units and 72 Blackwell graphics processing units via Nvidia’s NVLink system, presented substantial engineering hurdles. Chu Wei-Chia, a Taipei-based analyst at consultancy SemiAnalysis, told the Financial Times the technical challenge was immense, as no company had previously attempted to synchronize so many AI processors in a server so quickly.
Chu also suggested to the Financial Times that Nvidia initially hadn’t provided the supply chain sufficient time, contributing to delays, but projected that inventory risks for the GB200 should now decrease as rack output rises.
Specific issues like liquid cooling were major obstacles. While Liquid cooling was a major hurdle, collaborative efforts of Nvidia and technological partners optimized the assembly and testing processes, leading to much better yields. This, along with improved yields for high-speed interconnects, has been pivotal.
Consequently, shipments of the GB200 racks, which commenced at the end of the first quarter of 2025, are now rapidly scaling. Suppliers have also enhanced testing protocols before shipping to ensure functionality for demanding AI workloads, according to the Financial Times. The earlier production difficulties had even led Microsoft to reportedly reduce its GB200 orders by 40% in late 2024.
Strategic Shifts and Future Hardware
Nvidia is also preparing for the third-quarter launch of its next-generation GB300 AI rack. This system features enhanced memory capabilities designed for more complex AI reasoning models.
To accelerate the GB300 deployment, Nvidia reportedly compromised on an initial design plan. The company instructed partners in April to revert to the so called “Bianca” chip board layout—currently used in the GB200—instead of a new design called “Cordelia”, due to installation challenges, supplier sources told the Financial Times.
While the Cordelia board might have offered better margins and easier maintenance, the decision to stick with Bianca could help Nvidia achieve its sales targets. Nvidia reportedly intends to implement the Cordelia redesign in subsequent AI chip generations.
Looking further ahead, Nvidia’s roadmap shared at GTC 2025 includes the Blackwell Ultra GPU, expected in the second half of 2025. This GPU is slated to offer 20 petaflops of AI performance and 288GB of HBM3e memory.
Following this, the Vera Rubin AI architecture is anticipated for a 2026 launch. The demand for such powerful hardware extends to new customers, with Saudi Arabia and the United Arab Emirates having announced plans to acquire thousands of Blackwell chips.
Market Landscape and Ecosystem Control
The accelerated Blackwell rollout occurs as Nvidia navigates a complex global market. The company is working to offset a significant financial impact from U.S. government restrictions on exports of its H20 chip to China. This resulted in an anticipated $5.5 billion charge for the company.
Bank of America analyst Vivek Arya, cited by the Financial Times, suggested this China sales hit would drag down Nvidia’s gross margins, though a faster Blackwell rollout might help offset this in the second half of the year.
U.S. export controls have been a contentious topic. AI developer Anthropic has advocated for robust controls, stating that “maintaining America’s compute advantage through export controls is essential for national security and economic prosperity”.
Nvidia, however, pushed back, with a spokesperson saying that “American firms should focus on innovation and rise to the challenge, rather than tell tall tales that large, heavy, and sensitive electronics are somehow smuggled in ‘baby bumps’ or ‘alongside live lobsters.'”.
Amidst these debates, Nvidia CEO Jensen Huang has emphasized that “China is a very important market for Nvidia”, while the company also stated that they “are not sending any GPU designs to China to be modified to comply with export controls,” regarding GPU designs for China. The restrictions have also fueled competitors, with analyst Patrick Moorhead predicting “Chinese companies are just going to switch to Huawei.”
To further solidify its market position, Nvidia launched its NVLink Fusion program at Computex 2025. This initiative opens its proprietary interconnect technology to partners. CEO Jensen Huang highlighted the fundamental industry shift, stating that AI is compelling a re-architecture of data centers.
However, he also quipped during the event, “nothing gives me more joy than when you buy everything from Nvidia”. While NVLink Fusion aims to cultivate a more adaptable AI hardware ecosystem, some analysts, like AInvest in a commentary, have raised concerns that it could create a “closed-loop of dependency,” potentially leading to market monopolization.
The fifth-generation NVIDIA NVLink platform, integral to systems like the NVIDIA GB200 NVL72 and GB300 NVL72, provides 1.8 terabytes per second of total bandwidth per GPU.
