Nvidia itself is facing delivery issues in the AI chain. This impacts the number of new GPUs the manufacturer can ship. As a result, the older generation Blackwell continues to dominate the market.
Nvidia will sell fewer new Rubin processors than expected this year. This is what TrendForce predicts. Rubin will account for approximately 22% of Nvidia’s GPU sales, not 29% as previously assumed.
The popularity of the old Blackwell architecture is increasing accordingly: 71% of GPUs will be equipped with this architecture (instead of 61%). The remaining 7% of Nvidia’s high-end GPU sales are made up of Hopper.
Irony
Part of the reason for Nvidia Rubin’s delayed shipments is the hype around AI that Nvidia itself created. The memory supply chain is under immense pressure due to the high demand for memory for AI chips. This initially leads to exorbitant prices and shortages in the PC and smartphone market, but little by little, AI players themselves are faced with limited production capacity.
The amount of HBM memory, so critical for AI chips, is therefore limited. Added to this is the fact that the war against Iran is impacting global helium production. This is also being felt in the microchip supply chain.
Technical challenges also come into play. These include HBM4 memory validation, moving from CX8 to CX9 network interconnects, higher power consumption, and additional adjustments for more advanced liquid cooling.
This complexity, limited memory availability and geopolitical tensions in the Gulf mean that Nvidia can deliver fewer Rubin chips than hoped.
Not the first time
This isn’t entirely a surprise. After Nvidia announced the first Blackwell chips in 2024 with great fanfare, it also took longer than expected to launch volume production. At the time, the problems were exclusively technical: a design flaw caused several months of delay.
Subsequently, Nvidia continued to experience difficulties with Blackwell production. Due to the complexity of the chip, the number of compliant GPUs coming off TSMC’s production lines was initially lower than expected.
Blackwell reste pertinent
Eventually all issues were resolved. Blackwell chips are now leaving factories in large volumes. Due to Rubin’s slow start, these GPUs are now enjoying additional time in the spotlight. That’s because big AI companies want as much computing power as possible, as quickly as possible. If Rubin is not available, we will have to settle for Blackwell.
If Rubin is not available, we will have to settle for Blackwell
According to TrendForce, the share of this series will increase from 61 to 71% of Nvidia’s high-end GPU sales in 2026. It is mainly the GB300 and B300 systems which should drive this growth.
The GB200 and B200 series also remain relevant. Volumes are certainly lower, but existing orders and demand from more price-sensitive customers can support deliveries until the second half of 2026.
Versatile policy towards China
Nvidia is suffering the consequences of US policy, and it’s not just the impact of the war in Iran on production. The Trump administration’s unstable policy towards China is reflected in the drop in sales of Hopper chips.
Hopper is the predecessor of Blackwell and already dates from 2024. However, Hopper chips are still efficient. Nvidia especially hoped to be able to sell the Hopper series in China. As the chips are less powerful, they theoretically do not fall under US export restrictions. China itself does not yet have large volumes of its own AI chips capable of matching Nvidia’s hardware, making the Hopper H200 an attractive component for powering AI data centers.
However, Hopper’s share drops from 10 to 7%. The United States first completely restricted the export of Nvidia chips, before turning around and authorizing the sale of the H200 in China. The national security risks mentioned above suddenly evaporated, provided Nvidia ceded 25% of its revenue to the government.
After all this pirouettes, the Chinese were in no hurry to allow Hopper on the market. After much hesitation, the situation began to improve and production of the H200 was being increased to meet real demand.
Perspectives positives pour Nvidia
Despite all obstacles, Nvidia’s total high-end GPU shipments continue to grow. TrendForce only slightly reduces the expected annual growth, from 26.8% to around 26%.
Demand remains high: if it’s an Nvidia AI chip coming off a TSMC production line, there’s bound to be a deep-pocketed AI specialist lining up to buy the component. This immediately explains why the lower sales forecast does not have a noticeable impact on Nvidia’s stock price.
