Is NVIDIA Restructuring Its Roadmap? Less Hopper and Rubin, More Blackwell in 2026
According to the latest analyses from TrendForce on the AI server market, 2026 will see a significant change in the mix of high-end GPU shipments from NVIDIA. Overall demand remains strong, but the distribution among different architectures will undergo a reshuffle due to technical, industrial, and geopolitical factors.
The Blackwell platform is set to strengthen its dominant position, with an estimated share rising from 61% to 71% of total shipments. The main drivers will be solutions based on GB300 and B300, while the GB200 and B200 models will continue to be shipped to meet existing orders and cost-sensitive segments.
On the other hand, the outlook for the future Rubin architecture has been scaled down. TrendForce has revised its forecast for this year downward, from 29% to 22%. This revision is based on several technical challenges.
Among the key obstacles are the time needed for the validation of HBM4 memory, the transition to the new ConnectX-9 network interconnects, and a significant increase in power consumption. Additionally, the complexities associated with adopting more advanced liquid cooling solutions, necessary for managing increasingly high power densities, are also slowing down large-scale production and distribution, directly affecting the availability of Rubin-based systems.
The Hopper generation, including the H200 accelerators, will also see a contraction in its share of the overall mix, dropping from 10% to 7%. In this case, the determining factor is geopolitical uncertainty, particularly in trade relations between the United States and China. Despite a partial easing of export restrictions to the Chinese market, the approval and implementation timelines of agreements have slowed down deliveries. NVIDIA has, however, restarted production of H200 accelerators destined for that market, supported by already acquired orders, but the overall impact remains limited in the short term.
The overall context remains expansive. The increasing demand for AI infrastructure and the push for high-density integrated rack solutions—such as GB and VR platforms—will lead to a significant increase in total high-end GPU shipments. TrendForce has, however, slightly reduced its annual growth estimates, adjusting them from a previous +26.8% to around +26%, reflecting operational difficulties related to the technological transition.
In parallel with the consolidation in AI training, NVIDIA is expanding its presence in the inference segment. In this area, new LPUs (Language Processing Units) designed to accelerate the token generation phase are emerging, at the center of the latest GTC.
According to estimates, the demand for these solutions could reach several hundreds of thousands of units by 2026, potentially doubling in 2027. However, the limited amount of on-chip SRAM memory implies the need to employ numerous chips for large-scale workloads.
The company is also pushing products such as the RTX PRO 4500/6000 series to cover mid-range, entry-level, and edge AI segments. This is expected to lead the share of mid-low-end solutions to exceed 32% of total shipments in 2026.
Further complicating the picture is the memory market. TrendForce indicates a possible increase in DRAM prices between 45% and 50% in the second quarter, following increases already between 75% and 80% in the first quarter.