NVIDIA Plans to Charge Annual AI Agent Software Fee per GPU; Multiple Bandwidth Specs for Vera Rubin Lowered

NVIDIA Plans to Charge per GPU, Vera Rubin Specs Adjusted in Tandem Source: Semiconductor Horizons Published: 2026-10-27 15:36 NVIDIA AI-generated poster From AI Agent software licensing to the next-generation AI computing platform Vera Rubin, NVIDIA is simultaneously advancing software commercialization and hardware platform iteration.

AI Agent Software to Be Charged per GPU, at About $2,000 per Year

At the GTC Berlin conference held on October 22, NVIDIA CEO Jensen Huang, speaking about the company's software business, said that software licensing will increasingly become an important part of NVIDIA's business. He gave the example that the NVIDIA software needed to run various AI Agents in an AI factory costs about $2,000 per GPU per year. This statement means that NVIDIA is exploring room to further expand its software licensing business. As AI Agents gradually enter enterprise applications, the related infrastructure not only needs to provide computing power but also needs to support agent operation, tool invocation, resource access, and security management. Software and supporting services have thus become an important part of the AI infrastructure business model.

However, the $2,000 figure is currently still a price reference cited by Jensen Huang, not a formally announced uniform pricing standard. NVIDIA has not yet clarified which software the fee covers, which GPU products it applies to, or the specific implementation timeline. By comparison, NVIDIA's existing NVIDIA AI Enterprise enterprise software platform is publicly listed at $4,500 per GPU per year. Because the software scope and licensing terms corresponding to the two are not yet clear, the $2,000 figure cannot simply be regarded as a price cut for the existing platform.

In terms of potential revenue, if 5 million GPUs were to adopt a software licensing model of $2,000 per year in the future, this could theoretically generate $10 billion in revenue per year; if renewed continuously for five years, the cumulative scale could reach $50 billion. However, this is only an estimate based on assumptions, and actual revenue will still depend on licensing coverage, customer adoption rates, and renewal conditions.

Vera Rubin Specs Adjusted, HBM4 and NVLink Bandwidth Lowered

While exploring a new charging model for its software business, some hardware specifications of NVIDIA's next-generation AI computing platform Vera Rubin have also been adjusted. According to media comparisons of NVIDIA product pages and historical archives, some specifications of the Vera Rubin platform have changed, but NVIDIA has not yet issued a formal explanation for this adjustment. The relevant information shows that the HBM4 memory bandwidth of the Rubin GPU has been adjusted from the previously announced 22 TB/s to 19.2 TB/s, a decrease of about 12.7%; the sixth-generation NVLink interconnect bandwidth has been adjusted from 3.6 TB/s to 3.0 TB/s, a decrease of about 16.7%.

At the larger system level, the HBM4 bandwidth of the Vera Rubin superchip has been adjusted from 44 TB/s to 38.5 TB/s, and the NVLink bandwidth from 7.2 TB/s to 6.0 TB/s. For the Vera Rubin NVL72 rack-scale system using 72 Rubin GPUs, total HBM4 bandwidth has been adjusted from 1,580 TB/s to 1,400 TB/s, and total NVLink bandwidth from 260 TB/s to 216 TB/s. At the same time, system network bandwidth has been increased from 28.8 TB/s to 32.4 TB/s, an increase of about 12.5%. For the eight-GPU system HGX Rubin NVL8 aimed at enterprise customers, the adjustment is more pronounced. Its total HBM4 capacity has dropped from 2.3 TB to 2.0 TB, and total bandwidth from 176 TB/s to 130 TB/s, a decrease of about 26.1%; NVLink switch bandwidth has been adjusted from 28.8 TB/s to 24 TB/s.

It should be noted that the above data come from a before-and-after comparison of public specification documents, not actual application test results. Some specifications remain inconsistent across different product pages, so the relevant parameters should still be subject to NVIDIA's subsequent official updates.

At the software level, NVIDIA hopes to further expand its revenue sources from hardware sales to ongoing software and services through enterprise software licensing. Especially against the backdrop of large-scale AI Agent deployment, runtime environments, security management, and enterprise-grade support may become new business opportunities.

At the hardware level, HBM4 bandwidth, GPU interconnect, and system networking jointly affect the operating efficiency of large-scale AI workloads. In this specification adjustment, some internal bandwidth indicators have been lowered while network bandwidth has increased, indicating that system parameters at different levels are changing. For large-model training tasks that rely on frequent data exchange between GPUs, memory bandwidth and interconnect capability are particularly important; the actual impact still needs to be judged in combination with the final hardware, software optimization, and independent testing. Some analysts believe that lowering HBM4 bandwidth requirements may also help ease manufacturing pressure in the memory supply chain and improve product supply capacity. However, NVIDIA has not yet formally explained the reasons for this specification adjustment, and this speculation cannot be regarded as a confirmed adjustment goal.

Overall, NVIDIA is advancing an overall layout spanning GPU chips, complete systems, network interconnection, and enterprise software. In the future, competition in AI infrastructure will depend not only on chip computing performance but also on system coordination, software ecosystems, supply capacity, and long-term operating costs. Whether the software charging model can be implemented, and the actual performance of Vera Rubin after the adjustments, will become two directions worth watching going forward.

This article is reproduced from media reports or online platforms and represents the author's personal position or views. Our republication is solely for sharing and does not mean that we approve of or endorse it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly, and we, as a neutral platform service provider, will promptly correct, delete, or handle it in accordance with the law. Comments No user comments yet Latest News More Related Recommendations

Moore Threads × Beijing Radio and Television Station | First Fully Xinchuang 4K Ultra HD Post-Production Network Officially Put into Operation 2026-10-27 15:49

The Strategic Significance of TSMC CoWoS-S to the AI Industry Chain 2026-10-27 15:44

PCB Supply Shortage Makes 2027 Tough for PC Makers 2026-10-27 15:43