NVIDIA: Rubin CPX debuts! Ready to arm wrestle with Broadcom ASIC
In the recent scenario where Broadcom is advancing rapidly with ASIC, the GPU camp led by NVIDIA is evidently facing challenges. Broadcom's stock price surged nearly 10% on the day of the earnings release, while $NVIDIA(NVDA.US) and AMD fell by 3% and 6% respectively. Especially, Broadcom's billion-dollar order from its fourth customer has directly brought expectations to the market for ASIC to capture the AI core chip market.
Comparing the relevant revenue of core chip manufacturers in the current data center market, Broadcom's share in the AI chip market has reached nearly 10%, surpassing Intel and AMD. Broadcom currently has "3 mass production customers + 4 potential customers," and as the products of potential customers enter mass production (such as the fourth customer), it is expected to bring more increments to the company.
In the context of Broadcom's continuous pursuit with ASIC, NVIDIA suddenly announced the launch of the "Rubin CPX" GPU at the AI Infra Summit, a new generation GPU specifically designed for massive context processing. This processor is dedicated to enabling AI systems to handle tasks with millions of tokens, including software coding and video generation, with breakthrough speed and efficiency.
For this "Rubin CPX" GPU new product, NVIDIA mainly highlighted three aspects:
a) Computing power and memory: Computing power up to 30 PFLOPS (NVFP4 precision); equipped with 128GB cost-effective GDDR7 memory, accelerating high-demand context workloads.
Although the computing power is lower than the 50 PFLOPS (FP4 precision) of the Rubin GPU released in March, this dedicated GPU provides 3 times faster attention mechanism compared to NVIDIA GB300 NVL72 system.
The company plans to launch this product at the end of 2026 after the regular Rubin GPU is released in 2026.
b) Configuration and collaboration: Rubin CPX can work in conjunction with NVIDIA Vera Rubin NVL144 CPX platform's NVIDIA Vera CPU and Rubin for generation phase processing, forming a complete high-performance decomposed service solution.
NVIDIA Vera Rubin NVL144 CPX—integrating 36 Vera CPUs, 144 Rubin GPUs, and 144 Rubin CPX GPUs, providing 8 EFLOPs of NVFP4 computing power, which is 7.5 times that of GB300 NVL72. A single rack can provide 100 TB of high-speed memory and 1.7 PB/s memory bandwidth.
Rubin CPX offers multiple configurations, including Vera Rubin NVL144 CPX, which can be combined with NVIDIA Quantum-X800 InfiniBand horizontal expansion computing architecture or NVIDIA Spectrum-XGS Ethernet technology and ConnectX-9 SuperNIC's Spectrum-X Ethernet network platform.
Vera Rubin NVL144 CPX can achieve unprecedented scale monetization, with a $100 million investment yielding $5 billion in token revenue.
c) Positioning of "dedicated GPU": Rubin CPX is specifically used for context and pre-fill computing, significantly enhancing the performance related to massive context. The original Rubin/Rubin Ultra GPU is responsible for the generation phase of computing.
Since the inference phase is often divided into context phase and generation phase, the former's limitation lies in computing power, requiring high throughput to process and analyze massive data; the latter's limitation lies in memory bandwidth, relying on memory transmission and high-speed interconnection to maintain output performance.
NVIDIA's launch of the dedicated Rubin CPX GPU this time is mainly aimed at improving the speed of the context phase in the case of massive data. In the specific Rubin system, optimization is achieved through the combination of GPU + dedicated GPU. CPX undertakes the work of the context phase, using GDDR7 memory, which can meet the demand without the need for HBM memory. The original Rubin GPU is still responsible for the generation phase.
Considering (a+b+c), Dolphin Research believes that NVIDIA's sudden announcement of Rubin CPX is mainly a response to Broadcom's ASIC competition. Previously, the market had speculated about NVIDIA's possible layout in the ASIC direction, and the announcement of Rubin CPX is also a clear answer.
Although Rubin CPX is still within the GPU category, from the specific responsibilities, it is mainly designed for inference, playing a role in accelerating the context phase. With the release of Rubin CPX, the product form in the Rubin phase of the company is also clearer. The NVIDIA Vera Rubin NVL144 CPX equipped with "CPU+GPU+CPX GPU" will be the main product in the Rubin phase, and specific customers will mainly target core cloud vendors and other large companies, which are also the main target customers of ASIC.
Compared to NVIDIA's GPU, ASIC is gradually binding cooperation with core cloud vendors due to its cost advantage. More notably, Broadcom already has 7 customers/potential customers (including Google, Meta, ByteDance, etc.). Large cloud service providers are the largest buyers of cloud services and AI chips, and if Broadcom is allowed to develop, it will inevitably affect NVIDIA's growth space in the future.
The company's release of the Rubin CPX dedicated to "inference" is a direct response to Broadcom and other ASIC products. Previously, some customers tried to customize ASIC chips from the "cost-effectiveness" perspective and achieved good results. The launch of Rubin CPX can also play a role in "cost reduction and efficiency improvement" for large-scale cloud vendors. Instead of facing the uncertainty of self-research, adopting NVIDIA's CPX and a complete Rubin solution is also an option.
In the context of core cloud vendors increasing capital expenditure, Broadcom recently received a billion-dollar order and Oracle's remaining contract value also soared by 300 billion. All signs indicate that the current major American companies are infrastructure maniacs. In the context of strong AI demand, the entire market cake is getting bigger and bigger. The "GPU vs ASIC" issue between NVIDIA and Broadcom will return to who can ultimately get a bigger piece of the cake.
<End here>
Dolphin Research's recent financial report review of NVIDIA and Broadcom (AVGO.O):
September 5, 2025 Broadcom conference call "Broadcom (Minutes): AI revenue growth will exceed 60% in the next fiscal year"
September 5, 2025 Broadcom earnings review "NVIDIA, AMD lackluster? Broadcom boldly takes over as AI flag bearer"
September 1, 2025 NVIDIA IR meeting "NVIDIA (small meeting): 2026 target relatively cautious, 2030 market reaches 3-4 trillion"
August 28, 2025 NVIDIA conference call "NVIDIA (Minutes): GB300 has been shipped, Rubin will be mass-produced as planned next year"
August 28, 2025 NVIDIA earnings review "NVIDIA: Universe's top stock, not explosive enough is a crime?"
Risk disclosure and statement of this article: Dolphin Research Disclaimer and General Disclosure
The copyright of this article belongs to the original author/organization.
The views expressed herein are solely those of the author and do not reflect the stance of the platform. The content is intended for investment reference purposes only and shall not be considered as investment advice. Please contact us if you have any questions or suggestions regarding the content services provided by the platform.
