--- title: "TrendForce: CSP self-developed ASIC scale upgrade, NVIDIA diversifies product line to target AI training and inference demand" type: "News" locale: "en" url: "https://longbridge.com/en/news/279607160.md" description: "TrendForce's latest research shows that NVIDIA will shift its focus to AI inference applications at the GTC 2026 conference, promoting a variety of products such as GPUs, CPUs, and LPUs to meet the demands of AI training and inference. It is expected that by 2030, the shipment proportion of ASIC AI Servers will rise from 27.8% to nearly 40%. NVIDIA is consolidating its leadership position in the AI market by launching integrated solutions like GB300 and VR200, with an expected shipment share of 80% by 2026" datetime: "2026-03-18T13:11:44.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/279607160.md) - [en](https://longbridge.com/en/news/279607160.md) - [zh-HK](https://longbridge.com/zh-HK/news/279607160.md) --- # TrendForce: CSP self-developed ASIC scale upgrade, NVIDIA diversifies product line to target AI training and inference demand According to the latest AI Server research by TrendForce, with major cloud service providers (CSPs) increasing their efforts in self-developed chips, NVIDIA (NVDA.US) has shifted its focus at the GTC 2026 conference to the implementation of AI inference applications across various fields, differing from its previous emphasis on the cloud AI training market. By promoting a diverse product line of GPUs, CPUs, and LPUs to address AI training and inference demands, and leveraging Rack integration solutions to drive supply chain growth. TrendForce indicates that as the trend of self-developed chips led by CSPs such as Google (GOOGL.US) and Amazon (AMZN.US) expands, it is estimated that the shipment proportion of ASIC AI Servers will rise from 27.8% in 2026 to nearly 40% by 2030. ![Image](https://imageproxy.pbkrs.com/http://img.zhitongcaijing.com/images/contentformat/6ed35bc8a471d3068ddce07dac4999d1.jpg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg) To consolidate its leadership position in the AI market, one of NVIDIA's strategies is to actively promote integrated cabinet solutions like the GB300 and VR200, emphasizing scalability for AI inference applications. The Vera Rubin announced at this GTC is defined as a highly vertically integrated complete system, encompassing seven chips and five cabinets. Observing the supply chain progress for Rubin, it is expected that in the second quarter of 2026, memory manufacturers will be able to provide HBM4 for use with Rubin GPUs, assisting NVIDIA in shipping Rubin chips successively around the third quarter. As for the shipment progress of NVIDIA's GB300 and VR200 Rack systems, the former is expected to replace the GB200 as the main product in the fourth quarter of 2025, with an estimated shipment proportion reaching nearly 80% by 2026, while the VR200 Rack is anticipated to gradually release shipping capacity by the end of the third quarter of 2026, with further developments depending on the actual progress of ODMs. Additionally, as AI transitions from generation to the era of agent models, there are significant delays and memory bandwidth bottlenecks during the decoding phase of generating tokens. To address this, NVIDIA has integrated Groq team technology to launch the Groq 3 LPU, designed specifically for low-latency inference, featuring 500MB SRAM per unit and up to 128GB for the entire cabinet. However, the memory capacity of the LPU itself cannot accommodate the vast parameters and KV Cache of the Vera Rubin level. Therefore, NVIDIA proposed a "Disaggregated Inference" architecture at this GTC, which divides the inference pipeline into two parts through an AI factory operating system called Dynamo: when processing agent-type AI, a large number of mathematical operations and the storage of a vast KV Cache during the Pre-fill and Attention computation stages will be handled by Vera Rubin, which possesses extremely high throughput and massive memory Due to bandwidth limitations and extreme sensitivity to latency, the decoding and token generation stages are directly offloaded to the LPU cabinet, which has been expanded with massive storage. In terms of supply chain progress, the third-generation Groq LP30 is being manufactured by Samsung and has entered full-scale production, with official shipments expected in the second half of 2026. There are also plans to launch a more powerful LP40 chip in the next-generation Feynman architecture ### Related Stocks - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [XSD.US](https://longbridge.com/en/quote/XSD.US.md) - [IXN.US](https://longbridge.com/en/quote/IXN.US.md) - [SMH.US](https://longbridge.com/en/quote/SMH.US.md) - [NVDU.US](https://longbridge.com/en/quote/NVDU.US.md) - [NVDL.US](https://longbridge.com/en/quote/NVDL.US.md) - [PSI.US](https://longbridge.com/en/quote/PSI.US.md) - [NVDY.US](https://longbridge.com/en/quote/NVDY.US.md) - [SOXX.US](https://longbridge.com/en/quote/SOXX.US.md) - [XLK.US](https://longbridge.com/en/quote/XLK.US.md) - [SOXL.US](https://longbridge.com/en/quote/SOXL.US.md) - [NVDX.US](https://longbridge.com/en/quote/NVDX.US.md) - [07788.HK](https://longbridge.com/en/quote/07788.HK.md) - [07388.HK](https://longbridge.com/en/quote/07388.HK.md) - [NVDD.US](https://longbridge.com/en/quote/NVDD.US.md) - [NVDQ.US](https://longbridge.com/en/quote/NVDQ.US.md) ## Related News & Research - [The AI Stock Wall Street Can't Stop Talking About in 2026](https://longbridge.com/en/news/282407351.md) - [Korean AI chip startup DEEPX, Hyundai work on robots powered by generative AI](https://longbridge.com/en/news/282774224.md) - [Acer Veriton GN100 to Power The Spark Hack Series - New York with New Capabilities](https://longbridge.com/en/news/282191048.md) - [Wall Street Financial Group Inc. Acquires 2,961 Shares of NVIDIA Corporation $NVDA](https://longbridge.com/en/news/282537372.md) - [Siemens Accelerates AI Chip Verification to Trillion‑Cycle Scale with NVIDIA Technology](https://longbridge.com/en/news/282222056.md)