「AI新世代」从GPU到LPU：英伟达大举进攻推理芯片，黄仁勋再落关键一子

Core Insights - NVIDIA is making a significant move into the inference chip market with the introduction of the Groq 3 LPU, as announced by CEO Jensen Huang at GTC 2026 [1] - The AI industry is shifting focus from model training to inference, with NVIDIA aiming to capture this market opportunity [1][6] - By the end of 2027, NVIDIA's Blackwell and Rubin product lines are projected to generate annual revenues of $1 trillion, doubling previous forecasts [1] Group 1: Product Launch and Features - NVIDIA has officially launched the Vera Rubin platform, which includes seven chips, such as Rubin GPU, Vera CPU, and the new Groq 3 LPU, designed to enhance AI inference capabilities [2] - The Groq LPU is expected to increase token throughput from 100 tokens per second to over 1500 tokens, supporting interactive AI agent scenarios [2] - A new rack, Groq LPX, has been introduced to accommodate the Groq accelerator, enhancing decoding performance for AI models [2] Group 2: Market Trends and Strategic Positioning - NVIDIA's interest in the inference chip market has been long-standing, highlighted by its $20 billion acquisition of Groq's core technology assets in December 2025 [3] - The market share of non-GPGPU chips in AI servers is expected to rise from 36% in 2024 to 45% by 2027, while GPGPU chip share will decline from 64% to 55% [3] - The shift in AI computing demand from training to inference is a strategic response by NVIDIA to market changes and competitive pressures [3][8] Group 3: Ecosystem and Infrastructure Development - NVIDIA is addressing the growing demand for inference with new initiatives, including a partnership with OpenAI for specialized inference chips [4] - The company has introduced the Vera Rubin DSX AI Factory reference design, which outlines how to build and operate AI factory infrastructure for optimal performance [7] - NVIDIA's advancements in AI infrastructure aim to maximize productivity and energy efficiency in generating AI tokens [7] Group 4: Competitive Landscape and Future Outlook - The introduction of the LPU does not imply a decline in NVIDIA's GPU business; rather, it is expected to create broader market opportunities through synergy [7] - The ASIC market is becoming increasingly competitive, with several challengers emerging, including Cerebras and Chinese companies like Cambricon and Huawei [8] - The entry of NVIDIA into the inference chip sector is seen as both a challenge and a catalyst for domestic manufacturers, potentially accelerating industry reshuffling and technological upgrades [8]