重磅！黄仁勋罕见出手，欧美AI芯片独角兽集体谢幕

Core Viewpoint - The article discusses the strategic partnership between Groq and NVIDIA, highlighting the shift in the AI chip landscape, particularly in the context of Groq's technology and the broader implications for the AI chip industry [4][8][12]. Group 1: Partnership Details - Groq announced a non-exclusive licensing agreement with NVIDIA, allowing NVIDIA to utilize Groq's inference technology, which aims to expand the application of high-performance, low-cost inference technology [4]. - Groq's team members, including co-founders Jonathan Ross and Sunny Madra, will join NVIDIA to help scale the licensed technology [4][9]. - The agreement is not a full acquisition; NVIDIA is paying for the technology license rather than purchasing Groq outright [8][9]. Group 2: Financial Aspects - Reports suggest that NVIDIA may have agreed to a $20 billion (approximately 140.2 billion yuan) deal for Groq's assets, although this figure has not been confirmed by either party [7][38]. - Groq's revenue expectations have been significantly revised downwards, with projected 2025 revenue reduced from $2 billion (approximately 14 billion yuan) to $500 million (approximately 3.5 billion yuan) [16]. - Groq's revenue for the previous year was reported at $90 million (approximately 600 million yuan), with future projections indicating growth to nearly $1.2 billion (approximately 8.4 billion yuan) by 2026 and over $1.9 billion (approximately 13.3 billion yuan) by 2027 [16]. Group 3: Technology Insights - Groq's custom AI inference chip, LPU, claims to run large language models faster than GPUs and can achieve up to 10 times the energy efficiency of GPUs [21]. - The LPU architecture is designed with four core principles: software-first, programmable streaming architecture, deterministic computation and networking, and on-chip memory [23][24]. - The on-chip SRAM memory bandwidth of LPU exceeds 80 TB/s, significantly outperforming the 8 TB/s bandwidth of GPU's external HBM, contributing to its performance advantages [28]. Group 4: Industry Context - The article notes a trend of consolidation in the AI chip market, with many startups facing challenges in scaling independently, leading to increased acquisition activity among major tech companies [12][46]. - The fate of the "four AI chip unicorns" in the West has diverged, with some being acquired and others struggling, reflecting a shift in the market dynamics for AI chips [43][45]. - The article emphasizes that AI inference will become a primary battleground for commercial AI, with companies needing to focus on system efficiency and software collaboration to remain competitive [49][50].