半导体行业观察·2026-02-14 01:37Core Viewpoint - The article discusses the emerging potential of d-Matrix, a chip startup supported by Microsoft, which aims to revolutionize AI inference by creating chips that are faster, cheaper, and more efficient than current GPU-based solutions, potentially reducing inference costs by about 90% [2][5][7]. Group 1: d-Matrix's Approach - d-Matrix focuses on designing chips specifically for inference rather than repurposing training hardware, emphasizing the architectural differences between training and inference tasks [3][5]. - The company aims to reduce latency and increase throughput by integrating memory and computation more closely, which contrasts with traditional GPU architectures that separate these functions [4][5]. - d-Matrix's chip design is modular, allowing for scalability based on workload requirements, similar to Apple's unified memory design [5][6]. Group 2: Market Dynamics - NVIDIA currently dominates the AI chip market, with a market capitalization of $4.5 trillion, but there is growing interest in alternatives as companies seek to hedge against NVIDIA's dominance [7][8]. - Several startups, including Groq and Positron, are gaining traction in the inference space, indicating a shift in the market dynamics as companies explore different memory types for faster responses [8][9]. - The competition is intensifying, with major players like OpenAI and Anthropic exploring partnerships with various chip manufacturers to enhance their AI capabilities [9][10]. Group 3: Future Outlook - d-Matrix plans to ramp up production significantly, aiming for millions of chips by the end of the year, which could position it as a key player in the AI inference market [6][9]. - The article suggests that while NVIDIA remains a formidable leader, the rapid growth of dedicated hardware for AI inference could lead to a more fragmented market where multiple players thrive [10].