AI算力上市潮中，谁是幕后最大赢家？这家"独中四元"的机构，如何用十年完成这场精准伏击？答案 都在这里。 作者丨 巴里 编辑丨 关雎 图源 丨Midjourney 2015年，当深度学习算法在ImageNet竞赛中一鸣惊人时，大多数人看到的是AI应用的曙光，而联想 创投看到的，是一场即将到来的、更为底层的算力革命。 洪荒时代的"冒险者" 时间拨回2017年。AI芯片在中国还停留在实验室论文和少数巨头的内部项目。市场上最炙手可热的 是O2O和共享经济。 就在这个时间点，联想集团提出了"端-边-云-网-智"的新IT战略蓝图。在这幅蓝图中，联想创投扮 演着"科技瞭望塔"的角色。 他们很早就形成了一个核心判断：智能世界的万物生长，都依赖于一颗强大的"心脏"——算力。未来 的大计算将由应用驱动，从云端的数据中心到边缘的智能设备，对算力的需求将在十年内增长百倍。 然而，这颗"心脏"的形态将发生根本性变革：它不再仅仅是中心化、通用化的，而将变得多元化、专 属化，并无处不在。正是基于这一超前的产业洞察，联想创投没有随波逐流，而是在AI算力尚处洪荒 时代时，便开始了系统性布局。 投资，尤其是早期投资，最忌讳的是盲目类比和追逐风 ...

Group 1: Market Growth and Trends - The AI ASIC market is expected to grow from $12 billion in 2024 to $30 billion by 2027, with a compound annual growth rate (CAGR) of 34% [1] - The demand for AI servers is driving major cloud service providers in the US to accelerate the development of ASIC chips, with new products being launched every 1-2 years [2] - In China, the market share of imported chips is projected to drop from 63% in 2024 to about 42% by 2025 due to new export control policies, while domestic chip manufacturers' market share is expected to rise to 40% [2] Group 2: ASIC Technology and Performance - AWS's Trainium 2 ASIC chip can complete inference tasks faster than NVIDIA's H100 GPU, with a cost-performance improvement of 30%-40% [3] - Google's TPU has become a typical representative of ASIC technology, with the latest version, Ironwood, capable of achieving 42.5 exaflops of AI computing power [5][6] - The Ironwood chip features significant enhancements in memory and bandwidth, with each chip equipped with 192GB of high-bandwidth memory [6] Group 3: Competitive Landscape - Broadcom holds a market share of 55%-60% in the ASIC market, with AI-related revenue reaching $4.1 billion, a 77% year-on-year increase [7] - Marvell's ASIC business is a core growth driver, with data center business accounting for approximately 75% of its revenue [9] - Domestic companies like Cambricon and Baidu are actively developing their own ASIC chips, with Baidu's Kunlun chip outperforming traditional GPUs in terms of cost and performance [11][12] Group 4: Challenges and Considerations - The increasing costs of advanced chip design pose challenges for companies looking to develop their own ASICs, with TSMC's 2nm wafers costing around $30,000 each [15] - The question arises whether every company truly needs its own CPU, given the high costs associated with chip design [16]