华尔街在近期的报告中指出，人工智能对内存的需求高峰预计在2027年第四季度出现。一份新的报告指出，在人工智能的需求高峰之后，人形机器人有望接棒，成为全球内存芯片的主要新需求来源。市场研究公司Counterpoint Research上周日发布的数据显示，用于人形机器人的DRAM市场规模预计将在五年内增长约20倍，从今年的8300太比特增长到2030年的172000太比特。研究指出，每个人形机器人预计需要数百GB的内存和存储容量，用于机器人的低功耗、高容量的DRAM有望成为全球内存制造新的收入来源。
Wall Street noted in a recent report that AI's peak demand for memory is expected to appear in Q4 2027. A new report points out that after AI's demand peak, humanoid robots are expected to take over as the main new source of demand for global memory chips. Data released by market research firm Counterpoint Research last Sunday shows the DRAM market size for humanoid robots is expected to grow about 20 times in five years, from 8,300 terabits this year to 172,000 terabits by 2030. The research notes each humanoid robot is expected to need hundreds of GB of memory and storage capacity, and low-power, high-capacity DRAM for robots is expected to become a new revenue source for global memory manufacturing.
人形机器人的大脑是一个计算系统，它实时处理从摄像头、激光雷达传感器和触觉传感器收集的信息，并将这些数据转化为物理动作。Counterpoint Research表示，机器人必须独立存储和处理的数据量正在急剧上升，这导致其内存需求快速增长。Counterpoint Research预计，到2030年，每台机器人的DRAM容量将从今年的19GB增加一倍以上，达到39GB。这也与一些内存厂商此前发表的观点相吻合。美光科技首席执行官Sanjay Mehrotra曾预测，到2020年代末，包括人形机器人在内的物理人工智能将成为存储器行业的主要增长驱动力。
The humanoid robot's brain is a computing system that processes information collected from cameras, lidar sensors, and tactile sensors in real time, converting this data into physical actions. Counterpoint Research says the amount of data robots must independently store and process is rising sharply, causing rapid growth in memory demand. Counterpoint Research expects that by 2030, each robot's DRAM capacity will more than double from this year's 19GB to 39GB. This also aligns with views previously expressed by some memory manufacturers. Micron Technology CEO Sanjay Mehrotra predicted that by the late 2020s, physical AI including humanoid robots will become the main growth driver for the memory industry.
他补充称，L4及以上的自动驾驶车辆的内存容量通常超过200GB，而存储容量达到数TB，每一项都比当今的L2和L3自动驾驶车辆高出一个数量级以上。不过，这也对内存厂商提出了新的技术要求。由于人形机器人依靠电池供电，并且对功耗高度敏感，分析人士认为它们将主要使用低功耗DRAM和高容量NAND闪存，而不是数据中心服务器中使用的HBM。韩国DS投资证券分析师Lee Su-rim表示，内存需求与部署的机器人数量成正比，而每台机器人中安装的内存数量也在稳步增加。人形机器人产业发展越快，存储器产业就越有可能演变成一个结构性增长的行业。
He added that L4 and above autonomous vehicles typically have memory capacity exceeding 200GB, with storage capacity reaching several TB—each more than an order of magnitude higher than today's L2 and L3 autonomous vehicles. However, this also imposes new technical requirements on memory manufacturers. Because humanoid robots rely on battery power and are highly power-sensitive, analysts believe they will mainly use low-power DRAM and high-capacity NAND flash, rather than HBM used in data center servers. Lee Su-rim, an analyst at Korea's DS Investment Securities, said memory demand is proportional to the number of robots deployed, while the amount of memory installed per robot is also steadily increasing. The faster the humanoid robot industry develops, the more likely the memory industry will evolve into a structurally growing industry.
这个判断放在前面讨论的产业脉络里，是AI算力需求向机器人转移的一个具体表现。前面讨论过英伟达投资Figure、DeepSeek接触具身智能创业者、亚马逊推出物理AI工具链，这些动作都指向同一个方向，AI的能力正在从数据中心走向物理世界。数据中心里的AI需要GPU和HBM，物理世界里的AI需要低功耗芯片和低功耗内存。人形机器人的内存需求增长，是这个转移过程在存储行业的具体体现。Counterpoint给出的20倍增长，对应的是人形机器人从千台级走向万台级甚至十万台级的过程。单台机器人的内存容量从19GB涨到39GB，对应的是机器人大脑处理的数据量在增加，感知、决策、执行这条链路上需要缓存和计算的数据越来越多。
This judgment, placed in the industrial thread discussed earlier, is a concrete manifestation of AI compute demand shifting toward robots. Earlier we discussed NVIDIA investing in Figure, DeepSeek contacting embodied intelligence entrepreneurs, and Amazon launching a physical AI toolchain—these actions all point to the same direction: AI capabilities are moving from data centers to the physical world. AI in data centers needs GPUs and HBM; AI in the physical world needs low-power chips and low-power memory. The growth in humanoid robot memory demand is a concrete manifestation of this transfer process in the storage industry. Counterpoint's 20x growth corresponds to humanoid robots moving from the thousands to the tens of thousands or even hundreds of thousands. Per-robot memory capacity growing from 19GB to 39GB corresponds to increasing data processed by the robot brain—more data needs caching and computation along the perception, decision, and execution chain.
