高通AR1 Gen2将更新，留给中国AI眼镜芯片商的时间不多了

摘要：如果中国产AI眼镜芯片不能在2025年Q2前形成规模可用方案，等AR1 Gen2铺开后，整个市场将再度被锁死两年。 评论：据业内消息，高通计划在2025年上半年推出AR1 Gen2，工艺从4nm升级至3nm，图像ISP支持更高帧率、更强人脸识别与多目标追踪，NPU算力也将同步提升，用于复杂环境感知、AR叠加与实时AI Agent处理。 Meta第三代Ray-Ban智能眼镜预计将首发搭载AR1 Gen2，而中国市场若无可比替代方案，将继续处于“被高通定义”的状态。 当大多数人还在质疑AI眼镜是否是“伪需求”时，Meta用200万台Ray-Ban智能眼镜的销量，悄然为整个行业划出了生死分界线。支撑其爆款背后的，是高通专为AI眼镜打造的AR1芯片平台。 如今，这枚AR1芯片即将迭代升级，AR1 Gen2预计在2025年登场，工艺从4nm提升至3nm，AI推理与视觉处理能力大幅跃升。对于那些仍在打磨样片、调试算法、苦寻客户的中国芯片厂商而言，这更像是一记倒计时警钟：留给中国产AI眼镜芯片突围的窗口，正在迅速关闭。 高通AR1：从“存在感稀薄”到AI眼镜的事实标准2023年9月，Meta与高通联合发布第二代Ray-Ban Meta智能眼镜，搭载AR1 Gen1平台。这款产品具备AI识图、音频交互、第一视角拍摄、照片生成等能力，并通过LLaMA模型接入ChatGPT与Whisper，在硬件与AI之间构建起完整闭环。 它的成功超出了所有人预期： 三个季度出货超过200万台； 占据2024年全球AI智能眼镜销量的95.7%； 成为首款真正“出圈”的AI穿戴设备。 Meta的成功也带火了AR1。在AI眼镜仍处于概念模糊、市场教育早期的阶段，AR1成为OEM厂商最容易、最安全的选择。无论是雷鸟V3、Rokid Glasses，还是小米刚发布的AI眼镜，几乎清一色地押注高通AR1。 Wellsenn XR数据显示，AR1芯片在整机BOM中的占比高达31.6%，约为55~60美元，几乎锁死了千元以下市场的空间。然而，厂商依然愿意选择它——因为这是目前为数不多“能用”“稳定”“出过货”的AI眼镜主控方案。小米退货率高达40%：不只是体验问题小米AI眼镜发布之初，以雷军亲自站台、1999元起的定价、24小时破万台的销量，一度被视为AI眼镜平价化、规模化的重要节点。但很快，用户体验的真相暴露： 拍照延迟、语音识别卡顿； 配戴夹头、续航不到一天； 文件无法直接在设备中删除； 高达40%的退货率。 这些问题背后，其实并不只是软件没打磨好，而是芯片架构本身就“绑手绑脚”。 高通AR1虽强，但它并非真正意义上的“单芯片方案”，需要外挂ISP、协处理器和多路电源管理器件，导致发热大、板子厚、续航差。这对必须追求“全天候佩戴”的眼镜产品来说，几乎是原罪。 Rokid也遇到了同样的问题。联合蓝思科技量产的25万台AI+AR眼镜，在用户手中迟迟无法交付，京东页面的发货时间被延迟至“2025年Q3”。核心原因之一，依然是硬件体系未能完全准备好，尤其是在芯片平台集成度与供应节奏上。恒玄、全志的突围之路中国产厂商也不甘心。恒玄科技早在2023年就开始开发AI眼镜专用芯片BES2800，一度被视为“中国产破局者”。 这颗芯片尝试将图像ISP、语音唤醒、AI推理能力集成在一个SoC中，实现“拍+算+传”的一体化设计。但现实很残酷： 初版仅支持800万像素，且未内建完整ISP； 拍照模块依然需外挂视觉芯片，功耗难控； 样片打磨数月，未能找到大客户下单。 据业内人士透露，小米虽参与了BES2800的初期讨论，但并未最终采用，华为也未将其纳入2025年新品计划。这颗芯片目前仍停留在“方案测试”阶段，尚未进入真正的量产周期。 另一边，全志科技则走了“弯道超车”路线。今年上半年，在深圳悄然发布了AI眼镜专用SoC——V821与V881，定价分别为3美元与5美元，远低于AR1。 其中，V821采用22nm制程，集成RISC-V CPU、64MB DDR2、Wi-Fi直连、摄像头接口和基础ISP，主打5~8MP级别低功耗拍照+语音交互。现场样机售价仅200元，面向的是白牌市场，目标是快速出货、抢占海外电商。 但这种“白牌突围”，也面临诟病： “用低功耗手表芯片拼接摄像头、扬声器，这种产品撑不起生态，只会劝退用户。”一位从业者直言。 简而言之，恒玄在往高走，全志在往快走，而真正可比肩AR1的平台，至今未现身。AI眼镜芯片的三条技术路线当前AI眼镜芯片的技术演进，大致形成三类路径： SoC单芯片集成路线：高通AR1、展锐W517 ，特点是成本高，体积大，功耗高，需更先进工艺 SoC + 协处理器双芯片路线：AR1 + 恒玄2700，特点是动态分工，系统复杂，适配困难 主控 + 外挂ISP路线：BES2800 + SSC309QL，特点是 成本低，功耗高，图像能力瓶颈 其中，真正的“终极形态”应是：一颗SoC同时具备图像ISP、NPU、语音唤醒、连接能力与AI推理，满足动态成像、低延迟响应与长时间续航。 然而，这需要： 芯片厂商自研NPU与ISP； 工艺制程达到7nm以下； 全栈算法适配与软件平台； 与模组厂、ODM厂协同验证； 这不是一颗芯片可以完成的工程，而是一整条从底层IP、工具链、生态到整机设计的长周期系统战。高通AR1 Gen2已在路上，留给中国芯片的时间不多了据业内消息，高通计划在2025年上半年推出AR1 Gen2，工艺从4nm升级至3nm，图像ISP支持更高帧率、更强人脸识别与多目标追踪，NPU算力也将同步提升，用于复杂环境感知、AR叠加与实时AI Agent处理。 Meta第三代Ray-Ban智能眼镜预计将首发搭载AR1 Gen2，而中国市场若无可比替代方案，将继续处于“被高通定义”的状态。 “我们不是没技术，而是没人敢先下10万颗流片。”一位芯片企业负责人苦笑道，“没有头部品牌先站出来试错，中腰部客户谁都不敢赌。” 换言之，如果中国产芯片不能在2025年Q2前形成规模可用方案，等AR1 Gen2铺开后，整个市场将再度被锁死两年。Qualcomm AR1 Gen2 Update: Time is Running Out for Chinese AI Glass ChipmakersAbstract:If Chinese-made AI glass chips cannot develop scalable solutions by Q2 2025, the entire market will be locked down again for the next two years after AR1 Gen2 is rolled out. While many still question whether AI glasses are a "pseudo-demand," Meta has quietly drawn a clear line between life and death for the entire industry with the sales of 2 million units of Ray-Ban smart glasses. Behind this successful product is Qualcomm's AR1 chip platform, specifically designed for AI glasses. Now, this AR1 chip is about to be upgraded. The AR1 Gen2 is expected to launch in 2025, with the manufacturing process shifting from 4nm to 3nm, significantly improving AI inference and visual processing capabilities. For Chinese chip manufacturers still refining prototypes, debugging algorithms, and desperately seeking customers, this upgrade serves as a countdown warning: the window for Chinese AI glasses chipmakers to break through is rapidly closing.Qualcomm AR1: From "Low Visibility" to the De Facto Standard for AI GlassesIn September 2023, Meta and Qualcomm jointly launched the second-generation Ray-Ban Meta smart glasses, equipped with the AR1 Gen1 platform. This product supports AI image recognition, audio interaction, first-person photography, and photo generation, and integrates with the ChatGPT and Whisper models through the LLaMA framework, establishing a complete closed loop between hardware and AI. Its success exceeded all expectations: Over 2 million units shipped in three quarters; Dominating 95.7% of the global AI smart glasses sales in 2024; Becoming the first truly mainstream AI wearable device. Meta’s success also boosted the AR1 chip. At a time when AI glasses were still in the early stages of market education with unclear concepts, AR1 became the easiest and safest choice for OEM manufacturers. Whether it’s Thunderbird V3, Rokid Glasses, or Xiaomi’s newly launched AI glasses, they almost all rely on Qualcomm AR1. According to Wellsenn XR, the AR1 chip accounts for 31.6% of the total bill of materials (BOM) for these devices, priced at approximately $55–$60, essentially locking out the sub-thousand-yuan market. Yet, manufacturers still prefer it because it’s one of the few AI glasses main control solutions that is “usable,” “stable,” and has been mass-produced.Xiaomi’s Return Rate Reaches 40%: It’s Not Just About the Experience IssuesWhen Xiaomi first launched its AI glasses, featuring a personal endorsement from Lei Jun, a starting price of 1999 RMB, and a 24-hour sales volume exceeding 10,000 units, it was seen as a significant milestone for the affordability and scaling of AI glasses. However, the reality about user experience soon surfaced: Delays in photography and lagging voice recognition; Pressure on the temples and less than a day of battery life; Unable to delete files directly on the device; A return rate as high as 40%. These issues aren’t just due to underdeveloped software, but because the chip architecture itself is "handicapped." Although Qualcomm AR1 is powerful, it is not truly a "single-chip solution." It requires an external ISP, co-processors, and multiple power management components, which lead to high heat generation, thick boards, and poor battery life. This is almost a fatal flaw for a product that must achieve "all-day wearability." Rokid faced similar issues. The 250,000 units of AI+AR glasses, jointly mass-produced with Lens Technology, still haven't been delivered to users. The delivery date on JD's platform was delayed to "Q3 2025." One of the core reasons is that the hardware system wasn’t fully ready, especially in terms of chip platform integration and supply rhythm.The Breakthrough Path of bestechnic and AllwinnerChinese manufacturers are not giving up. bestechnic started developing the AI glasses-specific chip BES2800 in 2023, which was once seen as the "Chinese breakthrough." This chip attempted to integrate image ISP, voice wake-up, and AI inference into a single SoC, achieving an integrated design of "photography + computing + transmission." But reality proved harsh: The initial version only supports 8MP and lacks a complete ISP; The photography module still requires an external vision chip, and power consumption is hard to control; Months of prototyping, but no major customers have placed orders. Industry insiders reveal that although Xiaomi was involved in the early discussions of BES2800, it ultimately didn’t adopt it, and Huawei also did not include it in their 2025 product plan. This chip is still in the "solution testing" stage and has not entered the mass production cycle yet. On the other hand, Allwinner Technology took the "bypass" approach. Earlier this year, they quietly released two AI glasses-specific SoCs—V821 and V881—in Shenzhen, priced at $3 and $5 respectively, much lower than the AR1. The V821 uses a 22nm process, integrating an RISC-V CPU, 64MB DDR2, Wi-Fi direct connection, camera interfaces, and basic ISP, targeting the white-label market. The goal is to quickly ship and dominate overseas e-commerce. However, this “white-label breakthrough” also faces criticism: "Using low-power smartwatch chips to assemble cameras and speakers, this kind of product cannot support an ecosystem and will only discourage users." In short, bestechnic is aiming for higher performance, while Allwinner is aiming for faster shipments, but a truly comparable platform to AR1 has yet to emerge.Three Technological Routes for AI Glass ChipsCurrently, the technological evolution of AI glasses chips has broadly formed three paths: SoC single-chip integration: Qualcomm AR1, UNISOC W517, characterized by high cost, large size, high power consumption, and the need for more advanced processes. SoC + co-processor dual-chip approach: Qualcomm AR1 + BES 2700, characterized by dynamic division of labor, complex systems, and difficult adaptation. Main control + external ISP approach: BES2800 + SSC309QL, characterized by low cost, high power consumption, and image capability bottlenecks. The true "ultimate form" should be: a single SoC that integrates ISP, NPU, voice wake-up, connectivity, and AI inference, meeting dynamic imaging, low-latency responses, and long battery life. However, this requires: Chip manufacturers to develop their own NPU and ISP; Process technology reaching below 7nm; Full-stack algorithm adaptation and software platforms; Collaboration with module manufacturers and ODM factories for validation. This is not a project that can be completed by just one chip but rather requires a long-cycle system strategy involving the entire ecosystem, from IP, toolchains, to complete machine design.Qualcomm AR1 Gen2 is Coming, and Time is Running Out for Chinese ChipsIndustry sources reveal that Qualcomm plans to release the AR1 Gen2 in the first half of 2025. The process will be upgraded from 4nm to 3nm, with enhanced image ISP support for higher frame rates, stronger facial recognition, and multi-target tracking. The NPU processing power will also be upgraded for complex environmental sensing, AR overlay, and real-time AI agent processing. Meta's third-generation Ray-Ban smart glasses are expected to debut with the AR1 Gen2. Without a comparable alternative from China, the Chinese market will continue to be “defined by Qualcomm.” “We don’t lack technology, but no one dares to commit to 100,000 wafer runs,” said one chip company executive with a wry smile. "Without a leading brand stepping up to test the waters, mid-tier customers won’t dare to bet on it." In other words, if Chinese-made chips cannot form scalable solutions by Q2 2025, once the AR1 Gen2 is released, the market will once again be locked down for two more years.