文 | 光子星球
今年以来，AI眼镜似乎撞上了第一堵墙。
IDC数据显示，2026年第二季度，国内智能眼镜出货约61.1万台，同比下降8.0%，是这条赛道狂飙以来的首次负增长。而同一时间，全球市场还在以35.3%的速度飞驰。
海外热火朝天，国内先打了个寒颤。
在此之前，没有哪个品类，比AI眼镜更相信自己会是“下一代手机”；也没有哪个品类，被这份相信，绊得如此之狠。
厂商为其All in，资本为其扎堆，发布会一场接一场，仿佛AI眼镜取代手机只是时间问题。可热闹全在供给端，需求端却显得有些不温不火，天天把眼镜戴出门的，似乎翻来覆去就是那批喜欢技术、乐于尝鲜的人。
品牌一年比一年多，新品一副接一副，但用户圈层并没有迎来爆发——AI眼镜太多，极客似乎快不够用了。
于是，曾经把“人人都戴”挂在嘴边的玩家，开始了一场集体后撤：有些玩家秉持着减法，试图把产品做回普通眼镜；有些玩家则一头扎进了车载、翻译、会议、无障碍等窄缝。
减法和窄缝能让行业活下来，却喂不饱他们对入口的胃口。在此背景下，这批玩家一边退守，一边又把筹码，悄悄压在了Personal Agent身上。
从撞墙、退守，再到当下的反攻，前两幕戏已渐渐落幕，而第三幕，似乎才刚开场。
撞墙：AI眼镜太多，极客快不够用了
尽管单季数据受新品发布、上市节奏影响，可能出现短期波动；仅凭一个季度的下滑，就给行业判刑也不公平。
但一个被寄予高增长预期的赛道，在厂商密集发布、营销不断加码的当口，作为基本盘的中国市场却不增反降，终究透着一股虚火的意味。
从产品侧来看，扩张确实有目共睹。两年里，AI眼镜重量从五十克压到三十几克，芯片从一颗堆到两三颗，功能从十几个堆到两百多个。
这场狂热的一大催化剂，是资本。
投资人对“入口”两个字向来饥渴，也愿意为一个尚未兑现的未来提前下注；但掏钱的消费者却冷静得多，他们打量产品，至今用的还是普通消费电子的标准。
品类早期，先把钱备足、把身位卡住，虽是一种活法，但问题在于，AI眼镜上游高度成熟，主控、光波导、传感器方案高度雷同，除了极少数有真正自研能力的玩家，品牌之间真正能拉开差距的地方并不算多。
所谓百镜大战，打到消费者眼前，只剩配件、功能的排列组合，以及定价单上的彼此试探。
从需求侧来看，愿意天天把AI眼镜戴出门的，翻来覆去就是同一批人，即科技极客、数码博主、硬件发烧友。
艾媒咨询数据显示，AI眼镜的主要消费者是AI爱好者和科技尝鲜派，分别占比32.5%和32.2%，合计约65%。
这群人非但不吝啬，反而最舍得掏钱，碰上新品就手痒，有人手上同时囤着好几副，有人出掉旧的、再买新的。
但这一小撮人，并不能代表普罗大众。
产品思维里有个原则，叫“防呆”，即设计时默认用户是“傻子”，不要求用户读透说明书，也能用好产品。越是成功的大众产品，越敢把用户想得“笨”，让操作尽可能符合直觉。
按这个标准看，很多所谓的AI眼镜，几乎是个反面教材。
一位从业者告诉光子星球，此前的产品，配重和框型默认照着男性的头做，头型更窄、对重量更敏感的女生，基本被排除在外，出现“割耳”等问题。直到厂商们逐渐意识到了这个问题，才纷纷在后续产品里添加镜腿角度调节的功能。
XREAL创始人徐驰，也曾在一次采访中表示，很多国内厂商并不“专业”，对一个复杂硬件品类缺乏足够的耐心与尊重——看到前沿玩家拿出了产品，便迅速跟进、拆解、复刻，再用供应链组合把产品“实现”出来，但实际上对底层问题的理解并不深。
而就现阶段AI眼镜高企的退货率来看，AI眼镜的整体体验仍不够厚重。
尽管很多发烧友，享受着参与改善一项新技术的乐趣，主动为找BUG、想场景、提反馈，但大众没有这个耐心，也没有这个义务。他们要的是开箱即用、不用伺候的产品，一旦产品需要其主动适应、容忍失误、还得自己找理由戴上，他的选择往往很简单——不戴。
资本催熟的供给，撞上还没打开的需求，这样的繁荣，自然撑不起“下一代手机”的分量。
极客这口井见底，第一堵墙也就竖在了所有厂商面前。
退守：大入口，先做小生意
撞墙之后，行业做了一件诚实的事，悄悄把目标人群，从“所有人”改成了“某一类人”。曾经张口闭口下一代手机，如今退回去，老老实实地做垂直人群、场景的专用工具。
这是被现实逼出来的清醒。
退守的路径，归结起来有两条，一条往产品上做减法，一条往场景里做垂直。
出货第一的雷鸟，在其雷鸟iO上动了最狠的两刀，把摄像头和扬声器都砍了，显示退回单绿色，整机压到34克。
对很多用户而言，摄像头、扬声器都是AI眼镜的标配，少了就不完整。雷鸟iO发布时，社交媒体上很多用户因此被劝退，嫌弃其“阉割”。
尽管如此，雷鸟iO还是卖爆了。
逻辑很简单，即承认它首先得是一副眼镜，再谈别的。而扬声器这块，TWS耳机比它做得更好；摄像头方面，多模态还没到能把第一视角真正用起来的地步。这两样占着重量和成本，却不在自己的主场。真正属于眼前这两块玻璃的，只有近眼显示。
砍完之后，筹码反而清晰了，即显示、麦克风、AI。
XREAL则做了另一种减法。它不急着让用户接受一个取代手机的未来，而是把XR收回到一个古老得多的需求上，即人总是需要一块屏幕，只是这块屏幕还不够自由。
家里放不下投影仪，戴上眼镜就是一块巨幕；出差、旅行，熟悉的画面随身带走。其逻辑，是把原本钉在客厅里的需求，从固定的空间里松绑。故事不性感，却更像消费电子该有的样子。
一位XR从业者的体会是，近眼显示恰好踩在这个品类最朴素的起点上：“很多人买Vision Pro之前，幻想自己在里面健身、做三维设计，等真买回来，绝大多数时候还是把它当外接屏办公，或者干脆看片。”
第二条，是往手机够不着的缝隙里钻。
通用场景看似宏大，实则全在手机厂商的射程之内，创业公司硬碰硬，半点胜算没有。能落脚的，是那些双手腾不出、感官有缺口或者手机很难适配的地方。
于是玩家们各找各的缝。Rokid钻进车里，和广汽联手把导航、车况投进驾驶视野，又接上腾讯WorkBuddy、微信支付，咬住通勤和办公；纳米盒扎进课堂，主打课堂笔记、抗遗忘背词；亮亮视野深耕会议同传和听障辅助，一套系统，顶下了一屋子同传设备。
Rokid创始人、CEO祝铭明曾说过：“很多人买眼镜已经不再是买眼镜，而是买能力，就像很多人买Switch就是为了玩塞尔达。”
这意味着，玩家们似乎暂时放下了“无所不能”的执念，转而追求“缺一不可”，即一副眼镜只要在实时翻译、提词、无障碍字幕这种场景里做到不可替代，就能俘获第一批大众用户。
还有一种退守更彻底，干脆不争用户，转头去往数采赛道，去给另一个风口当“铲子”。
眼下，具身智能颇为稀缺的，是能看懂人类怎么干活的第一视角数据。而眼镜恰好长在人眼位置，基于此，渐渐有玩家把给人戴的终端，改造成给机器人喂数据的“眼镜”。
比如PICO在GTC上，接入英伟达Isaac Teleop架构，PICO 4 Ultra能直接录制第一人称操作数据；小派则与高校共建具身数据平台，并拿到智元旗下觅蜂科技的战略投资。
这意味着，面对那道横在极客和大众之间的鸿沟，玩家们选择绕开它，即不与主流市场硬刚，直接跳到鸿沟对岸的一个个利基市场里，先活下来。
这在商业上是务实的，产品复购、黏性和口碑，都比泛泛的“极客玩具”扎实得多。但选择退守，亦将背负对应的代价。毕竟，一旦AI眼镜把自己定位成“垂直场景的工具”，就等于主动放弃了“下一代计算入口”的叙事。
一副更好用的助听器、提词器、点单器，能让行业活下去，却不能点燃行业。于是，所有不甘心只做工具的人，都得再找一面新的旗帜。
反攻：新的故事，旧的难题
退到墙角，行业把为数不多的筹码，押给了Personal Agent。
今年9月，Meta发布Muse后不久，其便宣布将Muse接入自家AI眼镜。10月，紧跟Meta，阿里千问亦在硬件发布会上，将包括AI眼镜在内的AI穿戴视为Personal Agent在物理世界的延伸。
这是少数能把眼镜从“功能集合”抬回“智能入口”的锚点，也是不甘心只做工具的AI眼镜玩家，重塑叙事时，难得的机会。
AI眼镜在Personal Agent时代的一大优势，是“在场”。
Agent要懂你、替你办事，前提是它得持续地感知你、理解你。手机在口袋里，要用户主动掏出记录，它所看到的，是用户愿意让它看到的那部分；能持续贴在第一视角上，同步地看、同步地听、同步记录，给Agent喂上这口最稀缺的饭的，只有眼镜。
除了Personal Agent，AI眼镜攒了三年也没攒够的应用生态，亦撞上了上一轮AI叙事热点——AI Coding。
在此之前，没有开发者愿意为了这么小的盘子做App。一位XR应用创业者告诉光子星球，相比硬件，这几年XR应用侧的融资一直处在低迷状态。
但随着AI Coding愈发成熟、门槛降低，个性化应用势必将同AI眼镜展开一轮“亲密接触”。
据光子星球了解，目前已有头部厂商，尝试在下一代产品中加入AI Coding能力。而Rokid，在今年9月上线的AIUI Studio，解决的亦是同一件事。
不过，叙事归叙事，就产品而言，AI眼镜究竟应该如何生长，目前仍悬而未决。
雷鸟创新CEO李宏伟曾给过一个数字，即所谓的AI功能，长期使用率不到6%。Meta本该是这套逻辑的标杆案例，结果却也混成了反面教材。
据Meta可穿戴副总裁Alex Himel此前透露，用户戴眼镜的时间里，只有约10%在用智能功能，剩下九成，是冲着外观和重量把它当普通眼镜戴。长期佩戴的用户也发现，纯AI问答是新鲜感退得最快的部分，头一个月还主动问，半年后宁可掏手机。
另一方面，愿景成立是一回事，眼镜受不受得起是另一回事。
Personal Agent试图实现的“全天智能”，和眼镜赖以成立的“全天佩戴”，在今天AI眼镜的硬件条件下，实现难度都比较高。
眼镜为了戴在脸上，重量、体积、散热、电池全被压到极限，能塞进的，本就是手机或手表芯片降规格后的版本。端侧那点算力，只够跑唤醒、识别这类轻活，跑不动一个要实时理解画面和声音、还要做推理的智能体。
端侧跑不动，就上云。可音视频先上传、结果再传回，延迟、网络环境、Token成本，每一项都得有人买单。
电量亦然。Personal Agent的前提是Always-On，眼镜恰恰是最不允许Always-On的终端。若AI、录音和拍摄火力全开，一两个小时电量就会见底。换电、随身带充电盒固然算得上解法，但仍无法支撑高频使用。
工程的世界里，没有死局，有的是绕开死局的路。一颗几毫瓦的协处理器，或许就能让眼镜始终睁着半只眼，让Agent维持“在场”。
只是，上一轮关于重量、光学、设计和应用的打磨还未收官，围绕Agent的新一轮探索，却已经开场。这意味着，厂商们注定要在一副“还不够好”的眼镜上，去追一个“尚未成熟”的Agent——通往终局之路，依然道阻且长。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Zen, Photons.
Since this year, AI glasses seem to have hit the first wall.
According to IDC data, in the second quarter of 2026, about 611,000 smart glasses came out of the country, representing a decline of 8.0 per cent over the same period, the first negative growth since the race track was raging. At the same time, the global market was flying at 35.3 per cent.
There was a chill in the country when there was heat and fire.
Before that, no class believed in “the next generation of mobile phones” more than the Ai glasses; nor was any class believed in this belief, which was so hard.
The manufacturer's All in, the capital's pegged, and publish's going to go on and on, as if the replacement of the phone by the Ai glasses was just a matter of time. But it's all on the supply side, but the demand side seems a little warm, and it seems that those who wear glasses every day are the ones who love technology and enjoy taste.
The brands are more than a year old, and the brands are new, but the user circles are not bursting — there are too many Ai glasses and the crowds seem to be running short.
As a result, a group retreat began with “everyone wears” as a player: Some players try to turn their products back into ordinary glasses with a reduction approach; others stick them in the narrow edges of vehicle carriages, translations, meetings, accessibility, etc.
Reductions and narrows allow the industry to survive without feeding their appetites at the entrance. Against this background, the players retreated while putting their chips on Personal Agent.
The first two scenes have come to an end, and the third appears to have just begun.
Breaking the wall: Too many Ai glasses, too fast to use.
Although single-season data are influenced by the new publish, listing rhythms, there may be short-term fluctuations; a one-quarter drop alone would also result in unfair sentences for industry.
But a race track that is pinned on high growth expectations, at a time when producers are dense with publish, marketing is increasing, and the Chinese market, which is the base plate, is not going to rise or fall, and ultimately has been bogus.
On the product side, expansion is indeed visible. In two years, the weight of the AI glasses went from 50 to 30 grams, the size of the chips went from one stack to two or three, and the function went from a dozen to over 200.
A major catalyst for this fanaticism is capital.
Investors have always been hungry for the word “entry” and have been willing to bet in advance for a future that has not yet materialized; yet paid consumers have been much more calm, measuring their products and still using normal electronic consumption standards.
In the early stages of the product, the availability of the money and the holding of the position is a living method, but the problem is that there is a high degree of maturity upstream of the Ai glasses and a high degree of similarity between the master, the light wave, and the sensor programmes, and that there is no real gap between the brands except for a very few players with real self-study capabilities.
The so-called Spectrum War reached consumers with only fittings, a combination of functions and mutual testing on price sheets.
On the demand side, those willing to wear AI glasses every day are the same people, i.e. technocrats, digital bloggers, and hardware-crazed friends.
Data from the Aids consultation show that the main consumers of AIS glasses are ATI fans and technologists, accounting for 32.5 per cent and 32.2 per cent, respectively, and a total of about 65 per cent.
Instead of being stingy, the group spends the most money, it tickles when it comes to new things, and some have their hands piled up and others buy new ones.
But this small group does not represent the public.
There is a principle in product thinking that is called "preventive," that is, that the user is a "wit" when designed, and that does not require the user to read the instructions or use them well. The more successful the mass product, the more bold the user is to think "mute" so that the operation is as intuitive as possible.
By this standard, many of the so-called Ai glasses are almost a negative educational material.
An practitioner told the photon planet that previously products, with weight and frame, were customarily done to the head of men, with narrower heads and more sensitive weights for girls, largely excluded, such as “heart cutting” and so forth. It was not until the manufacturers became aware of the problem that they added a mirror leg-adjusted function to the subsequent product.
The founder of XREAL, Xu Zhi, also stated in an interview that many domestic manufacturers were not “professional” and lacked sufficient patience and respect for a complex hardware class — seeing the front player produce the product, quickly follow up, dismantle, reset and “realize” the product with a supply chain combination, but in practice the bottom-up problem was not well understood.
At this stage, the overall experience of AI glasses is still not strong enough.
Although many of them enjoy the pleasure of participating in the improvement of a new technology, actively looking for BUGs, looking for scenes, giving feedback, the public does not have the patience and obligation to do so. What they want is an open-box, unserving product, and when the product requires its own adaptation, tolerance and justification, his choice is often simple — no.
The boom of capital-intensive supplies that hit unopened demand naturally cannot sustain the weight of the next generation of mobile phones.
The first wall was laid in front of all the manufacturers.
Stand down: Big entrance, small business first.
After the wall was hit, the industry did an honest thing, quietly changing the target group from “all” to “some kind of person”. The next generation of mobile phones, which had been closed to mouths, now retreated to work as a tool for vertical crowds and scenes.
It was driven out of reality.
The retreating path, in sum, is two, one down to the product and one vertically to the scene.
The first thunderbird, with two sharpest knives on his thunderbird io, cut off both the camera and the speaker, showing a return to single green and a total pressure of 34 grams.
For many users, cameras and loudspeakers are the object of AI glasses and are incomplete without them. As a result, many users of social media were discouraged from “embracing” as a result of the thunderbird iopublish.
Despite that, the thunderbird is still selling.
The logic is simple: to admit it must be a pair of glasses first, and then to say something else. And the loudspeaker, the TWS headphone, is better than it is; on the camera, the multiplicity is not as good as it really uses the first perspective. Both account for emphasis and cost, but are not in their own home. It really belongs to the glass in front of us, but it's only visible.
After the cut, the chips were clear: display, microphone, AI.
XREAL made another reduction. It does not rush to get users to accept a future that replaces mobile phones, but to take XR back to a much older demand that people always need a screen, which is not free enough.
There is no projector at home, and glasses are a big screen; travel, travel and familiar images are taken with them. The logic is to untie the demand that was originally pinned to the living room from a fixed space. It's not sexy, but it's more like consuming electronics.
One of the XR practitioners has the impression that close-eyes happen to be on the simplest starting point for this type of product: “Many people, before buying Vision Pro, fantasize about working out in it, doing three-dimensional designs, and when they actually buy it back, mostly use it as an off-screen office or simply watch films.”
The second is to drill into a gap where the cell phone cannot reach.
The general scene appears to be large, but it is all within the range of a mobile phone manufacturer, where the start-up company is hard-core and has no chance of winning. What can be done is where there are no hands, sensory gaps or mobile phones that are difficult to fit.
So the players look for different stitches. Rokid drilled into the car and joined forces with the public to put the navigation and condition into the driver ' s view, then joined with TencentWorkBuddy, WeChat to pay for it, biting on commuting and working; the nanobox was inserted into the classroom, taking notes in the classroom and writing against oblivion; the luminous-looking deep-earthed conference was accompanied by an acoustic and hearing-impaired aid, and a system was set up to top off a room with a delivery device.
The founder of Rokid, CEO, once said: “Many people buy glasses no longer for glasses, but for abilities, as many people buy Switch to play Zelda.”
This means that players seem to have put aside their “infinity” obsession for the time being and instead to pursue “infinity”, i.e. a pair of glasses that can capture the first public users if they are irreplaceable in a real-time translation, quotes, barrier-free Z100Z.
There is also a more thorough retreat, which simply goes back to the user and goes back to the race track and becomes a shovel for the other wind.
At present, intelligence is scarce and is the first perspective data that can understand how human beings work. And the glasses happen to be in the eye position, and as a result, there is a gradual shift by the player to a “glasses” that feed data to robots.
On the GTC, for example, PICO has access to the Isaac Teleop structure in Weidar, where PICO 4 Ultra can directly record first-person operating data; and PICO has developed a personal data platform with higher education institutions and received strategic investments in bee-seeding technology under the flag of intellectual capital.
This means that, in the face of the divide between the extremists and the general public, the players choose to bypass it, i.e. not to go hard with the mainstream market and jump directly to a niche market across the divide to survive.
This is commercially pragmatic, with product buy-backs, sticky and lip service much stronger than a broad “bots toy”. But opting out will carry the corresponding costs. After all, once AI glasses have positioned themselves as a tool for “vertical scenes”, it is tantamount to abandoning the narrative of “the next generation of calculators”.
A better-used hearing aid, verbs, and singles can keep the industry alive without lighting it. So, all those who don't want to be a tool have to find a new flag.
Counterattack: new stories, old dilemmas
Back to the corner, the industry gave the few chips to Personal Agent.
In September this year, shortly after MetapublishMuse, it announced that Muse would be connected to its own Ai glasses. In October, close to Meta, Ali thousands asked at the hardware publish meeting that the wearing of AI, including Ai glasses, was considered an extension of Personal Agent in the physical world.
This is a small number of anchors that can lift glasses back from the "functional assembly" to the "intelligence portal" and a rare opportunity for an Ai glasses player who is reluctant to do nothing but make tools to recreate narratives.
One of the main advantages of the AI glasses in the period of Personal Age is to be present.
Agent needs to understand you and work for you, provided that it continues to feel you and understand you. The mobile phone is in its pocket and the user is asked to take out the record, and what it sees is the part that the user is willing to see; it is able to keep on the first view, to read, listen and record, and to feed Agent the most scarce meal, only the glasses.
In addition to the Personal Age, the AI lenses, which have not been available for three years, also hit the previous round of AI narrative hotspots — AI Coding.
Until then, no developer was willing to do App for a plate of this size. An XR application entrepreneur told the photoplanet that finance on the side of the XR application had been in a low state over the years compared to hardware.
As AI Coding matures and lowers the threshold, however, personalized applications will inevitably lead to a round of “intimate contact” with AI glasses.
According to Photons, there are now head manufacturers who are trying to add AI Coding capabilities to the next generation of products. And Rokid, on the line in September this year, AAUI Studio, solved the same thing.
However, as far as narratives are concerned, how the AI glasses should grow in terms of products remains unresolved.
Thunderbird Innovation CEO Lee Hwaa gave a figure of less than 6 per cent of the so-called AI function over the long term. Meta was supposed to be the standard case for this logic, but the result was the opposite.
According to Meta's Vice President, Alex Himel, only about 10 per cent of the user's time in glasses was using smart functionality, with 90 per cent remaining to wear them as normal glasses for appearance and weight. Long-held users also found that pure AI questions and answers were the fastest part of freshness, with the first month asking on their own initiative, preferring a cell phone after six months.
On the other hand, the establishment of a vision is one thing, and the unaffordability of glasses is another.
The “full-time intelligence” that Personal Agent is trying to achieve, and the “full-day wear” of the glasses that they are built upon, are more difficult to achieve today under the hardware conditions of AI glasses.
To wear glasses on the face, the weight, volume, heat, and battery are all pressed to the limit and can be plugged in, which is the version of the reduced size of a mobile phone or watch chip. The tip of the line is only enough to wake up, recognize such light work and not move a AI agent that needs to understand images and sound in real time and to reason.
If you can't run, you'll be in the clouds. The video is uploaded, the results are sent back, the delay, the network environment, Token costs, each of which has to be paid for.
The same goes for electricity. Personal Agent's premise is Always-On, and the glasses are precisely the terminal where Always-On is the least allowed. If AI, the recording and the shooting are fully operational, the electricity will be at the bottom of one or two hours. The switch, the charge box with the sideband, although decomposition, still does not support high frequency use.
There's no dead end in the engineering world. There's a way around it. A milliwatt co-processing device might keep the glasses open half-eyed and Agent “at the scene”.
It was just that the last round of grindings on weight, optical, design and applications had yet to be completed and a new round of exploration around Agent had begun. This means that manufacturers are destined to pursue an “unmaturized” Agent — the road to the end — on a pair of “not good enough” glasses, which is still blocked and long.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.