2026AI硬件半场复盘:活下来的产品都做对了这3件事2026 AI Hardware Midseason Review: The Surviving Products All Got These 3 Things Right


AI free token/AI免费积分

 

图片
摘要
Humane AI Pin 停产了。
距离它被惠普以不到2.5亿美元收购,刚好一年。曾经融资2.3亿美元的明星公司,最后打折出售。Rabbit R1 也没好到哪去,退款潮一波接一波,用户社区里骂声一片。
2024年AI硬件那波狂热,到2026年终于冷静下来了。
大模型这波浪潮里,所有人都想造硬件。手机厂、耳机厂、手表厂、甚至做充电宝的都想往产品里塞个NPU,然后讲一个"AI赋能"的故事。但一年半过去,真正能活下来的产品,两只手数得过来。
我做了7年硬件PM,今年拆了不少AI设备,也正在推进一款AI墨水屏的立项。聊聊我的真实观察——活下来的产品,到底做对了什么。
一、不卷大模型,卷场景
图注
图片
AI Pin 的死法很典型。
它把GPT塞进一个胸针里,然后告诉用户:"你可以跟它对话,它可以帮你拍照、搜索、翻译。"
听起来很酷。但用户拿到手才发现:对话延迟高得离谱,拍照画质还不如手机,翻译不如直接打开百度。每个功能都做了,每个功能都做不好。
问题出在哪?
"跟AI对话"不是场景,是功能。
没有人需要一个能对话的胸针。人们真正需要的是:开会时帮我录音转文字;出国旅游时帮我实时翻译;跑步时帮我记录心率变化提醒我别太猛。
这些才叫
场景
。
活下来的AI硬件,都是先有
场景
,再找AI。
  • 讯飞的录音笔卖得好,不是因为它有"AI转写"四个字,是因为它解决了一个具体问题:开完会,5分钟出纪要,不用重新听一遍录音。用户为"帮我省了半小时"买单,不是为"AI"买单。
  • 翻译耳机能火,是因为出国的人真的需要跟当地人沟通,而不是因为他们想体验"AI翻译技术"。
我拆机时发现一个规律:凡是在包装正面印
"AI-Powered"
的产品,大多卖不动。反而是那些不强调AI、只强调使用场景的产品,留存率还不错。
消费者很诚实。他们不在乎你的芯片算力多少TOPS,不在乎你背后跑的什么模型。他们只在乎一件事:这东西能不能帮我解决一个具体的问题。
能,就买。不能,再炫的技术也没用。
二、BOM成本砍到见骨
AI排版配图
图注
硬件和软件最大的区别:成本结构
硬件和软件最大的区别:软件多一个用户,边际成本约等于零;硬件多卖一台,每一颗螺丝都是钱。
很多AI硬件创业公司死在这上面。
立项时算BOM:芯片X元、屏幕Y元、电池Z元,加起来觉得毛利不错,信心满满。等产品真能量产了才发现——良率不够,废料吃掉10%毛利;渠道要抽成30%;售后预留5%;还有认证费、包装费、物流费……一通算下来,卖一台亏一台。
我拆阅星瞳X4的时候仔细看过它的物料策略。墨水屏品类里,屏体成本占BOM的40%-50%是行业常态。但X4在SoC选型上选了国产方案,把芯片成本压到了同档竞品的60%左右。省下来的钱投在墨水屏面板上,用了更好的屏体。
这就是老牌硬件公司的打法:不卷参数,卷成本结构。
AI硬件的BOM里,最容易失控的是AI芯片。很多团队一上来就选最贵的NPU方案,觉得算力越大越好。但用户根本感知不到15TOPS和30TOPS的区别。他们感知到的是:这个东西卖1999,有点贵了。
活下来的产品有一个共同特征:BOM极其克制。该花的地方花(直接影响体验的屏幕、续航),该省的地方省(芯片选型够用就行、外观件用成熟方案)。
硬件赚钱的本质从来没变过:成本控制 > 定价能力 > 技术领先。这个排序在AI时代也没变。
三、续航就是产品力
图片
图注
AI硬件最大的敌人不是竞品,是充电器。
我跟踪过几款AI可穿戴设备的使用数据(脱敏处理,不点名):续航不到1天的产品,7天留存率中位数不到15%。用户前三天觉得新鲜,第四天忘了充电,第五天就让它吃灰了。
续航差的AI硬件,本质上是在跟用户的习惯对抗。没人愿意每天给一个胸针充电,再给一个眼镜充电,再给一个耳机充电。充电这个动作本身,就是在提醒用户"这个东西很麻烦"。
这就是
墨水屏
品类天然的优势。
墨水屏的功耗极低,一颗电池撑一周甚至两周没问题。对于AI墨水屏这类产品,"长续航"本身就是最强的卖点——不是因为参数好看,是因为它不改变用户的使用习惯。
用户不需要记住"今晚得给XX充电"。拿起来就用,放下去就待机。这种零负担的体验,才是AI硬件该有的样子。
我判断2026下半年AI硬件的胜负手之一就是续航。谁的设备不需要用户操心充电,谁就能留住用户。这个逻辑朴素到不像AI时代该说的话,但硬件就是这样——尊重物理规律比尊重技术趋势更重要。
· · ·
我的判断
2026下半年,AI硬件进入"去伪存真"阶段。
泡沫挤完之后,活下来的不是技术最强的,而是最懂硬件生意的。具体来说:
  • 有机会的品类:垂直场景工具(会议记录、翻译、阅读辅助)+ 低功耗长续航 + 成本可控。这三个条件同时满足,就有机会。
  • 没机会的方向:通用型AI终端(什么都想做,什么都做不好)+ 高BOM高定价 + 续航拉胯。三个里踩中两个,基本就是在等死。
对还在做AI硬件的朋友说句实话:
别被大模型的参数焦虑传染到硬件端。硬件产品的成败,80%取决于供应链管理和成本控制,20%才是技术方案选型。这个比例,AI时代也没有变。
AI是能力,不是卖点。把AI藏进产品里,让用户感受到价值但感受不到AI的存在——这才是硬件PM该干的事。
聊两句:
你用过哪款AI硬件产品?体验如何?最打动你的是功能、续航还是价格?
评论区聊聊,我会逐条回复。
觉得有用的话,转发给身边做硬件的朋友。

 

 

 

图片

Summary

The Humane AI Pin has been discontinued. It has been exactly one year since HP acquired it for less than $250 million. The once-star startup that raised $230 million ended up being sold at a discount. The Rabbit R1 hasn't fared much better either—refund requests have come in wave after wave, and the user community is full of complaints. The AI hardware craze of 2024 has finally cooled down by 2026. In this wave of large language models, everyone wanted to build hardware. Phone makers, headphone makers, watch makers, and even power-bank companies all wanted to stuff an NPU into their products and tell an "AI-powered" story. But a year and a half later, you can count the products that actually survived on two hands. I have been a hardware PM for seven years. This year I have torn down quite a few AI devices, and I am also pushing forward on the development of an AI e-ink product. Let me share my real observations—what the products that survived actually did right.

1. Don't compete on large models; compete on scenarios

Figure caption

图片

The AI Pin's death is a typical case. It stuffed GPT into a brooch and told users: "You can talk to it, and it can take photos, search, and translate for you." It sounded cool. But when users got it in hand, they realized: the conversation latency was absurdly high, photo quality was worse than a phone, and translation was no better than just opening Baidu. It did everything, and did nothing well.

Where was the problem?

"Talking to AI" is not a scenario—it's a feature.

No one needs a brooch you can talk to. What people actually need is: help me transcribe audio into text during meetings; help me translate in real time when traveling abroad; help me track my heart rate while running and remind me not to push too hard.

Scenario

Scenario

iFlytek's voice recorder sells well not because it has the words "AI transcription," but because it solves a concrete problem: after a meeting, you get the minutes in five minutes without having to listen to the recording all over again. Users pay for "it saved me half an hour," not for "AI."

Translation earbuds took off because people traveling abroad really need to communicate with locals, not because they want to experience "AI translation technology."

"AI-Powered"

Consumers are honest. They don't care how many TOPS your chip has, and they don't care what model is running underneath. They care about only one thing: can this thing help me solve a concrete problem?

If it can, they buy it. If it can't, no matter how flashy the technology, it's useless.

2. Cut the BOM cost to the bone

AI排版配图

Figure caption

The biggest difference between hardware and software: cost structure

The biggest difference between hardware and software: for software, one more user costs almost nothing at the margin; for hardware, every additional unit sold means every screw costs money.

Many AI hardware startups died on this issue.

At the project kickoff, they calculated the BOM: chip costs X yuan, screen costs Y yuan, battery costs Z yuan. Added together, the gross margin looked good, and they felt confident. But when the product actually went into mass production, they realized—yield was too low, scrapped units ate 10% of gross margin; channels took a 30% cut; after-sales needed a 5% reserve; plus certification fees, packaging costs, logistics costs... After all that, they lost money on every unit sold.

When I tore down the Yuxingtong X4, I looked carefully at its component strategy. In the e-ink category, it's the industry norm for the screen to account for 40%-50% of the BOM. But the X4 chose a domestic solution for the SoC, pushing chip costs down to about 60% of comparable competing products. The money saved went into the e-ink panel, using a better screen.

This is how veteran hardware companies play: don't compete on specs; compete on cost structure.

In the BOM of AI hardware, the AI chip is the easiest place to lose control. Many teams immediately pick the most expensive NPU solution, thinking bigger compute is always better. But users can't even perceive the difference between 15 TOPS and 30 TOPS. What they perceive is: this thing costs 1,999, which feels a bit expensive.

Surviving products share one common trait: extremely restrained BOMs. They spend where it matters (screens and battery life that directly affect experience) and save where they can (chip selection just needs to be good enough, and exterior parts use mature solutions).

The essence of making money in hardware has never changed: cost control > pricing power > technical leadership. This ranking hasn't changed in the AI era either.

3. Battery life is product strength

图片

Figure caption

The biggest enemy of AI hardware isn't competitors—it's the charger.

I tracked usage data on several AI wearable devices (anonymized, no names named): for products with less than one day of battery life, the median 7-day retention rate was below 15%. Users found it novel for the first three days, forgot to charge it on the fourth, and let it collect dust by the fifth.

AI hardware with poor battery life is essentially fighting against users' habits. No one wants to charge a brooch every day, then charge glasses, then charge earbuds. The act of charging itself is a reminder to the user that "this thing is a hassle."

E-ink

E-ink has extremely low power consumption—one battery can easily last a week or even two. For products like AI e-ink devices, "long battery life" is itself the strongest selling point—not because the specs look good, but because it doesn't change users' habits.

Users don't need to remember "I have to charge XX tonight." Pick it up and use it; put it down and it sleeps. This kind of zero-burden experience is what AI hardware should feel like.

My call is that one of the key deciding factors for AI hardware in the second half of 2026 will be battery life. Whoever's device doesn't require users to worry about charging will keep its users. This logic is so simple it sounds out of place in the AI era, but that's hardware—respecting the laws of physics matters more than respecting technology trends.

· · ·

My assessment

In the second half of 2026, AI hardware enters the "separating the real from the fake" stage.

After the bubble is squeezed out, the survivors won't be the ones with the strongest technology, but the ones who understand the hardware business best. Specifically:

Categories with a chance: vertical scenario tools (meeting notes, translation, reading assistance) + low power consumption and long battery life + controllable costs. If all three conditions are met at the same time, there's an opportunity.

Directions with no chance: general-purpose AI terminals (wanting to do everything, doing nothing well) + high BOM with high pricing + poor battery life. Hit two out of three, and you're basically waiting to die.

To friends who are still building AI hardware, I have a piece of advice:

Don't let the parameter anxiety from large models infect the hardware side. The success or failure of a hardware product is 80% determined by supply chain management and cost control, and only 20% by technology solution selection. That ratio hasn't changed in the AI era either.

AI is a capability, not a selling point. Hide AI inside the product, let users feel the value without feeling the presence of AI—that's what a hardware PM should be doing.

Let's talk:

Which AI hardware product have you used? How was the experience? Did the feature, battery life, or price impress you the most?

Share in the comments, and I'll reply to each one.

If you found this useful, forward it to friends who work in hardware.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry