
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。

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。
-
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"。


-
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。
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
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
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
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.