AI Hardware 2027: Why the Second Wave Is Different

The first generation of standalone AI devices did not fail because the AI was bad. It failed because it tried to replace the phone. The question Blazon gets asked most about this category is whether the wave arriving in 2027 ends differently from the one that arrived in 2024, and the answer is yes, for a reason most people miss.

A standalone AI device has to clear four structural bars: a job worth doing, thermal and battery headroom, unit economics, and distribution. Three of those four moved measurably during 2026. The fourth, distribution, decides who wins, and it is the one almost nobody is building for.

This is the argument, not a roll-call. If you want the roll-call, we keep a separate list of AI hardware startups worth watching.

The first wave failed on commercialisation, not on intelligence

Humane is the documented case, and it is worth being precise about what happened rather than reaching for the easy narrative.

By its March 2023 Series C, TechCrunch reported that Humane had raised approximately $230 million from existing and previous investors, among them Kindred Ventures, SK Networks, LG Technology Ventures, Qualcomm Ventures, Tiger Global, Microsoft and Volvo Cars Tech Fund. The Ai Pin launched in April 2024 at $699, later cut to $499.

The Verge, citing internal sales data, reported that Humane sold roughly 10,000 Ai Pins against a goal of shipping 100,000 devices by April 2025, that between May and August 2024 more Pins were returned than were purchased, and that the device and its accessories generated just over $9 million in revenue. That reporting also noted the returned units could not be refurbished, so they became e-waste.

In October 2024 Humane recalled approximately 10,500 Ai Pin Charge Case Accessories after lithium cells from a third-party supplier were found to pose a fire hazard, refunding $149 for standalone purchases and $129 where the case came bundled.

On 18 February 2025, HP announced it would acquire assets from Humane for $116 million, covering the CosmOS operating system, more than 300 patents and patent applications, and most of the technical staff, folded into a division called HP IQ. The Ai Pin device business was explicitly not part of that deal. It was wound down, and the servers went dark at 3pm ET on 28 February 2025, at which point calling, messaging and AI queries stopped working on every Pin in the world.

Read that sequence carefully and the lesson is not that the technology was premature. The team built genuinely novel hardware and sold the intellectual property for $116 million, which is not the outcome of a company with nothing.

The lesson is that the device had no job it did better than the phone in the buyer's pocket, so nothing held the customer through the first bad week. Returns outran sales inside four months.

Rabbit is the retention lesson, and it is not over

Rabbit announced the r1 at CES on 9 January 2024 and sold well. In June 2024, Forbes reported CEO Jesse Lyu telling the Collision conference in Toronto that the company had sold 130,000 units.

Then the r1 did what a device with no defensible job does. Three months later, in September 2024, Lyu put the installed base at over 100,000 units after returns and said only around 5,000 were in daily use.

Reception was hostile on a specific point. Reviewers argued the functionality could run as a smartphone app, an argument that gained force when Android Authority ran the rabbit launcher as an Android APK on a standard Google Pixel 6a. Lyu rejected the characterisation, saying rabbitOS and the large action model run in the cloud on heavily modified AOSP, and that a local copy of the APK cannot reach the service.

Rabbit has since pivoted the software twice, shipping a "teach mode" and LAM Playground in late 2024 and a full rabbitOS 2 redesign in September 2025 that repositioned the device as an agent rather than an assistant.

In November 2025, TechSpot, summarising reporting by Tom's Guide, said three of Rabbit's 26 employees had been on strike since October, having not been paid since July, with Lyu attributing the strain to regulatory complications that collapsed an Indian launch and saying a funding round was expected to close.

We are not writing an obituary. Rabbit sold hardware, retained almost nobody, and has spent two years trying to build the job in software after the device shipped. That is the actual failure mode and it is a commercialisation failure: the sale was won and the customer was lost.

The four bars a standalone AI device has to clear

Three of those four are now addressable. The first is a product decision, the second is an engineering constraint that is easing, the third is a business model choice.

The fourth is the one that money and model quality cannot buy quickly. It is where the 2027 cohort will separate.

The silicon constraint genuinely lifted in 2026, and you can date it

The argument that on-device inference is "getting cheaper" is usually hand-waving. In 2026 it became datable.

At MWC 2026, Qualcomm launched Snapdragon Wear Elite, which Qualcomm says is its first 3nm wearable processor and the first wearable platform with a dedicated Hexagon NPU alongside a separate always-on eNPU for keyword detection, activity recognition and noise suppression. It runs models of up to 2 billion parameters on-device at roughly 10 tokens per second, targets the 300 to 600mAh batteries that wearables actually carry, claims 30% longer battery life than the previous generation, and charges to 50% in ten minutes.

Google and Samsung were named partners. First commercial devices were expected within months of the announcement.

The detail that matters for this argument is how Qualcomm positioned it. Qualcomm frames the platform for personal AI devices beyond smartwatches, naming pins and pendants that operate as standalone AI nodes rather than phone companions. That is the exact category that died in 2024, being handed purpose-built silicon in 2026.

At the other end of the scale, AMD used its IFA 2026 opening keynote in Berlin to launch the Ryzen AI Max PRO 400 series, supporting up to 192GB of unified memory shared across CPU, GPU and NPU. AMD claims it can run models of more than 300 billion parameters locally, without a discrete GPU and without sending data off the machine. That is AMD's own launch claim, not an independent benchmark, and systems ship from Q3 2026 through HP, Lenovo and ASUS.

At CES 2026 in January, Intel launched its Core Ultra Series 3 mobile processors, codenamed Panther Lake, which Intel says are the first client processors built on its 18A process, aimed at the same on-device shift. Both January and September pointed the same way.

We were on the IFA floor in Berlin on 4 and 5 September, and the change was visible below the keynote level too. RayNeo showed the GT Max, billed as the first AR glasses with Dolby Vision playback, and the RayNeo iO, a 33g pair of AI glasses with a heads-up display and deliberately no camera, with the GT series starting at $329 and rolling into 40 countries from 4 September.

Xreal confirmed its Android XR headset would ship before the end of 2026. Anker, a company with distribution most startups will never have, showed MindBase, a local AI home hub with a 64GB SSD and two drive bays, and AI note-taker earbuds with transcription. Our fuller write-up from the show is in our IFA 2026 report.

Silicon stopped being the excuse in 2026. That is the good news, and it is also why 2027 gets crowded.

Power and thermals still decide which form factors survive

The constraint that has not lifted is heat, and it determines which 2027 devices are viable.

The Humane Ai Pin ran hot. Reviewers documented it overheating within minutes of setup and cutting off contact to cool down during back-to-back requests.

Humane itself said the laser projector was designed for brief interactions of up to six to nine minutes rather than prolonged use, and that it had set conservative thermal limits for the first release. The heat was attributed in part to limited dissipation room around the processor.

That is a physics problem wearing a product review. A device worn against the body, with no airflow and a battery you can hide under a jacket, has a hard ceiling on sustained compute.

Every form factor that has worked commercially respects it. Capture devices run a microphone and a radio, and do the heavy work elsewhere. Glasses put the weight on a frame with surface area and keep the display small or leave it out entirely.

The camera-free RayNeo iO at 33g is that trade made explicit, and it happens to solve a social problem at the same time.

If a 2027 pitch deck describes continuous multimodal inference on the body with all-day battery, the thermal claim is the one to interrogate first. A launch programme cannot fix a device that overheats.

Distribution is the variable that actually decides outcomes

This is the part of the argument the category keeps skipping, and the evidence is unusually clean.

EssilorLuxottica reported selling over 7 million Meta AI glasses in 2025. A year earlier it had told investors the Ray-Ban Meta glasses had sold 2 million units in total, covering the period from the October 2023 launch to February 2025.

It described smart glasses as the dominant driver of Ray-Ban wholesale growth in the second half of the year, with North American growth described as exponential. In the second quarter of 2026, revenue from Meta AI glasses nearly doubled year on year.

Meta did not build a retail network to achieve that. It partnered with a company that has almost 18,000 retail stores worldwide, running LensCrafters, Sunglass Hut, Pearle Vision and Target Optical.

Then it did something even more instructive with the Meta Ray-Ban Display: launched 30 September 2025 at $799, sold in-store only at selected US retailers, with a mandatory demo appointment before purchase. In January 2026 Meta paused the planned rollout in the UK, France, Italy and Canada because of what Meta called unprecedented demand and limited inventory.

Compare that with the alternative model. Humane sold online, direct, with a carrier plan attached and no way to try the thing.

Friend spent more than $1 million on a New York subway campaign, which Outfront, the company that handles advertising in the subways, confirmed ran to more than 11,000 ads across stations and trains. Friend's own account of the buy, given to Adweek, breaks that down as 11,000 car cards, 1,000 platform posters and 130 urban panels. The ads were comprehensively defaced.

We make no claim about Friend's commercial outcome, because none is documented. The narrower point stands on its own.

Attention is not distribution. Eleven thousand subway ads put a product in front of people. A booked demo appointment puts it on their face.

IDC's tracker data says the same thing structurally for the one category with real volume. It forecasts approximately 13.6 million display-less smart glasses units in full-year 2026 rising to 27.3 million by 2030, with average selling price falling from $376 to about $229, and optical see-through display glasses growing from 3 million units in 2026 to 12.2 million by 2030.

Those are smart glasses figures specifically, not a measure of AI hardware as a whole. They describe a category that already has shelves.

The phone got the same capability for free

One more bar got higher, not lower.

At WWDC 2026 Apple opened its Foundation Models framework to third-party model adapters, so providers can implement the public LanguageModel protocol from iOS 27 onward. A Gemini implementation was available through the same protocol in preview from day one, which means developers reach on-device models and frontier cloud models through the same API, with on-device inference carrying no per-query cost and working offline.

Every standalone AI device in 2027 is competing against capability the buyer already owns and did not pay extra for. That does not make standalone devices pointless.

It does mean the job has to be one the phone genuinely cannot do: hands-free capture in a meeting, a display in the line of sight, a sensor the phone does not carry, a context the phone is not in. "It has an AI in it" is not a reason to buy a second object.

Why 2027 and not 2026

The timing argument has a concrete anchor. OpenAI's first Jony Ive-designed consumer device, following OpenAI's acquisition of io in May 2025, has been reported by The Information as a camera-equipped smart speaker priced between $200 and $300, designed to learn about its user and surroundings, and not shipping to customers before the end of February 2027. Reported friction includes the pace of design revisions.

Every specification there is reporting, not confirmation, and none of it has been announced.

That device matters to everyone else in the category whether or not it is any good. It sets the reference point every buyer, retailer and journalist will compare against, and its supply chain pulls components and attention with it. A 2027 launch window is a crowded window.

Meanwhile the platform companies moved in rather than out. Amazon acquired Bee, the AI wearables startup, in July 2025, a $49.99 bracelet with a $19 per month subscription.

The category was not abandoned after 2024. It was absorbed by organisations with distribution.

What the second generation has to get right

This is market analysis, not a launch plan. If you want the operational version, we have written separately on how to launch a hardware product.

Companion, not replacement. Every documented commercial success in this category sits beside the phone. Every documented failure tried to supersede it.

One job, done reliably. Plaud has shipped more than 2 million AI notetakers and passed $100 million in annualised software revenue on a single job: capture a conversation, return a usable record. Narrow beats general when the buyer has to justify a second device.

Hardware plus recurring revenue. Plaud sells its $179 Plaud Pro notetaker, and every new device comes with a starter plan carrying 300 free transcription minutes a month.

Roughly half of device users upgrade to a paid tier. Bloomberg reported in June 2026 that the company is planning a new wearable and targeting $500 million in 2026 sales.

A device that earns once and a device that earns monthly are not the same business, and they cannot afford the same customer acquisition cost. This is also why we build the Shopify store before the campaign rather than after it: the subscription attach happens in the store, not on the campaign page.

A channel where the product can be demonstrated. The $799 Ray-Ban Display sells by appointment. A device you cannot try is a device you can only return.

Honest thermal design, stated up front. Say what the device does for how long. Humane's own six to nine minute figure only became damaging because it surfaced after reviewers hit it.

A returns path that is not e-waste. Returned Ai Pins could not be refurbished. That is a cost line, a sustainability story and a reputational exposure in one.

The commercial argument, stated plainly

In 2027 the hard part stops being whether the device works and becomes whether anyone can buy it, and whether the launch converts. Every documented winner in this category won on commercialisation rather than on model quality.

Meta did not out-engineer anybody on distribution. It partnered with a company that owns the shelves and made the premium product a booked appointment. Plaud paired a device with a subscription and got roughly half its buyers onto a paid tier.

Humane raised approximately $230 million, built novel hardware, sold direct online, and moved on the order of 10,000 units.

That gap is a go-to-market gap, and closing it is what a launch agency is hired for. Blazon has run 500+ product launches and 300+ crowdfunding campaigns since 2016, driving $250M+ in sales and $120M+ raised, and our founders were on the IFA 2026 floor in September rather than reading the write-ups afterwards.

The work we do as a hardware launch agency is the commercialisation half. Positioning the device as a companion with one defensible job. Building the pre-launch audience before tooling is committed.

Running crowdfunding as a paid market test that proves demand and funds the first production run rather than as a last resort.

Building the demo-led retail and channel narrative, running the paid media, and carrying the store and post-launch growth after the campaign closes so the subscription attach rate actually materialises.

Our hardware launch work starts at a $15,000 minimum engagement, with full launch programmes from $40K.

We will also say the honest part. A launch programme cannot fix a device that overheats, and it cannot invent a job the product does not do. If those two things are right, the rest is commercialisation, and commercialisation is solvable.

Frequently asked questions

Why did the Humane Ai Pin fail?

It failed on commercialisation rather than on technology. The Verge, citing internal data, reported roughly 10,000 units sold against a 100,000 target, more returns than purchases between May and August 2024, and just over $9 million in device revenue. It ran hot enough to throttle, it was sold direct online with no way to try it, and it had no job the buyer's phone did not already do. HP acquired assets, IP and most of the technical staff for $116 million in February 2025, excluding the device business, and the servers were switched off on 28 February 2025.

Is Rabbit still operating?

Yes, as of the most recent reporting available to us. Rabbit shipped rabbitOS 2 in September 2025, repositioning the r1 as an agent. In November 2025, TechSpot, summarising Tom's Guide, reported that three of Rabbit's 26 employees had been on strike since October over wages unpaid since July, with the CEO saying a funding round was expected to close and teasing new hardware. Anyone citing Rabbit's status should re-verify it at the time of writing rather than assume an outcome.

What changed in silicon that makes 2027 different?

Wearable-class silicon got a purpose-built neural engine. Qualcomm says its Snapdragon Wear Elite, launched at MWC 2026, is its first 3nm wearable processor and the first with a dedicated NPU plus a separate always-on eNPU, running models of up to 2 billion parameters on-device at about 10 tokens per second within typical 300 to 600mAh wearable batteries. Qualcomm positions it for pins, pendants and other standalone AI nodes rather than phone companions. At the laptop end, AMD's Ryzen AI Max PRO 400 series at IFA 2026 supports up to 192GB of unified memory, and AMD claims it can run models above 300 billion parameters locally.

Are AI glasses actually selling, or is that hype?

They are selling. EssilorLuxottica reported over 7 million Meta AI glasses sold in 2025, having told investors a year earlier that the Ray-Ban Meta glasses had sold 2 million units in total between the October 2023 launch and February 2025, and said revenue from those glasses nearly doubled year on year in the second quarter of 2026. IDC forecasts approximately 13.6 million display-less smart glasses units in 2026, rising to 27.3 million by 2030. The reason is distribution: EssilorLuxottica has almost 18,000 retail stores worldwide, and the $799 Ray-Ban Display is sold in-store only with a mandatory demo appointment.

Should an AI hardware startup use crowdfunding?

It depends on what you want from it. Used well, a crowdfunding campaign is a paid market test that proves demand and funds a first production run before you commit tooling, and it gives you an audience that has already paid rather than a waiting list that has not. Used as a rescue for a product with no job or no channel, it exposes the same weakness faster and in public. Blazon has run 300+ crowdfunding campaigns, and the ones that work are planned as a launch mechanism rather than a funding mechanism of last resort.

Will AI devices replace smartphones?

No, and that assumption is precisely what sank the first generation. Every commercially documented success in this category sits beside the phone rather than replacing it, and the bar rose again at WWDC 2026 when Apple opened its Foundation Models framework to third-party adapters, with a Gemini implementation available through the same protocol in preview from day one. On-device AI is now a built-in feature of the device the buyer already owns. A separate object has to earn its place with a job the phone genuinely cannot do.

What is the biggest risk for a 2027 AI hardware launch?

Launch congestion combined with an undemonstrable product. The Information has reported that OpenAI's first Jony Ive-designed device, a camera-equipped smart speaker at a reported $200 to $300, will not reach customers before the end of February 2027, which concentrates buyer and press attention in the same window. If your device cannot be demonstrated somewhere a buyer can touch it, and cannot state honestly what it does and for how long, the window will close over you.