An AI appliance is a dedicated box you buy once, plug in, and run AI models on, sold against the subscription you are already paying. The category is young, the advertising is running well ahead of the products, and almost every ad in it makes the same four claims.
Those four claims are checkable. Not by testing the hardware, which you do not have, but by reading what the vendor has already published about its own product. In every case below, the ad and the documentation were written by the same company and do not say the same thing.
This is not the same category as the screen-driving dongles we covered in physical AI agent devices, which watch your monitor and operate your apps. An AI appliance is about where the model runs. If you want the underlying technology, local AI models covers what actually runs on your own hardware, and hardware for AI agents covers what the hardware needs to do it.
The short version: check the four claims in the order below, and note that the one risk most people worry about, being stranded when the vendor folds, has not yet happened to anyone who bought a genuine local box. It has happened repeatedly to people who bought devices that needed a server to work at all. That distinction is the whole thing.
What an AI appliance actually is
A small computer with an unusual amount of memory, running an open-weight model locally, with a friendly front end on top.
The memory is the point. Model weights have to fit in memory to run at reasonable speed, so these machines carry 64GB to 128GB of unified memory where an ordinary desktop carries 16GB or 32GB. That is also why they cost what they cost, and why they got substantially more expensive this year, which we went through in local AI hardware prices.
The category currently divides into three groups. Chip vendors and PC makers sell developer boxes: NVIDIA’s DGX Spark, AMD’s Ryzen AI Halo, the Framework Desktop, various Strix Halo mini PCs. Apple sells Mac Studios and Mac minis that happen to be the best local inference machines most people will own, without marketing them as appliances at all. And a set of startups sells purpose-built boxes, usually on preorder, usually before anything ships.
It is that third group where the claims need checking, because they are the ones advertising to people who will not read a spec sheet.
Claim one: it runs offline
The test is not whether it can run offline. It is whether it does, on day one and on day five hundred, with the network cable out.
NVIDIA is the honest benchmark here, and worth holding others against. Its DGX Spark first boot documentation lists as a prerequisite "a fast, reliable internet connection (Wi-Fi or Ethernet) to download required updates during the initial setup process," and states that "the system will automatically download and install the full software image." It also warns against captive portals and phone hotspots. The documented setup flow creates a local username and password and no NVIDIA account appears in it, so after setup you are genuinely on your own hardware. But you cannot take the box out of the carton in a building with no internet and have it work. NVIDIA put that in the prerequisites rather than burying it, which is the standard to expect.
Compare that with the softer version. HP’s announcement for HP IQ describes "a local, 20B-parameter model" and says the product is "designed to run many experiences on-device while routing select tasks to the cloud only when permitted by enterprise policy and user permissions." That is a disclosure, and a fairly candid one for a launch release. It is also local-first rather than local-only, and those are different products. HP publishes no model name and no way to verify what stays on the machine, so treat it as a vendor claim.
Then there is the version that fails outright. There is a preorder campaign advertising on social platforms this month whose creative promises an AI that lives on your own device, over an image reading "works offline," for a box built on a low-end Intel N95 with 8GB of DDR4. The company’s own hardware page says the device "handles credentials, files, and tasks locally" and that "inference runs in the cloud, which means faster models, zero maintenance, and a lower price." Both statements are the vendor’s. Only one of them is in the advertisement.
That is the check. Find the vendor’s hardware or technical page, not the landing page, and look for the word inference. If a box has no meaningful AI accelerator and under 16GB of memory, it is not running a useful model locally, whatever the ad says. Ask directly what happens when the internet drops, and ask whether the product falls back to a cloud service when local inference fails, because a silent fallback is the most common way "local" stops being true without anybody lying.
Claim two: no subscription
Usually true of the hardware, and usually not the whole bill.
Separate three things: a subscription for the hardware or its firmware, a subscription for the app or front end, and metered charges for model usage. Vendors advertise the absence of the first and stay quiet about the other two. The preorder campaign above says "no hardware subscription" on the same page where it offers "$29/mo for app access" if you supply your own API keys, and pay-as-you-go model usage if you do not.
The clearest documented case of a one-time purchase acquiring a recurring fee is Friend, the AI pendant. It was $99 at preorder in 2024 and $129 when it actually shipped in 2025, with no subscription. It relaunched on 30 July 2026 at $249, and the new version charges roughly $10 a month to retain memories beyond thirty days. Nothing improper happened. The point is that "no subscription" describes the product you are buying today, not the product line, and the next generation is where the fee appears.
Plaud is worth looking at because it is the commercial success story in adjacent hardware and it is frankly subscription-shaped: a free tier capped at 300 transcription minutes a month, Pro at $17.99 a month, Unlimited at $29.99, Team at $28 per user. There is nothing hidden about it. It simply is not the model the appliance ads are selling against.
Ask what the device costs to operate for a year, not what it costs to buy.
Claim three: private by design
Running a model on your own hardware protects the inference. It does not by itself protect the front-end app, the tool integrations, the telemetry, the logs, or the fallback path.
Plaud is the fair test case here, because it is more forthcoming than most and still illustrates the problem. Its AI data usage transparency policy states plainly that "by default, your data is stored locally on your device and is not uploaded to any server," and then concedes that "when you use an AI-powered feature, only the data necessary to perform the requested task is transmitted to our AI service providers." On that page, those providers are not named. A separate support article does name them, as OpenAI, Google and Microsoft, with a zero-retention commitment and TLS in transit.
So the disclosure exists. It is just in the article a buyer is unlikely to find rather than in the policy written for buyers. That is a placement criticism and a fair one, and it is narrower than saying the company hides its subprocessors, which would be wrong.
The check is simple and it works on any vendor. Ask for the subprocessor list by name, in writing. A vendor that can produce one has thought about this. A vendor that answers with a sentence about being private by design has not. Then ask what is logged, where, and for how long, because a device that never sends your audio anywhere can still send a full record of what you asked it.
Claim four: yours to keep
This is the one with a real body of evidence, and reading it correctly matters more than any other part of this article.
Humane sold the AI Pin from April 2024 at $699 for one finish and $799 for the others, with a $24 monthly plan. The company had raised more than $230 million. HP acquired most of its assets for $116 million, announced 18 February 2025, and Humane shut the servers off ten days later, at noon Pacific on 28 February. Devices lost calling, messaging, AI queries and cloud access. Refunds went only to devices shipped on 15 November 2024 or later, claimed by 27 February, the day before shutdown. Subscriptions paid in advance were refunded pro rata.
Limitless sold a recording pendant. Meta acquired the company and the announcement went up on 5 December 2025. Existing customers were treated reasonably by most measures: sales stopped, support was promised throughout 2026, and the paid Unlimited plan was made free. But the same announcement says the service would no longer be available in "Brazil, China, the European Union, Israel, South Korea, Turkey, and the United Kingdom," and that "users in these regions will have until December 19, 2025 to download their information, after which, their account and all associated information will be permanently deleted."
Fourteen days. For a device sold on the promise of remembering everything, in the regions with the strongest data protection regimes, because of a corporate transaction on another continent. The announcement does not mention refunds, and the company’s standard 7-day return policy was not extended.
Both of those are worth carrying into a purchase conversation. Neither is an argument against AI appliances, which is the part most write-ups get wrong.
The graveyard that is not there
We went looking for a dedicated local model box that took consumer preorders and failed to deliver, or shipped and then stranded its buyers. We did not find one.
Every casualty in this space so far has been a device that needed a server to function. Humane’s pin was a thin client with a camera. Limitless was a microphone in front of a cloud transcription service. Both were switched off from somewhere else, because both could be. A box that holds the weights and runs the model does not have that failure mode, and the absence of a graveyard in that group is not luck. It is the architecture.
The nearest real analog predates the current category. Mycroft was an open-source voice assistant sold on local processing, and its Mark II Kickstarter ended on 12 February 2023 with rewards undelivered: more than 2,000 backers, about 52 units shipped, and a chief executive saying the company could no longer afford to build and ship them. That is a genuine case of money taken for a local-first device that never arrived. It is also three years old and a different product class, and it is a manufacturing failure rather than a switch-off.
Two corrections to stories in circulation, because getting them wrong undermines the useful parts. Rabbit has not shut down. The last substantive reporting, from November 2025, described employees and contractors going unpaid since July 2025 and three on strike, with the chief executive saying a binding term sheet had been signed with an investor he would not name. What happened after that is not publicly documented. As of today the company is still selling the r1 at $199 with a 30 day return window and the service is up. And Bee, the wearable Amazon acquired in July 2025, is the case where the feared outcome did not occur: its commitment that audio is processed in real time and never stored was publicly reaffirmed about six months after the acquisition, with the addition that Amazon has no access to transcripts. Worth noting that the strong version of that promise lives in statements to journalists, while the binding privacy notice is quieter.
For fairness, the good exits deserve naming too. Open Interpreter canceled its 01 Light device in September 2024 and refunded every order, writing that the hardware "didn’t offer enough value to justify manufacturing it." That is what a clean unwind looks like, and it is possible.
The one question that predicts the rest
Does the device do its main job with the company gone?
Not "is it private," which every vendor answers yes to. Not "is it local," which has four different meanings in this category. Just that. If the weights are on the box and the model runs on the box, a bankruptcy or an acquisition costs you updates and support, which is bad but survivable. If the box is a thin client, the same event costs you the product.
Everything else follows from the answer. A device that passes it can be evaluated on price, speed and memory like any other computer, and the break-even arithmetic in local AI hardware prices applies. A device that fails it is a subscription with a one-time fee attached, and should be priced as one.
Three things to do before a deposit leaves your account. Read the vendor’s hardware or technical page rather than the landing page, and find the sentence about where inference runs. Ask, in writing, what happens with the network unplugged and who the subprocessors are by name. And check what the deposit actually commits you to, because a refundable reservation and a non-refundable preorder against a ship date eighteen months out are not the same instrument, and the second one is where the money is.
Frequently Asked Questions
What is an AI appliance?
A dedicated computer sold to run AI models locally, typically with 64GB to 128GB of unified memory, a preinstalled model runtime, and a front end. The selling proposition is one purchase instead of a subscription, with data staying on the device.
Is an AI appliance the same as an AI PC?
No. An AI PC is a general-purpose laptop or desktop with a neural processing unit, sold to do everything a computer does. An AI appliance is a single-purpose box whose reason for existing is running models. In practice a Mac Studio or a well-specified mini PC does the job without either label.
Do these devices really run without internet?
Some do, after setup. NVIDIA’s DGX Spark documentation requires an internet connection at first boot to download its software image, and says so in the prerequisites. Others advertise offline operation while their own hardware pages say inference runs in the cloud. Read the technical page, not the ad.
What does “no subscription” usually leave out?
The app or front-end fee and the metered model usage. A vendor can accurately say there is no hardware subscription while charging monthly for access to the software and passing through per-token model costs.
How do I check a vendor’s privacy claim?
Ask for the list of subprocessors by name, in writing, and ask what is logged, where, and for how long. Running the model locally protects the inference. It does not automatically cover the app, the integrations, the telemetry, or a silent fallback to a cloud service when local inference fails.
Has any local AI box company gone under and left buyers stranded?
We could not find one. The devices that stranded buyers, such as the Humane AI Pin and the Limitless Pendant, all depended on a cloud service to function. A box that holds the model weights and runs them locally does not have a remote off switch.
What happened with the Humane AI Pin?
HP acquired most of Humane’s assets in a deal announced 18 February 2025, and the servers were switched off at noon Pacific on 28 February 2025. Devices lost calling, messaging, AI queries and cloud access. Refunds covered only devices shipped on 15 November 2024 or later and had to be claimed by 27 February. Prepaid subscriptions were refunded on a prorated basis.
What happened with Limitless?
Meta acquired the company, announced 5 December 2025. Sales stopped, support was promised through 2026, and the paid plan was made free. But service ended in Brazil, China, the European Union, Israel, South Korea, Turkey and the United Kingdom, and users in those regions had until 19 December 2025 to export their data before accounts and information were permanently deleted.
Is Rabbit still operating?
Yes. Reporting from November 2025 described unpaid staff and a strike, and what happened afterward is not publicly documented, but the company is still selling the r1 at $199 today and the service is running. Treat claims that it has shut down as unsourced.
How much memory does an AI appliance need?
Enough to hold the model plus its context cache. 64GB runs models in the 20B to 30B range comfortably. 128GB opens up larger sparse mixture-of-experts models. Dense models in the 70B class run but too slowly to be useful interactively on this class of hardware.
Is buying one cheaper than paying for an API?
Usually not, below roughly 50 million input tokens a month. Hardware prices rose substantially through 2026 while hosted prices held or fell, which pushed the break-even further out. The non-cost reasons, such as data residency, restricted networks and offline operation, have not changed.
What should I ask before paying a deposit?
What the deposit commits you to and whether it is refundable, what the full price and ship date are, whether specs can change before shipment, where inference runs, who the subprocessors are, and what the device does if the company is acquired or closes. A reservation you can cancel and a preorder against a distant ship date are different risks.
