Gallery Two: Production AI Hardware

What Makes a Production Box Different

By the time a prototype becomes a production appliance, most of the interesting detail has been hidden away. The debug headers are gone, the jumpers are replaced with firmware settings, and the whole thing is wrapped in a smooth enclosure designed to sit quietly on a shelf. That is not a criticism — it is the point of shipping a product — but it changes what a photographer can show.

For production hardware, the story moves from the copper to the whole object: the enclosure, the ports, the airflow, the noise, the heat, and the small choices that define the experience of living with the machine. A production box is judged by how it behaves over weeks, not by how it looks under a macro lens, and this collection is built around that longer view.

Where a device opens without voiding a warranty or breaking an anti-tamper seal, we document the internals too. But we respect the boundary: a production appliance belongs to its owner, and we never tear down something that was not meant to be opened.

The Measurements We Publish

Alongside every production entry, we record a small set of numbers that photographs cannot capture on their own. These are measured on the specific unit in the photograph, under the same ambient conditions each time, and we note any variance we observe.

• Idle power: what the device draws sitting at a desktop with no model loaded.
• Load power: sustained draw during a representative inference workload, recorded over at least ten minutes.
• Acoustic level: fan noise at idle and under load, measured at a fixed distance.
• Surface temperature: the warmest external point after a sustained run, so you know whether it will scorch a shelf.
• Boot-to-inference time: how long from power-on to a running model, because an appliance is judged by its readiness.

These figures are the practical complement to the visual record. A box can look beautiful and still be loud enough to be unbearable in a quiet room, and the notes exist so you find that out before you buy, not after.

The Categories in This Collection

Production AI hardware is not one category but several, and they are aimed at very different people. The gallery groups them so a reader can find their own use case quickly rather than wading through devices that will never apply to them.

Edge gateways

Fanless, low-power devices for factories and shops where a server is out of the question.

Local LLM servers

Tower and rack units with multiple accelerator slots, built to run large models on-premises.

Accelerator modules

Cards and modules that drop into an existing machine and give it dedicated inference silicon.

Each category gets its own treatment. A fanless gateway is photographed for its industrial restraint; a multi-slot server is photographed for its engineering ambition. The visual language follows the device, not the other way around.

The Collection

Box 01. A minimal inference appliance; one status LED is the only sign it is awake.

Box 02. A fanless edge gateway with four antenna ports, built for the factory floor.

Box 03. A local LLM server tower; the vertical strip glows through the grille under load.

Box 04. A single-slot accelerator with a blower cooler that keeps the profile slim.

Box 05. A rack-mount inference unit with redundant power, aimed at small data closets.

Box 06. A compact desktop box that runs entirely passive — silent, but warm to the touch.

Power and Thermal Notes

Because production hardware is judged over time, we keep the measurements in this collection especially disciplined. Idle and load figures are recorded on the same unit, in the same room, with the same workload where possible, so the numbers can be compared across entries without caveats piling up.

These numbers are indicative of the photographed units, not a promise about every unit of the same model. Thermal behavior varies with ambient temperature and firmware, and we note both where it matters.

Choosing the Right Production Box

With so many shipping appliances on the market, the hardest part is not finding one — it is choosing. The photographs in this collection will not make the decision for you, but they will help you ask the right questions, and the right questions are most of the battle.

The first question is about duty cycle. A device that runs a model for a few minutes a day has very different needs from one that is expected to serve inference around the clock. Duty cycle drives the thermal budget, which drives the noise, which drives where you can actually put the machine. A fanless gateway in a silent study is a pleasure; a rack unit with a loud fan in the same room is a mistake.

The second question is about model size. Match the device to the model you actually run, not the model you dream of running. Buying a machine sized for a large model you never load is the most common and most expensive error we see. The idle and load figures in our notes are meant to keep expectations honest.

The third question is about software. Check the driver and firmware situation before you check the clock speed, because a device with great silicon and poor software support will frustrate you daily. This is the one area where photographs cannot help, and we say so plainly rather than pretending a teardown tells the whole story.

Living With an Inference Appliance

A production box is not a review unit you run for a week and return; it is something you live with. Over time, small details that looked trivial in a photograph become the things that define the experience — the brightness of a status LED in a dark room, the sound of a fan spinning up at three in the morning, the way a cable nest grows behind the unit.

Our notes try to capture some of this longer view. Acoustic measurements matter because a quiet device can become an annoyance when it is the only thing running in a silent house. Surface temperatures matter because a warm enclosure can damage the shelf under it or the sticker on top. These are not glamorous observations, but they are the ones that determine whether you are happy with a purchase a year later.

We also record the less obvious ergonomics: where the ports are, whether the power brick is a brick or a wall wart, how the mounting holes are arranged. A device that mounts cleanly under a desk and keeps its cables tidy will earn its keep in ways that never appear on a spec sheet.

A Glossary of Production Terms

Production hardware brings its own vocabulary, and a short glossary keeps the captions and tables in this collection readable.

• Idle power: the wattage a device draws sitting powered on with no model loaded, a measure of its base cost to keep around.
• Load power: sustained draw during a representative inference workload, recorded over at least ten minutes.
• Acoustic level: the fan noise measured at a fixed distance, at idle and under load.
• Peak temperature: the warmest external surface after a sustained run, in degrees Celsius.
• Fanless: a design with no moving cooling parts, relying entirely on passive heat dissipation.
• Anti-tamper seal: a sticker or mechanism that reveals whether an enclosure has been opened, often voiding warranty.

These terms give the numbers in our tables their meaning, and they help a reader translate a technical note into a practical expectation about what living with the machine will be like. When you see a peak temperature or an acoustic level in a caption, you are looking at a measurement that was taken on the photographed unit under stated conditions — not a number lifted from a vendor's marketing page.

From Here to the Wider Catalog

This collection covers shipping hardware we have photographed, but it is not a catalog and does not attempt to be exhaustive. If you are ready to buy or want the complete, current product family with specifications and support, the official ClawBox gallery catalog is the authoritative source. Our archive shows the machines up close; the catalog shows them for sale, and the two are meant to be read together.

For the experimental and concept designs that have not shipped — the machines that hint at where the form factor is going next — continue to Gallery Three.