Why a Gallery Devoted to AI Hardware
Most coverage of artificial intelligence stops at software: model releases, benchmark charts, and prompt tricks that spread across social feeds in a weekend. The physical layer gets far less attention, even though it is where the rubber meets the road. Every token that a model generates has to be computed somewhere, and increasingly that somewhere is a small, silent box sitting on a desk or bolted under a workbench. This site exists to look closely at those boxes.
Our premise is simple. AI is migrating from vast remote data centers toward local, private, always-on devices. That migration has produced a wave of genuinely interesting industrial design — milled aluminum enclosures, custom cooling loops, tightly packed accelerator boards, and little status displays that glow in the dark. A gallery is the right format for this material because so much of it is visual: the shape of a heatsink, the arrangement of ports, the way a fan shroud tucks into a chassis.
We treat each device as an object worth seeing clearly. Photographs here are captured under controlled lighting, often with the enclosure removed, so the internals are legible rather than a blur of reflections. Accompanying notes record dimensions, thermal behavior, idle power draw, and the first-hand impressions of the person who opened the case.
What Counts as AI Hardware Here
The definition we use is deliberately narrow. To appear in this gallery, a machine must be primarily built for inference or agent workloads, not general-purpose desktop computing. That includes dedicated neural processing units, edge inference gateways, local LLM servers with multiple accelerator slots, and the occasional laptop-grade NPU that deserves a closer look.
- Dedicated accelerators: cards and modules whose entire die area is given over to matrix math and tensor operations.
- Inference boxes: self-contained appliances that run a quantized model out of the box, often with a web UI and an API.
- Edge gateways: fanless devices built to run small models in factories, shops, and homes where a full server would be absurd.
- Developer boards: the kits that arrive first, before a product exists, and tell us what a vendor is actually planning.
By keeping the scope tight, we avoid the trap of becoming yet another general hardware review site. The gallery is about a specific, fast-moving category, and every entry earns its place by being relevant to someone who runs models locally.
How the Collections Are Organized
The archive is split into three collections, each with its own emphasis and its own page in this gallery. Readers who want a broad survey can start on the home page and jump around; readers who know exactly what they are hunting for can go straight to the relevant collection.
Collection One — Prototypes
The earliest boards and reference designs, photographed before they were polished into products. This is where the ideas surface first.
Collection Two — Production Boxes
Shipping appliances and modules you can actually buy, documented with power measurements and thermal notes.
Collection Three — Concepts
Concept hardware, crowdfunded prototypes, and designs that push the agent form factor somewhere unexpected.
We refresh collections as new hardware arrives rather than on a fixed schedule. A device enters the archive when we can photograph it properly, not when a marketing calendar says so.
Our Photographic Standard
Gallery photography is easy to do badly and hard to do well. Glossy plastic, brushed metal, and raw PCB are three very different surfaces, and each needs its own light to avoid looking like a smudge. We photograph enclosures with soft, diffused light and internals with raking light that brings out trace routing and component labels.
Every entry includes at least one straight-on exterior shot, one three-quarter view, and — where the device opens without destroying it — an interior shot with the heatsink or shield removed. Scale is important too, so we include a reference object in the frame whenever the size is not obvious from the ports alone.
Images are processed minimally: white balance, exposure, and a gentle contrast curve. We do not composite multiple exposures or fake colors, because the point of the archive is to show what the hardware actually looks like on a desk, not what it could look like in a render.
The Reading You Should Do Before Buying
A gallery is a place to look, but it should also be a place to think. If you are evaluating hardware for a home lab or a small business, the visual archive is only the beginning. We encourage readers to combine what they see here with careful reading elsewhere, especially around specifications that photographs cannot reveal: sustained throughput, memory bandwidth, and software maturity.
For a long-running benchmark culture and deep silicon analysis, the broader hardware press remains invaluable. Long-established architecture and review desks have documented processor and accelerator generations for decades, and their archives pair well with the object-level photographs we publish here.
In the same spirit, if you want to see how one company has bundled agent hardware into a complete, shippable product line, the main ClawBox site is the canonical destination. It carries the current lineup, specification sheets, and purchasing details that a gallery, by its nature, does not attempt to duplicate.
Featured from the Archive
What the Gallery Does Not Do
It is worth stating the boundaries plainly, because clarity keeps an archive honest. This gallery does not sell hardware, does not take affiliate commissions on the devices it photographs, and does not publish benchmark numbers that we have not produced ourselves. Where a figure appears in our notes, it is a measurement we took on the specific unit in the photograph, not a number copied from a vendor slide.
We also avoid the trap of ranking everything against everything. A fifty-dollar inference stick and a five-thousand-dollar multi-GPU tower are not competitors, and pretending otherwise makes the archive less useful. Context matters more than leaderboards, and the notes next to each image try to preserve it.
Finally, this is a visual archive first and a review site second. We would rather publish twenty clear photographs of one machine than two paragraphs of opinion about twenty machines.
A Short History of Local AI Hardware
It is easy to forget how recently all of this existed only in server rooms. For more than a decade, the practical way to run a neural network was to send your data across the internet to a machine you would never see, operated by a company you had to trust. The physical hardware that made it possible lived behind locked doors, in rows of racks, cooled by infrastructure that cost more than a house.
The first meaningful crack in that model came from the edge. Factories, shops, and vehicles needed inference without the latency and cost of a round trip to a data center, so chipmakers began producing low-power accelerators that could run a small model on a few watts. These early edge parts were not glamorous — their benchmark numbers looked tiny next to a data-center GPU — but they proved something important: useful AI could run in a box small enough to hold in one hand.
Then came the consumer wave. Home labs, hobbyists, and small businesses discovered that a dedicated inference device could keep their data local, answer instantly, and keep working when the internet went down. What had been an industrial niche became a personal purchase, and the design language of the hardware changed to match. The anonymous rack unit gave way to objects meant to sit on a desk, with lights, speakers, and enclosures someone actually thought about.
That is the arc this archive documents. Every photograph in these collections is a frame from a larger story about compute moving from the cloud back into the room, and the machines that carried it there.
How to Read a Teardown Photograph
If you have never looked closely at a teardown image, the value is not always obvious. A bare circuit board under harsh light can look like a city seen from too high up — you know there is structure, but you cannot yet name what you are seeing. The trick is to read it in layers, the way a mechanic reads an engine bay.
Start with the power path. Near the input you will usually find a cluster of inductors and capacitors; their size and number tell you how much current the design expects to move and how seriously it treats stability. A board that pours silicon into voltage regulation is built to be hammered, not sipped.
Next, look for the memory. High-bandwidth parts sit close to the main die, often on the same substrate or in a tight ring around it. The distance between compute and memory is one of the most honest signals about a design's intent, because engineers do not spend board space on short traces unless the workload demands it.
Finally, notice what is absent. Unpopulated pads, missing connectors, and empty solder points are the ghosts of features that were planned and then cut. A production board is a record of compromises, and the gaps are where the compromises show. Reading them is how you learn what a company almost shipped — which is often more revealing than what it did.
This is a skill that accumulates, and the archive is arranged to help you practice it. The prototypes in Collection One are the most legible, because their debug headers and test points are still exposed; the production boxes in Collection Two hide their internals behind panels and require a more careful eye.
Frequently Asked Questions
Is this archive affiliated with any hardware vendor?
No. We photograph hardware from many makers and maintain no commercial relationship with any of them. When we point readers toward a vendor's own catalog, we do so because it is the authoritative source for specifications and purchasing, not because we are paid to.
Can I use these photographs in my own work?
Non-commercial reuse with attribution is generally welcome. Commercial or print use requires an individual agreement, because the hardware in the frame belongs to its makers even when the photograph is ours. Reach out through the contact page and we will respond to each request individually.
How do you decide what to photograph next?
We prioritize devices that run AI models locally and are either new, historically interesting, or unusually well designed. Loan offers from readers are the single biggest driver of what appears next, which is why the contact page invites submissions so directly.
Why are some numbers measured on a single unit?
Because that is the honest unit of measurement for an object-based archive. A thermal figure is a fact about the specific device in the photograph, not a statistical claim about an entire production run. We state this explicitly rather than pretending otherwise.
Start Exploring
The three collections are linked above and reachable from the navigation. If you are new to local AI hardware, the prototypes in Collection One are the gentlest introduction — they show the ideas before the compromises. If you want to see what you can actually buy and run this month, head to the production boxes in Collection Two. The companion archive at ClawBox One follows the single-device story in more depth, and the broader catalog lives at https://clawbox.com where the full product family is presented together.