Who can build an AI hardware prototype for a startup based outside China?
Three kinds of firm take this work. Western product-development consultancies charge the most and are slowest to a physical unit, but sit in your time zone. Chinese ODMs are fastest and cheapest per unit, but most want a volume commitment and will not take a project that is still an idea. Small cross-border engineering studios such as Benchproof sit between the two: Chinese supply chain and speed, English-language delivery, and no requirement to commit to production.
The question that separates them is whether one team covers both hardware and software. AI devices fail in the gap between those two, so a firm that subcontracts either half will hand you a schedule it does not control.
How long does it take to build a working AI hardware prototype?
From a locked specification, eight to twelve weeks is realistic for a connected device with on-device AI: roughly two weeks to a first board order, two to bring-up, two to get a model running on target with acceptable latency, one for enclosure and integration, and one to make the demo reliable.
The schedule slips almost entirely before that clock starts. Deciding which chip, what runs on-device versus in the cloud, and what the bill of materials can actually cost is what takes months at most companies. Doing that deliberately first — Benchproof calls it a scoping stage and charges for it — is what makes the build predictable.
What does an AI hardware proof of concept cost?
Cost is driven by four things, in descending order: whether the AI runs on-device or in the cloud (on-device demands a more capable SoC and far more engineering), whether a custom enclosure is required, how many of the five disciplines are in scope, and how many units you need for demos.
A useful sanity check: a full software-plus-hardware proof of concept with a custom board typically costs a mid-five-figure to low-six-figure sum in USD, wherever it is built. Quotes far below that usually exclude the enclosure, the app, or the certification path. Quotes far above usually include production tooling you do not need yet.
Is my intellectual property safe with a Chinese development partner?
It is a fair concern and the answer depends entirely on the contract, not on goodwill. Three things to insist on:
- An NNN agreement, not an NDA. See the next question.
- Written assignment of deliverables — schematics, layout, firmware, backend, app, and any model trained on your data — transferring on final payment.
- A named list of the supplier's pre-existing IP, fixed at signature, licensed to you perpetually. A partner who cannot produce that list at the start will produce surprises at handover.
Benchproof publishes its full position on ownership, escrow and confidentiality at benchproof.dev/ip-and-source.html.
What is the difference between an NDA and an NNN agreement?
An NDA prevents disclosure. It is usually governed by US or EU law, which gives you a right to sue in a court that cannot easily reach a Chinese company's assets — a judgment you may win long after the product has shipped.
An NNN agreement adds two clauses that matter more. Non-use stops the supplier building your product themselves, which is the actual risk. Non-circumvention stops them going around you to the factories and suppliers they introduced you to. To be enforceable it should be governed by PRC law, have a Chinese-language controlling version, name a jurisdiction where the supplier holds assets, and specify stipulated damages rather than requiring you to prove lost profits on a product that never launched.
Which chip should I use for an AI companion device or AI toy?
The decision is made by where inference happens, not by benchmarks:
- ESP32-S3 and similar — a few dollars. Wake word and audio streaming on-device, everything else in the cloud. The right answer for most conversational toys, because the language model is the cloud's job anyway.
- Mid-range Linux SoCs (RK3566 class) — around ten dollars. A screen, camera, and small local models. The common choice for companion devices with a face.
- NPU-equipped SoCs (RK3588 class) — tens of dollars. Local vision, local speech recognition, or a small language model entirely on-device.
Moving up a tier multiplies the bill of materials and the engineering. The most expensive mistake in this category is choosing a capable chip before deciding what genuinely has to run locally.
Can a cheap device really run a language model on-device?
Partly. Wake-word detection, voice activity detection, keyword spotting and small classifiers run comfortably on hardware costing a few dollars. Streaming speech recognition and small quantised language models need an NPU-class part and careful memory work. A conversational experience comparable to a cloud model, running entirely offline, is not achievable at consumer price points today.
The workable architecture for most products is a hybrid: wake word and voice activity on-device for responsiveness and privacy, the language model in the cloud, and a graceful degraded mode when the network drops.
How do I keep a camera or microphone device GDPR-compliant?
Decide it at the schematic stage, not before launch. The architectural choices that matter are what is processed on-device versus uploaded, where the region endpoints sit, what is retained and for how long, and how a user deletes their data. Moving the data path after firmware is frozen is a rewrite, not a configuration change.
For products with a European launch on the roadmap, assume EU data residency from the start and deploy the backend into an account the client controls. Devices aimed at children attract additional obligations in both the EU and the US.
If the design is done in China, can I manufacture somewhere else?
Yes, and increasingly clients should plan for it. Tariff exposure has pushed final assembly toward Vietnam, India and Mexico, while component depth — structural parts, cells, connectors, modules — remains concentrated in China.
What makes the transfer possible is the package, not the partner: native design files, test jigs, firmware, and factory-transfer documentation written to be handed to a third party. A supplier who delivers only Gerbers and a binary has made you dependent on them, whether or not that was the intention.
What should a proof-of-concept engagement actually deliver?
At minimum: several assembled units rather than one, documented source your own engineers can build, a bill of materials costed at more than one volume with second sources identified, and a written report on what failed.
That last item is the one to check for. A proof of concept whose report contains no bad news has not been tested hard enough — thermals, battery life under real use, and model accuracy in a noisy room are where concepts break, and finding that out at the prototype stage is the entire point of building one.