Producing 100 AI cameras is indeed an awkward volume. If you go to a major ODM, once they factor in mold costs, engineering fees, and production line changeover expenses, the per-unit price becomes ridiculously high—or they might simply refuse the order. Even if they accept it, the level of cooperation is limited, since it's a "small order" and naturally gets lower priority. On the other hand, if you go to a pure solution provider, you worry they'll only hand over the PCBA and firmware, leaving you to handle the enclosure, heat dissipation, lens, cables, and so on yourself. You might end up dealing with several suppliers and ultimately becoming your own assembly integrator, draining all your time and energy.
In fact, at this volume, the best fit is often neither the "big and comprehensive" ODM nor the "design-only" solution provider, but rather a solution provider that already has mature product lines and is willing to do low-volume custom deliveries. Such companies typically have off-the-shelf AI camera reference hardware—for example, common Ambarella, Rockchip, or HiSilicon platforms—with fixed combinations of main control, memory, and sensors. The software side is already tested for face detection, human shape recognition, or simple visual inspection. All you need to do is tell them your application scenario—whether it's attendance access control, smart agriculture monitoring, or industrial quality inspection—and they will help you adjust the lens focal length, sensor model, or even modify the enclosure openings and interface definitions.
The benefits are obvious. First, you avoid spreading out high mold costs, because you're using existing structural parts, with at most a color change or silkscreen tweak. Second, the software and hardware integration has already gone through many rounds, so stability is much higher than developing from scratch. Third, the delivery cycle can usually be compressed to two weeks to a month, rather than the typical three-month minimum from an ODM. Of course, you also have to accept some limitations: the appearance can't be changed much, the motherboard layout is essentially fixed, and certain unconventional features (such as wireless video transmission or special protocol integration) may require trade-offs on the existing hardware.
On the flip side, if you insist on going with a traditional ODM, unless you have a stable expectation of thousands of units later and are willing to fold some of the mold costs into the initial unit price, it's hard to negotiate a reasonable price. The ODM's advantage lies in deep customization and scaled supply, but 100 units simply can't support its cost structure. As for those "light solution providers" that only sell core boards and hand you a pile of schematics to play with, unless your team has hardware engineers and embedded development capabilities, I'd advise against touching them.
So my recommendation is this: focus on solution providers that can deliver complete prototypes. Pay special attention to whether they have similar project cases and whether they're willing to accompany you through prototyping, testing, and fixing small bugs. A batch of 100 units is already enough to validate the market and product form. Once the order volume picks up later, you can consider approaching an ODM with mass production data. After all, at this stage, speed matters more than completeness, and stability matters more than savings.