Cloud Vision API Pricing
Computer vision AI licensing begins with the cloud APIs, which cluster around $1.00–$1.50 per 1,000 images. Google Cloud Vision charges $1.50 per 1,000 units for the 1,001–5 million band, dropping to $1.00 above 5 million, after a free first 1,000 units per feature. Azure Computer Vision starts at $1.00 per 1,000 transactions, with Custom Vision inference at $2 per 1,000 ($0.002 per image) and training at $2 per compute hour; its per-inference model avoids paying for idle compute, which suits sporadic prediction. AWS Rekognition uses volume-tiered per-image pricing with 5,000 analyses free for the first 12 months.
At low volume these rates feel trivial. At scale they compound fast: 50 million images a month on a single feature is $50,000–$75,000 a year before any of the costs that follow. Treat the API rate as the entry fee, not the bill — the same total-cost discipline that runs through the AI contract negotiation deep dive and the consumption analysis in AI data pipeline licensing.
The Feature-Stacking Trap
The subtlest cost in cloud vision is feature independence. Providers price each feature separately, so a single image analysed for three things — say label detection, OCR and face detection — is billed three times. At Google Cloud Vision's $1.50 per 1,000 per feature, that image costs $4.50 per 1,000, not $1.50. Pipelines that run four or five detectors per frame quietly triple or quadruple the unit cost the business case assumed.
One image, three features, three charges. Before you sign, count the detectors per frame — the real unit price is the rate multiplied by the number of features you run, not the headline figure.
The fix is partly architectural — run only the detectors a use case needs — and partly commercial: negotiate bundled feature pricing or a blended per-image rate for multi-feature pipelines rather than accepting the stacked list price. This is the same "what unit am I really paying for" problem we dissect in conversational AI platform licensing.
Edge Deployment Changes the Economics
Most enterprise vision does not stay in the cloud. Running models at the edge — on factory floors, in retail stores, on vehicles or medical devices — changes the cost structure entirely, replacing per-image API fees with hardware, networking and operational burden. That is why full deployments range from about $10,000 for a proof of concept to over $500,000 for enterprise-grade systems, with most production projects landing between $50,000 and $250,000. The cost drivers are model complexity, data annotation, infrastructure, deployment environment and real-time processing needs — and the ongoing bill adds GPU compute, model retraining, monitoring, cloud storage and security maintenance.
The licensing implication is that the cloud API rate predicts very little about an edge programme's true cost. Buyers should model the full lifecycle — annotation, hardware, retraining cadence, monitoring — before negotiating, exactly as we recommend for custom models in AI fine-tuning costs and contracts and infrastructure in AI model hosting contracts.
The Subscription-Locked Camera Trap
The most expensive lock-in in computer vision is hardware. Cameras tied to subscription-locked systems can cost four to six times more per unit, adding roughly $90,000–$110,000 upfront for a 100-camera deployment. Worse, the five-year licence cost alone from a subscription-locked system often exceeds the total hardware investment of a complete open alternative — you pay a premium for the cameras and then pay again, every year, to keep them working.
The defence is to treat per-stream licensing as a negotiated term, not a fixed reality, and to keep an open, interoperable option credibly on the table. A documented alternative built on standards-based cameras is the single most effective lever against subscription-lock pricing — the same portability principle that underpins multi-model AI strategy.
Cloud vs Edge: Choosing the Model
The deployment model decides the cost structure, and the right answer is workload-specific. Cloud inference keeps initial costs low — you pay the per-image API rate (around $1.00–$1.50 per 1,000) with no hardware, and Azure's per-inference billing means you avoid paying for idle compute on sporadic workloads. It suits variable, lower-volume or experimental use cases where the convenience of a managed API outweighs the unit price.
Edge deployment inverts the economics. Running models on cameras, factory lines or devices removes the per-image fee but introduces hardware, networking, environmental and operational cost, plus ongoing GPU compute, retraining and monitoring. It earns its keep at high, sustained volume, where latency or data-residency rules forbid sending images to the cloud, or where connectivity is unreliable — and it is why most production programmes land between $50,000 and $250,000. The practical test is the crossover point: model your image volume against the cloud per-image rate, and the moment sustained throughput makes the cloud bill exceed amortised edge hardware, edge becomes the cheaper architecture. Decide this before negotiating, because the licence terms that matter differ sharply between the two.
What to Model and Negotiate
Four moves protect a computer vision budget. First, model the all-in cost — API or per-stream fees plus edge hardware, annotation, retraining, monitoring and storage — and negotiate against that number, not the rate card. Second, secure volume-tier pricing and a price lock so a multi-year programme does not absorb mid-term list increases, and push for bundled feature pricing on multi-detector pipelines.
Third, avoid proprietary camera lock-in by requiring interoperability or hardware portability, and benchmark the per-stream licence against an open alternative before you commit. Fourth, confirm data and model ownership: your annotated training data and any custom model are your assets, with export rights on exit and clear terms on where images are processed and retained — a point that matters for cost and for compliance alike. For the full clause set, work through the AI Procurement Checklist and the AI Contract Red Flags brief, benchmark the vision clouds via the AWS and Google Cloud hubs, and request a confidential briefing before signing a vision platform or camera contract.