Total cost of AI deployment
The software licence is rarely the dominant cost in an edge AI project. The hardware mandate is. Neural networks deployed at the edge almost always require specialist compute that is not in the existing product. The cost of that silicon — and the qualification, toolchain migration, and power infrastructure that comes with it — is invisible at proof-of-concept stage and significant at production scale.
Neural networks vs Logic-Based Networks: the hardware cost.
The LBN cost advantage does not come from cheaper training or a lower licence fee. It comes from running the same inference on silicon already in the product. The capital cost comparison is between paying for new silicon and paying nothing.
| Cost category | Neural network (edge) | Logic-Based Network |
|---|---|---|
| New silicon required | Yes — NPU or GPU-capable SoC | No — existing MCU or CPU |
| BOM uplift per unit | $2–10+ at volume | $0 |
| SoC qualification cost (automotive/industrial) | £50,000–£200,000 | £0 |
| Toolchain migration | Often required | SDK integrates with existing toolchain |
| Development timeline | Extended (new platform) | Shorter (existing platform) |
The qualification cost row deserves emphasis. In automotive applications, the cost of qualifying a new SoC ranges from £50,000 to £200,000 in engineering alone, before regulatory costs or timeline impact. An automotive Tier 1 customer evaluating an ADAS application found that the neural network approach required a 200% SoC cost increase. The LBN approach ran on the existing PowerPC hardware already in the vehicle electronic architecture. The qualification cycle was unchanged. The SoC cost increase was zero. The commercial case that had been marginal became straightforward.
Cloud inference, energy, and field service costs.
Capital cost is the upfront story. Operating cost is the long-term one. For edge AI deployments at scale, operating costs accumulate across three dimensions: cloud inference fees for deployments that push processing off-device, energy costs for battery-powered nodes, and field service costs for systems that require regular maintenance.
Costs avoided in the field.
The figures below come from real deployments where neural network approaches were evaluated and their cost structures made them commercially unworkable at production scale. In each case, the cost problem was identified before full deployment, and a logic-based alternative was used instead.
Wastewater network: UK water utility
A UK water utility needed anomaly detection across 100,000 sewer monitoring sites. Blockages, overflows, and structural failures had to be detected before they caused service disruption. The neural network approach required either reliable cloud connectivity — unavailable across remote sewer infrastructure — or NPU-capable hardware at each site.
Deploying mains-power infrastructure at remote sewer sites cost approximately £15,000 per installation. At 100,000 sites, this amounts to £1.5 billion in power infrastructure alone, before sensor hardware, software, integration, or maintenance costs are counted. The business case for the neural network approach never closed.
An LBN anomaly detection model running on the existing sensor hardware — powered by a lithium battery at 455 µJ per inference — produces predictions every five seconds. Projected battery life is ten years. Mains infrastructure cost: zero. The economics that made neural network deployment impossible made LBN deployment straightforward.
The battery-powered AI section covers the field energy figures in detail.
Automotive ADAS: Tier 1 supplier
An automotive ADAS Tier 1 needed to replace a physical sensor with an AI model for vehicle dynamics classification. The target hardware was NXP PowerPC e200 processors already embedded in the vehicle electronic architecture — hardware that the customer could not change without triggering a full qualification programme.
The neural network approach required a new NPU-capable SoC at approximately 200% of the existing SoC cost — a cost increase that entirely negated the bill-of-materials saving that justified the sensor replacement in the first place. The system integrator's alternative approach also failed to meet the performance requirement on the existing hardware.
An R-LBN implementation ran on the existing PowerPC e200. Inference completed in four microseconds. Memory footprint: 4 KB. The SoC cost increase: zero. The qualification impact: none. The business case: restored.
For the technical detail on why LBNs are compatible with this class of hardware where neural networks are not, see AI without a GPU.
The cost structure at production scale.
Individual deployment costs are manageable. Fleet-level costs — the same cost structure multiplied across tens of thousands to millions of units — are where architecture decisions become commercially decisive. The choice between a model that requires new silicon and one that does not is worth, at large volume, the difference between a viable product and one that cannot be profitably made.
A hardware cost of $5 per unit is unremarkable in isolation. At one million units, it is $5 million of additional bill-of-materials cost. This figure appears in a product's cost model as a direct reduction in gross margin. For products competing on thin margins in commodity IoT markets, an additional $5 per unit in unplanned silicon cost is commercially significant — sometimes the difference between a profitable and loss-making product line.
The qualification cost is less easily quantified but equally real. Automotive and medical programmes that trigger a re-qualification cycle due to an unexpected hardware change absorb engineering resource, extend programme timelines, and may delay revenue. Engineering time spent on platform migration is time not spent on the application features the product was supposed to deliver.
Cloud inference costs follow a different profile. They are not upfront; they accumulate continuously. A sensor array making one inference per second, running 24 hours a day, generates 86,400 inference calls per sensor per day. At a modest cloud inference price of $0.001 per 1,000 calls, this costs $0.086 per sensor per day. At 10,000 sensors, $860 per day, $314,000 per year. The cost grows linearly with fleet size and inference frequency. On-device inference eliminates this line item entirely.
Energy cost compounds in a similar way for battery-powered deployments. A sensor requiring quarterly battery replacement at a median field service cost of £50 per visit generates £200 per sensor per year in service cost. At 10,000 sensors, £2 million per year. Extending the replacement interval from quarterly to annually saves £1.5 million per year at that fleet size. The service cost differential between a three-month and a ten-year battery life is, across a large infrastructure deployment, a programme-defining number.
Deploy AI on the hardware you already have.
ModelMill trains Logic-Based Networks and generates a C-code SDK that runs on any standard 32-bit processor. No NPU. No platform migration. The bill-of-materials cost of adding AI is zero.
01
Train your model
02
Generate C-code SDK
03
Deploy on existing hardware