Updated 4 October 2026. Written in April 2026 around Qwen2.5; the method holds. Today you would use small Gemma 4 models for vision, Llama 4 Scout (or Maverick) for real-time defect detection, and quantised DeepSeek V4 for complex inspections on local servers.
Manufacturing is one of the clearest cases for running AI at the edge. Factories generate sensor, camera and machine data every second, and sending all of it to the cloud adds latency, bandwidth costs and data-sovereignty questions.
What edge AI means for a factory
Edge AI means running models on hardware at the factory: on the line, inside the machine enclosure, or in a local rack. No data leaves the facility, and inference takes milliseconds, not the hundreds a cloud round-trip needs. That matters because:
- Latency costs parts: on a line running 120 units per minute, a part passes every 0.5 seconds. Capture, inference and the reject signal must fit that window; a network round-trip eats into it.
- Bandwidth costs scale: cameras recording all shift produce constant video. Uploading it all means paying bandwidth and storage per camera; at the edge, only results and flagged images leave the machine.
- Data stays under your control: keeping production data, including operator images, on-premises avoids international transfers under the GDPR. Your EU AI Act obligations depend on what the system is for, not where it runs.
Six use cases worth piloting
| Use case | What it does | Typical hardware | What the pilot should measure |
|---|---|---|---|
| Predictive maintenance | Vibration and temperature sensors flag bearing or motor faults early | Jetson Orin Nano + sensors | Unplanned stops avoided vs. false alarms |
| Visual quality control | Vision model spots surface defects and misalignment at line speed | Jetson AGX + industrial camera | Defects caught vs. manual inspection |
| Frontline worker support | Operators ask about procedures, safety or machine settings | Mac Mini M4 + tablet | Time to answer; answers checked against the manual |
| Internal defect detection | X-ray or infrared images checked for hidden defects | GPU server + specialised sensor | Detection rate on a labelled reference set |
| Energy optimisation | Finds avoidable HVAC, compressor and lighting use | Raspberry Pi 5 + smart meters | kWh per unit, before vs. after |
| Supply chain anomalies | Flags unusual inventory, delivery or supplier-quality patterns | Server + ERP integration | Anomalies confirmed by the buyer vs. noise |
ZEDEDA’s analysis of edge AI in manufacturing documents real deployments across automotive, electronics and food production.
Vision language models on the line
A Vision Language Model (VLM) can look at a photo of a weld seam and answer “Does this weld show any surface imperfection that ISO 5817 does not allow at quality level B?” A photo only shows the surface; internal imperfections still need radiographic or ultrasonic testing.
Open VLMs such as PaliGemma 2 and Qwen2.5-VL have small variants (around 3B parameters) that, quantized, fit in the 8 GB of shared memory of a Jetson Orin Nano. The 7B variants or LLaVA-NeXT run more comfortably on a Jetson AGX Orin or a GPU server. Instead of training a classifier per defect type, you describe the standard in plain text, then validate on labelled parts from your line.
How Spanish SMEs can start today
The Kit Digital programme has offered vouchers of up to EUR 12,000 for companies with 10 to 49 employees; check which calls are open before you plan around it. At EU level, the AI Factories initiative, with one site at the Barcelona Supercomputing Center, gives companies access to compute and support, and the Commission says it prioritises access for AI startups and SMEs.
The deployment path: pick one use case, select hardware, deploy a small model locally, connect sensors or cameras, validate on production data, then scale to more lines and measure the ROI.
Starter hardware kit
| Component | Option | Why |
|---|---|---|
| Inference device | NVIDIA Jetson Orin Nano Super 8 GB | 67 INT8 TOPS at 7-25 W (NVIDIA) |
| Local server (multi-model) | Mac mini M4, 16 GB or more | Unified memory for 7-14B language models |
| Camera (quality control) | Industrial GigE camera, e.g. Basler ace 2 | Fixed exposure and trigger input for the line |
| Sensor gateway | Raspberry Pi 5 + industrial HAT | Collects vibration and temperature data |
Prices change by configuration and month, so quote the kit on the day you buy; our hardware guide compares the devices. Steady daily use usually favours owned hardware; occasional use can be cheaper per token in the cloud. Run your numbers in the cloud vs. local cost calculator.
The EU AI Act angle
Checked against Regulation (EU) 2024/1689 as amended by the “Digital Omnibus on AI”, Regulation (EU) 2026/1744, in force since 27 July 2026.
An earlier version of this article called most factory AI “limited risk”. The Act has no such category: it has prohibited practices (Article 5), high-risk systems (Article 6, Annexes I and III), transparency duties (Article 50) and AI literacy (Article 4). Ask these questions about each system.
1. Is it a safety component?
Most use cases above are not. The new Article 6(1a) excludes AI used solely for non-safety “user assistance, performance optimisation, service efficiency, automation or convenience or quality control”. Article 6(1b) sets the limit: if its failure would endanger health and safety, it is a safety component. Flagging scratches, predicting bearing wear or tuning compressors normally stays outside the route to high risk; a vision system that stops a press when a hand enters the danger zone does not.
2. A safety function in machinery: now Annex I, Section B
Regulation 2026/1744 (Article 1, point 41) moved machinery from Annex I, Section A, to Section B, which now lists the Machinery Regulation, 2023/1230. In practice:
- Classification still happens. Article 6(1) covers all of Annex I, so a machine-learning safety function is still high-risk where the product needs third-party assessment, as the Machinery Regulation (Annex I, Part A, points 5 and 6) requires for self-evolving safety components.
- The AI Act’s high-risk obligations do not apply directly. For Section B products, the new Article 2(2) applies only Articles 6(1), 60a and 102 to 112.
- The requirements arrive through machinery law, by Commission delegated acts under amended Article 8 of the Machinery Regulation that “shall apply by 2 August 2028”. Until then, the new Article 20(10) presumes conformity for AI meeting the AI Act’s harmonised standards.
The content of those delegated acts is not known yet. What is fixed is who carries the obligation: the manufacturer of the machine or safety component. A factory that only uses a CE-marked machine is not in that role; one that builds or substantially modifies a safety function may be.
3. Other products still in Annex I, Section A
Explosive-atmosphere equipment (Directive 2014/34/EU), pressure equipment (Directive 2014/68/EU), lifts and personal protective equipment stay in Section A. AI that is a safety component of such a product needing third-party assessment is high-risk from 2 August 2028. Under the new Article 6(1c), an assessment required only for non-safety risks, such as radio spectrum, does not count.
4. Does it touch workers?
This is where factory AI most often becomes high-risk. Annex III, point 4(b), covers AI that monitors and evaluates workers, allocates tasks based on individual behaviour or traits, or decides on terms of work, promotion or termination. A camera system scoring each operator’s productivity falls here, from 2 December 2027.
Separately, Article 5(1)(f) prohibits inferring emotions at work, except for medical or safety reasons. Whether an operator-fatigue camera fits that exception depends on what it infers and why.
5. Does it talk to people, and what applies to everyone?
The frontline assistant talks to operators, so Article 50(1) requires the provider to design it so people know they are interacting with an AI, unless that is obvious. If you deploy it under your own name, you may be the provider.
Article 4, as replaced by 2026/1744, asks providers and deployers to support the AI literacy of staff who operate AI on their behalf, without guaranteeing any individual’s level.
Key dates for a factory
| Date | What applies | Provision |
|---|---|---|
| 2 February 2025 | Prohibited practices (incl. emotion inference at work), AI literacy | Art. 113(a) |
| 2 August 2026 | General application, incl. Article 50 transparency | Art. 113 |
| 14 January 2027 | Machinery Regulation applies | Reg. 2023/1230, Art. 54 |
| 2 December 2027 | Annex III high-risk, incl. worker monitoring | Art. 113(c)(i) |
| 2 August 2028 | Annex I high-risk (Section A); machinery AI delegated acts | Art. 113(c)(ii); Reg. 2023/1230, Art. 8 |
Where edge deployment helps, and where it does not
Local hardware does not change a system’s classification; the Act looks at intended purpose and your role. It does give you the evidence: you hold the logs, know what the model sees and can reproduce its outputs.
For GDPR, local is the cleanest answer, but if an integrator maintaining the system can access personal data, you still need a data processing agreement.
Next steps
- Pick one use case and write down what the pilot must measure.
- Check the Kit Digital guide for open calls.
- Check a system with the AI Act checker and read the key-dates overview.
Sources
- ZEDEDA: Edge AI in Manufacturing.
- European Commission: AI Factories (checked 2026-10-09).
- Regulation (EU) 2024/1689 (AI Act), Articles 4, 5, 6, 50, 113; Annexes I and III.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), Article 1, points (2), (5), (8), (40), (41); Article 3.
- Regulation (EU) 2023/1230 (Machinery Regulation), Articles 8, 20, 25, 54; Annex I, Part A.
Work with us
We deploy small models on edge hardware, agree up front what the pilot has to prove, and measure it. Book a 15-minute call or see how we work in consulting.