EU AI Act for Semiconductor and Chip Companies: Why Most of Your AI Is Minimal-Risk

If you run a fab, design chips, or build the AI accelerators that other companies train models on, the honest headline is this: most of the AI in a semiconductor business is minimal-risk, with no mandatory obligations under the EU AI Act. The systems that actually carry duties are the same ones that carry them in any industry — workforce AI under Annex III point 4, and the rare case where your plant feeds critical infrastructure. The chips themselves, and the AI you use to make them, almost never trigger high-risk classification.

This matters because the assumption runs the other way. "We are at the centre of the AI supply chain, so the AI Act must land hard on us." It doesn't. The regulation classifies systems by use, not by how close a company sits to the silicon. A company that makes the GPUs the world trains models on can have a lighter AI Act profile than a 30-person HR-tech startup.

The two facts that resolve most semiconductor questions

Fact one: making chips that become AI accelerators does not make you an AI provider. Under Article 3 of Regulation (EU) 2024/1689, a provider is the actor that develops an AI system — software performing inference, generating outputs, making or supporting decisions — and places it on the market under its own name. A GPU, an NPU, a wafer, an ASIC, or a memory die is hardware. It is not an "AI system" within the meaning of Article 3(1). The customer who builds and ships a model on your accelerator may be a GPAI or system provider; you are their component supplier. Selling the engine is not the same as driving the car.

Fact two: the AI you use to make chips is operational, not regulated. Process-control AI that tunes etch and deposition parameters, yield-prediction models that flag a drifting tool, defect-detection vision on inspection stations, EDA and design-automation AI that places-and-routes a netlist — none of these appear in Annex III, none are prohibited under Article 5, and none are the limited-risk transparency cases of Article 50. They sit in the bottom tier: minimal risk, no compliance stack. Voluntary codes of conduct (Article 95) can apply; nothing mandatory does.

Classify by your own use — a four-question filter

Run every AI system in the business through these, in order:

1. Is it prohibited (Article 5)? Almost never in a fab — except one trap: AI that infers the emotional state of cleanroom operators (stress, frustration, engagement) is prohibited emotion recognition in the workplace under Article 5(1)(f), in force since 2 February 2025, unless it is intended for medical or safety reasons. Fatigue is not an emotion: recital 18 excludes physical states such as fatigue, so a drowsiness alert is not prohibited, though it can be high-risk under Annex III point 4(b) where it is used to monitor workers. The fine ceiling is €35,000,000 or 7% of total worldwide annual turnover, whichever is higher (Article 99(3)).
2. Is it a safety component of an Annex I product, or an Annex III use? This is where the real duties live (below).
3. Is it a limited-risk transparency case (Article 50)? A customer-facing chatbot on your support site, or synthetic media in marketing, needs disclosure — not the high-risk stack.
4. Otherwise — minimal risk. Document the reasoning and move on.

Where a semiconductor company actually picks up high-risk duties

There are only two realistic routes, and both are about use, not silicon.

Annex III point 4 — employment and worker management. AI that screens applicants for fab technician or engineering roles, ranks candidates, drives promotion or termination decisions, or allocates shifts through automated scoring is high-risk under Annex III point 4, exactly as it is for any employer. This is the single most common way a chip company lands in scope. It is a deployer duty when you buy the tool (Article 26) and a provider duty if you build it in-house (Article 16). The deadline is 2 December 2027 under the Digital Omnibus (originally 2 August 2026).

Annex III point 2 — critical infrastructure (narrow). Most fabs supply the general electronics market and are not critical infrastructure. But a facility whose AI manages a safety function in the supply of electricity, water, gas, or heating to a grid — for instance an on-site power or water-treatment plant feeding a public utility — can fall under Annex III point 2. The trigger is the safety-management function in critical infrastructure, not the value of the chips. A scheduling tool at a logic fab is not it; an AI controlling emergency shutdown across a grid-connected substation might be.

A note on the product route: if you embed an AI system as a safety component of a finished product covered by EU harmonised legislation in Annex I — say, machinery under the Machinery Regulation — that is the Article 6(1) product route, with a 2 August 2028 deadline. Since the Digital Omnibus, machinery sits in Annex I Section B, so the AI requirements reach it through delegated acts under the Machinery Regulation rather than through the AI Act directly. For most pure-play chipmakers this is theoretical; raise it only if you ship regulated equipment, not bare components.

The deadline picture (read this before you plan)

The high-risk deferral is law. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and has been in force since 27 July 2026. It moved the stand-alone Annex III high-risk deadline from 2 August 2026 to 2 December 2027, and the Annex I product-route deadline from 2 August 2027 to 2 August 2028. Each route now has one legal date to plan against.

Worked example: NordSilicon Fab GmbH, ~620 employees, ~€240M turnover

NordSilicon runs a 200mm analog fab in Saxony and licenses some in-house design IP. Its compliance officer inventories the AI in the building and runs each system through the filter.

• Process-control AI tuning plasma-etch recipes, yield-prediction flagging tool drift, defect-detection vision on the inspection line, and the EDA place-and-route assistant in the design group: all minimal-risk. No Annex III use, no safety-component role in a regulated product. No obligations.
• The power-management AI on the on-site substation: NordSilicon's substation is behind the meter and does not perform a safety-management function in the public electricity supply. Not Annex III point 2. Minimal. (Had it managed grid-side emergency shutdown, the verdict would flip.)
• The accelerator chips NordSilicon fabricates for a fabless customer who trains vision models on them: NordSilicon is a component supplier under Article 3, not the AI provider. The customer carries any model-provider duties. No AI Act obligation flows to NordSilicon from the chip itself.
• The applicant-screening tool HR bought to rank graduate-engineer candidates: high-risk, Annex III point 4. As the deployer (Article 26), NordSilicon must use it per the vendor's instructions, assign competent humans for oversight (Article 14 in operation), log use, and monitor outcomes. The vendor, as provider, holds the Article 11 / Annex IV technical documentation.
• The "fatigue-monitoring" system a line manager piloted via webcam turned out to score operators' stress and engagement as well: that part is emotion recognition under Article 5(1)(f), with no safety case behind it. The compliance officer shuts the pilot down; a drowsiness-only alert would have gone to the Annex III point 4(b) high-risk check instead.

Verdict: of NordSilicon's entire AI estate, exactly one system (applicant screening) carries the high-risk stack, one was prohibited and removed, and everything else is minimal-risk. That is the typical shape for a semiconductor company.

How Confir helps

A semiconductor business with dozens of AI systems needs a defensible record of why almost all of them are out of scope — not just an assertion that they are. Confir registers every AI system the company builds or deploys, classifies each under Articles 5 and 6 with Annex III logic through plain-English checklists, and derives the role — provider (Article 16) or deployer (Article 26) — for the few that matter. The engine is deterministic and rule-based: the same intake produces the same classification and names the rule that fired, so a "this is minimal-risk because it is operational process control, not an Annex III use" finding is reproducible and auditable. For the HR system that is high-risk, Confir structures the Article 11 / Annex IV documentation and the Article 47 Declaration of Conformity. No consultants, no multi-month rollout.

Related guides

• Article 6 high-risk classification
• What the EU AI Act is
• Critical infrastructure safety AI under Annex III
• Recruitment and screening AI as high-risk
• Manufacturing under the EU AI Act
• AI risk classification, explained