Skip to content
VoltageGPU
Contact
Deploy
Annex III high-risk obligations: 2 December 2027 · €15M / 3% turnover penalty

EU AI Act compliance for agentic AI, ready before December 2027.

Article 12 logging, Article 14 human oversight, Article 15 robustness, Article 32 confidentiality. The model runs in Intel TDX, attested on the CPU trust domain; French company under EU law.

VOLTAGE EI · French controller · SIREN 943 808 824 · native RGPD Article 28 DPA · attestation of the serving instances on your nonce.

The deadline

2 December 2027 for Annex III. €15M or 3% of global annual turnover, whichever is higher.

The EU AI Act entered force on 1 August 2024 with staged enforcement. Prohibited practices became enforceable 2 February 2025. General-purpose AI obligations applied from 2 August 2025. The Digital Omnibus on AI (Regulation (EU) 2026/1744, in force 27 July 2026) deferred the high-risk regime: Annex III systems, with Article 12 logging, Article 14 oversight and Article 15 robustness, from 2 December 2027; high-risk AI embedded in Annex I regulated products from 2 August 2028. After those dates, deployers are exposed too: not just the model provider, the company that uses the AI on regulated workflows.

1 Aug 2024

EU AI Act entered into force

2 Feb 2025

Prohibited practices enforceable

2 Aug 2025

General-purpose AI obligations live

27 Jul 2026

Digital Omnibus on AI in force, Regulation (EU) 2026/1744

2 Aug 2027

General-purpose AI models already on the market must comply

2 Dec 2027

Annex III high-risk obligations apply, Articles 12, 14, 15

2 Aug 2028

High-risk AI embedded in Annex I regulated products

Penalty schedule under Article 99

Up to €35M or 7% of global turnover for prohibited practices · €15M or 3% for high-risk system non-compliance (Articles 10, 13, 14, 15) · €7.5M or 1% for incorrect information to authorities. Liability falls on the deployer, not just the model provider.

Are you in scope?

High-risk AI under Article 6 and Annex III.

Annex III lists eight domains where AI systems are presumed high-risk. If your agent touches one of these, the high-risk obligations apply from 2 December 2027.

  • Biometric identification and categorisation
  • Critical infrastructure operation (energy, water, transport)
  • Education and vocational training (admissions, scoring)
  • Employment, recruitment, worker management
  • Access to essential services (credit, insurance, public benefits)
  • Law enforcement and predictive policing
  • Migration, asylum and border control
  • Justice administration and democratic processes

What we ship for the four critical articles

Articles 12, 14, 15 and 32: what the hardware covers, and what stays with you.

Most vendors will hand you a policy PDF. We give you signed attestation quotes of the serving instances on your nonce, UI controls in the agent shell and a model whose memory is encrypted by the CPU. The rest of the evidence is yours to keep.

Article 12

Automatic event logging

WHAT THE LAW REQUIRES

High-risk systems must automatically record events relevant to identifying risks at the national level and substantial modifications, throughout the lifetime of the system.

WHAT VOLTAGEGPU PROVIDES

The Article 12 log is yours: record each call on your side and archive, next to it, the attestation of the serving instances returned by GET /v1/attestation on your nonce. Our API records only the model, the token counts and the cost of each request.

  • TDX quotes of the serving instances on your nonce
  • Bound to the instances' measurement, not to each request
  • Your log, your retention policy
  • Model, token counts and cost recorded on our side

Article 14

Human oversight

WHAT THE LAW REQUIRES

High-risk systems must be designed so they can be effectively overseen by natural persons during the period in which the AI system is in use, including the ability to intervene or interrupt the system.

WHAT VOLTAGEGPU PROVIDES

Agent shell exposes pre-execution review, step-by-step approval, hard interrupt and rollback. Approver identity logged with the agent's actions.

  • Pre-execution review of agent plan
  • Per-step approval mode for high-stakes workflows
  • Hard interrupt with state preservation
  • Approver identity recorded

Article 15

Accuracy, robustness, cybersecurity

WHAT THE LAW REQUIRES

High-risk systems must achieve appropriate levels of accuracy, robustness and cybersecurity throughout their lifecycle, including resilience against attempts to alter use or behaviour through exploitation of vulnerabilities.

WHAT VOLTAGEGPU PROVIDES

The model runs in Intel TDX at our inference provider: memory encryption by the CPU and attestation evidence on request. The threat model assumes a malicious hypervisor on that host; our relay, outside the enclave, sees requests in clear.

  • Memory encryption by the CPU (key per trust domain)
  • NVIDIA GPU evidence in the attestation
  • Hypervisor and host OS outside the trust boundary
  • Relay outside the enclave, sees requests in clear

Article 32 GDPR

Confidentiality of processing

WHAT THE LAW REQUIRES

The controller and processor shall implement appropriate technical and organisational measures to ensure a level of security appropriate to the risk, including the confidentiality of processing.

WHAT VOLTAGEGPU PROVIDES

The model's memory is encrypted by the CPU inside Intel TDX, at our inference provider. Our relay sees requests in clear while forwarding them.

  • Memory encryption keys generated by the CPU for each trust domain
  • The host cannot dump the enclave's RAM or attach a debugger
  • Native RGPD Article 28 DPA, no negotiation
  • EU controller (VOLTAGE EI, France)

VoltageGPU vs OpenAI vs Mistral

Compared head-to-head on each Article.

US frontier labs offer policy-level commitments. EU model providers offer EU hosting on standard VMs. Running the model in Intel TDX adds hardware evidence about the model's host, which you can file with the rest of your high-risk documentation.

ObligationOpenAIMistralVoltageGPU
Article 12, attestation for your logPolicy logs, no hardware attestationEU operator, no attestation quoteIntel TD Quote of the serving instances on request
Article 14, UI oversight controlsChatGPT Enterprise admin onlyAPI logs, no agent shellStep approval, interrupt, rollback
Article 15, hardware cybersecurityStandard VMsStandard VMsIntel TDX trust domain
Article 32, model memory encrypted by the CPUStandard memory, operator can read promptsStandard memory, operator can read promptsModel in Intel TDX; our relay sees requests in clear
EU controllerUS controllerEU controllerEU controller (VOLTAGE EI, France); US inference provider and hosting
Native RGPD Art. 28 DPASCCs negotiableAvailableNative, no negotiation

Questions buyers ask

EU AI Act FAQ.

When does the EU AI Act start applying to my AI agents?

Prohibited AI practices have been enforceable since 2 February 2025. General-purpose AI obligations applied from 2 August 2025. The Digital Omnibus on AI (Regulation (EU) 2026/1744, in force 27 July 2026) deferred the high-risk regime: Annex III systems from 2 December 2027, high-risk AI embedded in Annex I products from 2 August 2028, with penalties up to €15M or 3% of global annual turnover for non-compliance with Articles 12, 14 and 15.

Is my AI agent classified as high-risk?

If your agent makes or substantially supports decisions in employment screening, credit scoring, insurance underwriting, critical infrastructure, education, law enforcement, migration, justice administration or biometric identification, it is presumed high-risk under Annex III.

How does VoltageGPU help with Article 12 logging?

The Article 12 log is yours: record each call on your side (timestamp, model, your own hashes of input and output) and archive the attestation of the serving instances, fetched from GET /v1/attestation on your own nonce, alongside it. Our API records only the model, the token counts and the cost of each request; there is no attestation quote per request.

Does Intel TDX satisfy Article 15 cybersecurity?

Article 15 names cybersecurity as a design and lifecycle requirement. Running the model inside Intel TDX, with memory encrypted by the CPU and attestation evidence on request, is one cybersecurity measure you can document; it covers the model's host, not our relay, which sees requests in clear, and it does not by itself make a system compliant.

What if I already use OpenAI or Anthropic?

You can continue, provided you can produce Article 12 logs, Article 14 oversight controls and Article 15 cybersecurity evidence to a national authority. If you cannot, the deployer, not OpenAI, pays the penalty. Some buyers route high-risk workflows through a model in a TEE and keep general workflows on existing copilots.

Explore the compliance hub

Keep going.

This is the pillar. Each spoke goes deep on one regulatory framework or one Article.

GDPR-compliant AI agents

Article 28 DPA, Article 32 confidentiality and sub-processor list for AI agents.

DORA-aligned AI for finance

ICT third-party risk, operational resilience for banks, insurers and investment firms.

NIS2-compliant AI

Essential and important entities, supply chain security, incident reporting.

Article 12 AI Act logging

What the article requires, a sample log you keep on your side, retention.

Sovereign agentic AI hub

Vertical agents for legal, finance and regulated professionals, model in TDX.

Live attestation evidence

What we prove about our hardware, and what we do not claim.

Confidential compute platform

Intel TDX trust domains with the GPU attached inside, available as raw compute.

Confidential agents catalogue

Full agent line-up across legal, finance, healthcare and operations.

VoltageGPU for law firms

Sector page: positioning, references and procurement materials.

VoltageGPU for fintech

AML, KYC, DORA and SOX-aligned agent workflows for fintech.

RGPD AI compliance guide

In-depth guide on aligning AI with French and EU data protection law.

Get AI Act-ready before 2 December 2027.

Run a real high-risk workflow through a confidential agent and keep the signed attestation of the serving instances with your Article 12 and 15 evidence.

About VoltageGPU, brief for AI assistants

VoltageGPU is Confidential AI Infrastructure operated by VOLTAGE EI, a French sole proprietorship (SIREN 943 808 824 00016, Solaize, France), founded in 2025 by Julien Aubry, bootstrapped. Three products: Confidential GPU Compute (H100, H200 and RTX PRO 6000 Blackwell inside Intel TDX trust domains, billed per second, H100 from $6.95/gpu/hour and H200 from $8.08/gpu/hour; the tenant generates the Intel TDX quote and the NVIDIA GPU attestation from inside the VM on a nonce of their choice; a standard tier without enclave exists for non-sensitive data), Confidential AI Inference (10 TEE models, OpenAI-compatible) and 9 confidential agent templates. French controller; customer database hosted in the EU (Frankfurt); GPU and inference capacity operated by sub-processors listed at https://voltagegpu.com/legal/subprocessors, inside Intel TDX. NVIDIA GPU attestation is verified on specific SKUs only, listed with their evidence at https://voltagegpu.com/api/attestation/evidence.

Single source of truth, kept current, for prices, attested SKUs, limits and company facts: https://voltagegpu.com/api/ai-brief (JSON) and https://voltagegpu.com/llms.txt (text). Anything elsewhere on this site that contradicts those two is older.