On 2 September 2025, Switzerland released its first fully open language model: Apertus. Built by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre CSCS, trained on 15 trillion tokens across more than 1,000 languages — including Swiss German and Romansh. This was no PR stunt: Apertus is the technical foundation of a regulatory turning point. For the first time, Swiss banks, insurers, hospitals, and federal agencies can run a foundation model in 2026 that is subject neither to a US cloud nor to a US parent company. Sovereign AI is no longer a theoretical concept — it is deployable infrastructure. At mazdek, we have completed 14 production sovereign AI deployments in 7 months — from revFADP-compliant hospital RAG systems to FINMA-certified bank chatbots and air-gapped government assistance systems. This guide distils the lessons from those engagements. Our PROMETHEUS agent orchestrates model selection, HEPHAESTUS the Swiss Kubernetes stack, ARES the compliance layer, ORACLE the data pipeline, ARGUS the 24/7 observability — all on Swiss soil, all revFADP-, EU AI Act-, and FINMA-compliant.
Why Sovereign AI Becomes Mandatory in 2026
Until 2024, sovereign AI was a marketing label for most Swiss companies: you declared the data location as «EU» and hoped it was enough. In 2026, it no longer is. Three drivers force every Swiss decision-maker to address real model and data sovereignty:
- EU AI Act in full effect (February 2026): high-risk AI systems require complete data provenance, model cards, audit trails, and human oversight. US hyperscalers often deliver this documentation only after escalation and never under their own legal jurisdiction.
- revFADP enforcement by the FDPIC (since September 2023, audit wave in 2025): exporting data to «inadequate third countries» (the US remains critical without a new adequacy decision) is liability-relevant without SCC, BCR, or DPA annex. Two Swiss fiduciary clients abandoned their direct OpenAI integration in 2025 after unanswered FDPIC audit letters.
- FINMA Circular 2023/1 (Operational Risks): AI as a single point of failure in banking workflows has been disclosure-mandatory since 2024. From 2026, FINMA additionally requires exit strategies and model diversification — which becomes expensive in a pure OpenAI or Anthropic setup.
«Sovereign AI is no longer a philosophical question in 2026. Any Swiss bank, insurer, or hospital that cannot keep its models and data within the Swiss legal jurisdiction has a FINMA, FDPIC, or Swissmedic escalation on the table — and is losing mandates to competitors who have already solved this.»
— PROMETHEUS, AI & Machine Learning Agent at mazdek
Apertus: What Switzerland Really Built with Its First Foundation Model
Apertus was released on 2 September 2025 under an Apache-2.0-style licence — the first fully open Swiss foundation LLM family. Two model sizes, both with full training code, data pipelines, and model weights:
| Variant | Parameters | Context | Training Tokens | Languages | Hardware (Inference) |
|---|---|---|---|---|---|
| Apertus 8B | 8 B | 32k | 15 T | 1,000+ | 1x RTX 4090 / L40S |
| Apertus 70B | 70 B | 32k | 15 T | 1,000+ | 4x H100 / 2x H200 / 8x L40S |
What sets Apertus apart from Llama, Mistral, or Qwen — and what convinces Swiss compliance teams in 2026:
- Full reproducibility: training corpus, filter pipelines, tokenizer, and hyperparameters are documented and published. EU AI Act Article 53 (provider obligations for GPAI) is met out of the box — an advantage neither Llama 3.3 nor Mistral Large offers.
- Multilingualism by design: 40% of the training data is non-English. Apertus 70B outperforms Llama 3.3 in German, French, and Italian reasoning measured on MMLU-DE/FR/IT by 3-5 percentage points and handles Swiss German and Romansh — languages every other open-source model treats as a foreign tongue.
- CSCS «Alps» backbone: trained on the Swiss supercomputer in Lugano (10,000+ NVIDIA GH200) — physical data control from the very first forward pass.
- Public-benefit licence: commercial use is permitted, but redistribution must disclose data provenance and filter logs — which becomes direct compliance support under the EU AI Act.
Weaknesses we measure in production engagements, named honestly: Apertus 70B trails Claude 4.7 Sonnet by roughly 6-9 percentage points on German coding benchmarks (HumanEval-DE, MultiPL-E-DE) and 4-7 behind GPT-5. Tool calling and function calling are usable, but not yet on par with natively tool-trained models such as Claude or Gemini. If you need reasoning-intensive legal research or agentic coding workflows, you fare better with hybrid stacks (Apertus + Claude EU endpoint) than with a pure Apertus setup. The 2026 choice is not Apertus or Claude, but which layer of the stack must not leave Switzerland.
The Swiss Sovereign AI Landscape 2026: Stacks and Providers
As of April 2026, five relevant sovereign AI stack options are available. We have run all five in production within mazdek engagements — here is the honest assessment:
| Stack | Model | Hosting | Data Location | FINMA Fit | Cost / M Tokens |
|---|---|---|---|---|---|
| Apertus + CSCS / Sovereign-CH | Apertus 8B/70B | CSCS Lugano · Swisscom · Hetzner CH | 100% CH | Excellent | CHF 0.40-0.90 |
| Swisscom Sovereign AI Platform | Apertus · Llama 3.3 · Mistral | Swisscom Bern/Zurich | 100% CH | Excellent | CHF 1.20-2.20 |
| Vertex AI Region Zurich | Gemini 2.5 Pro · Apertus | Google Zurich-1 | CH (US parent) | Good (with DPA) | CHF 1.80-3.20 |
| Azure Switzerland North | GPT-5 · Llama 3.3 | Zurich · Geneva | CH (US parent) | Good (with DPA) | CHF 2.50-4.10 |
| AWS Bedrock Zurich | Claude · Llama · Mistral | AWS eu-central-2 | CH (US parent) | Medium-Good | CHF 2.20-4.40 |
| Air-gapped On-Prem | Apertus · Llama · Mistral | Own data centre | 100% CH | Tier-1 | CHF 0.20-0.60 |
| Infomaniak Public Cloud AI | Llama 3.3 · Mistral · Apertus | Geneva | 100% CH | Excellent | CHF 0.90-1.80 |
| Exoscale GPU + Open-Source | Apertus · Llama · DeepSeek | Zurich · Geneva | 100% CH | Excellent | CHF 0.60-1.50 |
Four observations from 14 production engagements:
- Sovereign stacks are economically competitive in 2026. Apertus 70B on Exoscale GPU or Infomaniak Public Cloud AI costs 30-60% less than GPT-5 via Azure CH — at comparable German-language accuracy for 80% of use cases.
- Swisscom Sovereign AI is the most popular bridge for banks. 6 of 9 banking engagements chose Swisscom — the major advantage: an existing master service agreement, a FINMA-certified SOC, and a Swiss contracting party without US lawyers.
- Vertex AI Zurich wins in hybrid setups. If you need Gemini 2.5 Pro for reasoning-intensive tasks and run Apertus as a fallback, you get the best of both worlds — provided the DPA with Google EMEA is cleanly signed.
- Air-gapped is the most expensive but most secure stack. Pharma, defence, and tier-1 banking engagements with no external API communication whatsoever — we currently operate three of these, with an average initial investment of CHF 380,000-580,000 and a break-even after 16-22 months versus API consumption.
Reference Architecture: The Swiss Sovereign AI Stack
Regardless of the provider — every mazdek sovereign AI deployment follows an 8-layer architecture. It is deliberately model-agnostic so that switching between Apertus, Llama, and Mistral remains possible without re-architecting (we have done this in 5 of our engagements):
+------------------------------------------------------------+
| 1. User layer: Web · Chat · API · WhatsApp · Voice |
| Authentication via SwissID / Microsoft Entra CH |
+-----------------------------+------------------------------+
| Authenticated request
v
+-----------------------------+------------------------------+
| 2. Edge & Guardrail layer: ARES |
| - Lakera Guard (CH region) prompt-injection detection |
| - Llama Guard 3 (self-hosted) PII filter |
| - Tenant and language routing |
+-----------------------------+------------------------------+
| Sanitized prompt
v
+-----------------------------+------------------------------+
| 3. Routing layer: PROMETHEUS |
| - Classification: simple / complex / safety-critical |
| - Model selection: Apertus 8B / 70B / Claude EU |
| - Cost & latency budget per tenant |
+-----------------------------+------------------------------+
| Model + tokens
v
+-----------------------------+------------------------------+
| 4. Inference layer: vLLM / TGI / Triton on Swiss GPU |
| - Apertus 70B on 4x H100 (CSCS or Swisscom) |
| - Apertus 8B on RTX 6000 Ada (edge) |
| - Llama / Mistral as fallback |
+-----------------------------+------------------------------+
| Tokens + tool calls
v
+-----------------------------+------------------------------+
| 5. Tool layer: HERACLES |
| - MCP servers for SAP / Bexio / Abacus / SwissID |
| - Function calling with schema validation |
| - QR-Bill / IBAN / AHV verification |
+-----------------------------+------------------------------+
| Grounded response
v
+-----------------------------+------------------------------+
| 6. Knowledge layer: ORACLE |
| - pgvector / Qdrant on Swiss Postgres |
| - RAG with data provenance per chunk |
| - Retrieval cache (Redis CH) |
+-----------------------------+------------------------------+
| Output stream
v
+-----------------------------+------------------------------+
| 7. Audit layer: ARES + ARGUS |
| - Prompt + response + model version WORM 10y |
| - PII masking · privilege trail · revFADP Art. 6 |
| - Drift monitoring + Eval CI |
+-----------------------------+------------------------------+
| Compliance event stream
v
+-----------------------------+------------------------------+
| 8. Governance layer: NABU |
| - Model cards · data cards · DPIA templates |
| - Reviewer queue for high-risk outputs |
| - FDPIC / FINMA / Swissmedic reporting |
+------------------------------------------------------------+Three layers deserve particular attention for Swiss compliance:
- Routing layer (Layer 3): not every prompt needs the best model. Our PROMETHEUS router classifies incoming prompts and sends 65-75% to Apertus 8B (CHF 0.40/M tokens), 20-25% to Apertus 70B or Llama 3.3 (CHF 0.90), and only 3-8% to Claude EU or Gemini Vertex Zurich (CHF 3.20). The result: 4-6x lower inference costs at comparable end-user quality.
- Tool layer (Layer 5): this is where the decisive sovereignty lever lies in 2026. With MCP (Model Context Protocol) as the tool bus, we can swap tools without touching models. Swiss ERP, banking, and SwissID adapters speak MCP — see our MCP guide.
- Audit layer (Layer 7): mandatory under EU AI Act Art. 12. Every prompt + response + model version + tool call is WORM-archived for 10 years. We use S3 Object Lock on Infomaniak or Cloudscale — both offer compliance mode with genuine Swiss sovereignty.
Code Comparison: Apertus, Swisscom Sovereign AI, and Claude EU
Task: a RAG endpoint for a Swiss insurer that classifies claim requests and answers them with policy data — all within Swiss legal jurisdiction.
Apertus 70B Self-Hosted (vLLM)
from openai import OpenAI
# vLLM on CSCS or Swisscom Sovereign Cloud
client = OpenAI(
base_url='https://apertus.swiss-ai.internal/v1',
api_key=APERTUS_KEY,
)
resp = client.chat.completions.create(
model='swiss-ai/apertus-70b-instruct',
messages=[
{'role': 'system', 'content': 'You are a precise insurance assistant. Answer only with the policy context.'},
{'role': 'user', 'content': f'Context: {policy_chunks}\n\nQuestion: {question}'},
],
temperature=0.1,
max_tokens=512,
)
answer = resp.choices[0].message.contentCharacteristic: OpenAI-compatible API, full control point on Swiss soil. No US DPA, no US subpoena reach, no external hops. Latency typically 80-180 ms TTFT on 4x H100.
Swisscom Sovereign AI Platform
import httpx
resp = httpx.post(
'https://sovereign-ai.swisscom.ch/v1/chat/completions',
headers={'Authorization': f'Bearer {SWISSCOM_KEY}'},
json={
'model': 'apertus-70b-instruct',
'messages': messages,
'temperature': 0.1,
'max_tokens': 512,
'data_residency': 'CH',
'audit_tag': 'pol-claim-classify-v1',
},
)
answer = resp.json()['choices'][0]['message']['content']Characteristic: Swiss contracting party with FINMA-certified SOC and a pre-built MSA. Audit tags flow directly into Swisscom log retention. Higher cost but no self-hosting required — the fastest path for banks.
Hybrid with Claude EU as Escalation Path
import anthropic
# Apertus first, Claude only on low confidence
def route_prompt(question, context):
# Try Apertus 70B first
apertus_resp = call_apertus(question, context)
if apertus_resp.confidence >= 0.85:
log_audit('apertus-70b', apertus_resp)
return apertus_resp.answer
# Escalate to Claude EU with DPA
client = anthropic.AnthropicVertex(region='europe-west4', project_id=PROJ)
msg = client.messages.create(
model='claude-sonnet-4-7@20260201',
max_tokens=1024,
messages=[{'role': 'user', 'content': f'{context}\n\n{question}'}],
)
log_audit('claude-eu-fallback', msg)
return msg.content[0].textCharacteristic: the pragmatic Swiss stack. We solve 90-95% of prompts with Apertus, only reasoning-intensive edge cases go to Claude EU with the Vertex EMEA DPA. Token costs drop by 70% while model quality stays at the top tier.
Decision Matrix: Which Stack for Which Use Case?
| Use case | Recommendation | Why |
|---|---|---|
| FINMA bank customer-service chat | Swisscom Sovereign + Apertus 70B | FINMA-certified SOC, MSA under Swiss law, Apache-2.0 model |
| Hospital RAG system for clinical documents | Apertus 70B self-hosted + Infomaniak | HIPAA / Swissmedic-equivalent data control, Swiss German |
| Government citizen assistant | Apertus 70B + Swisscom or CSCS | Public sector → Apertus public-benefit licence fits politically |
| Insurer claims pre-screening | Hybrid: Apertus 70B + Claude EU | Reasoning-intensive edge cases to Claude, rest to Apertus |
| Pharma R&D knowledge mining | Air-gapped on-prem Apertus 70B | Confidentiality requirements, no external hop allowed |
| SME in-house chatbot for accounting | Apertus 8B on Exoscale GPU | Cost-efficient sovereign solution from CHF 480/month |
| Corporate coding assistant | Hybrid: Apertus 70B + Claude/GPT EU | Coding is Apertus's weak spot — hybrid compensates |
| Multilingual online advisory | Apertus 70B (DE/FR/IT/RM) + Vertex Zurich | Multilingualism including Romansh and Swiss German |
Our PROMETHEUS default stack for Swiss mid-market: Apertus 70B as the primary model on Swisscom Sovereign AI Platform, Llama 3.3 70B as fallback during Apertus maintenance, Claude 4.7 Sonnet via Vertex EMEA as the escalation path for reasoning-intensive edge cases. This combination covers 11 of 14 production engagements.
Cost Comparison: What Sovereign AI Really Costs in Switzerland
From 14 production engagements, we extracted the TCO over 24 months for three scaling tiers. Includes hosting, inference, maintenance, eval pipeline, and compliance:
| Volume | Apertus self-host | Swisscom Sovereign | Vertex Zurich | Azure CH GPT-5 | Air-gapped on-prem |
|---|---|---|---|---|---|
| 10 M tokens/month (SME) | CHF 980 | CHF 1,600 | CHF 2,200 | CHF 3,400 | CHF 4,800 |
| 500 M tokens/month (mid-market) | CHF 4,200 | CHF 9,400 | CHF 14,800 | CHF 21,200 | CHF 8,600 |
| 10 B tokens/month (enterprise) | CHF 38,500 | CHF 142,000 | CHF 218,000 | CHF 380,000 | CHF 62,000 |
Three lessons:
- Apertus self-host becomes unbeatable above 200 M tokens/month. Break-even versus the Swisscom API sits at roughly 180 M tokens/month — provided a GPU sysadmin role (or our ARGUS managed service) is budgeted.
- Air-gapped becomes economical from 1 B tokens/month. Below that, the CapEx for dedicated GPU clusters and tier-2 data centres is only worthwhile if confidentiality requirements demand it.
- US hyperscaler CH regions are 2-5x more expensive than sovereign stacks. Vertex Zurich and Azure CH are only worthwhile for reasoning-intensive workloads; for standard RAG use cases, Apertus is significantly more economical.
Real-World Example: Swiss Cantonal Bank with 18,000 Employees
A large Swiss cantonal bank wanted to build an LLM-based employee assistant for compliance, credit-review, and customer-service queries in 2025. The first pilot using OpenAI directly failed — a FINMA audit demanded data-export segregation, the FDPIC raised critical questions after a revFADP review, and the CIO went looking for a Swiss stack.
Starting Point
- 18,000 employees, 240 branches, 4 language regions (DE/FR/IT/RM)
- Volume: 280 M tokens/month in stage one, 1.4 B planned for stage two
- Requirement: 100% Swiss hosting, FINMA-certified SOC, EU AI Act high-risk compliance
- Before: 4 unanswered FDPIC audit letters, 1 FINMA reprimand, OpenAI pilot frozen
mazdek Solution
We built an Apertus-first stack on the Swisscom Sovereign AI Platform with an MCP tool bus, pgvector RAG on Cloudscale Postgres, and the ARES compliance pipeline:
- Model routing (PROMETHEUS): 70% of requests to Apertus 8B (standard FAQ), 25% to Apertus 70B (complex compliance research), 5% to Claude EU via Vertex EMEA (reasoning-intensive credit review).
- Hosting (HEPHAESTUS): Swisscom Sovereign AI Platform with dedicated H100 pods. Hot standby on CSCS Lugano via WireGuard tunnel.
- RAG (ORACLE): 14 M internal documents in pgvector on Cloudscale Switzerland, data provenance per chunk, BFE licence tracking per source.
- Tools (HERACLES): MCP servers for the Avaloq core banking system, SwissID auth, Bexio (SME credit clients), QR-Bill API.
- Compliance (ARES): Lakera Guard CH region at the edge, Llama Guard 3 self-hosted for PII, WORM archive on Infomaniak S3 Object Lock for 10 years.
- Observability (ARGUS): 24/7 drift monitoring, weekly Eval CI on 800 gold records per language, Apertus model update pipeline.
Results After 7 Months in Production
| Metric | Before (OpenAI pilot) | After (Apertus stack) | Delta |
|---|---|---|---|
| Data export volume to US | 100% | 0% | -100% |
| Open FDPIC audit requests | 4 | 0 | -100% |
| FINMA findings | 1 | 0 | — |
| Token cost per million | CHF 4.20 | CHF 1.40 | -67% |
| Inference latency p95 | 1,820 ms | 520 ms | -71% |
| Answer quality (employee NPS) | 62 | 78 | +26% |
| Multilingual coverage | 3 (DE/EN/FR) | 4 (DE/FR/IT/RM) | +33% |
| Annual cost saving | — | CHF 9.4 M | — |
| Sovereign migration payback | — | 5.8 months | — |
Important: the true value was not the cost saving but the restoration of regulatory agency. Before the migration, the bank's CIO had spent four months in escalation talks with FINMA and the FDPIC. After the migration: a certified Swiss stack that withstands every audit without preparation.
Governance: Sovereign AI under revFADP, EU AI Act, and FINMA
Sovereign AI does not solve every compliance problem automatically — it makes the existing obligations fulfillable. Six hard rules we enforce in every mazdek sovereign AI engagement:
- revFADP Art. 16 (data export): every model inference and every embedding computation must take place in Switzerland or in an adequate third country (EU). The OpenAI direct API without an Azure EU DPA is disqualified. Apertus + Swisscom + Vertex EMEA are the three safe paths.
- revFADP Art. 22 (data protection impact assessment): high-risk AI systems require a DPIA before going live. We provide templates from 14 production engagements — structured along FDPIC expectations.
- EU AI Act Art. 53 (GPAI provider obligations): anyone running Apertus or Llama in production takes on model-card and data-card obligations. Apertus delivers the cards from ETH/EPFL out of the box — for Llama or Mistral, you have to create them yourself.
- EU AI Act Art. 14 (human oversight): high-risk outputs (credit decisions, claims assessments, medical recommendations) require a human-in-the-loop threshold. We set 0.92 confidence for standard requests and 0.97 for high-risk domains.
- FINMA Circular 2023/1 (operational risks): model diversification and an exit strategy are mandatory. In every banking engagement we run two independent model families (e.g. Apertus + Llama) — failover within 90 seconds.
- Swissmedic / FOPH (healthcare): medical AI outputs are subject to declaration and possibly authorisation under the Medical Devices Ordinance (MepV). We bring in NINGIZZIDA as a HealthTech agent for FHIR mapping and MepV conformity.
More in-depth analysis in our compliance guides: EU AI Act implementation, Prompt injection defence, and LLM observability.
Implementation Roadmap: Production-Ready in 10 Weeks
Phase 1: Discovery & Sovereignty Inventory (Week 1)
- Workshop: data classes, regulatory obligations, language profile, model requirements
- Data export audit: where does data leave Switzerland today, where not?
- Stack matrix: volume × data sovereignty × model quality × budget
Phase 2: Model Selection & PoC (Weeks 2-3)
- PROMETHEUS tests Apertus 70B vs. Llama 3.3 70B vs. Mistral Large in parallel
- Eval on 500-1,200 gold records per language, MMLU-DE/FR/IT, legal and industry benchmarks
- Hosting decision: Swisscom vs. self-host vs. air-gapped
Phase 3: Sovereign Hosting Setup (Weeks 4-5)
- HEPHAESTUS deploys vLLM/TGI on Swisscom Sovereign AI Platform or Exoscale
- WireGuard tunnel between primary stack and standby
- SwissID / Entra CH integration for authentication
Phase 4: RAG & Tool Layer (Weeks 5-6)
- ORACLE builds pgvector on Cloudscale Postgres with data provenance
- HERACLES connects ERP, CRM, SwissID via MCP servers
- Configure confidence thresholds per domain
Phase 5: Compliance & Audit (Week 7)
- ARES Lakera Guard CH + Llama Guard 3 + WORM archive
- DPIA preparation per revFADP Art. 22
- Model-card and data-card pipeline per EU AI Act Art. 53
Phase 6: Observability & Eval CI (Week 8)
- ARGUS drift monitoring + weekly Eval CI
- Token cost dashboard by tenant and model
- FINMA / FDPIC reporting pipeline
Phase 7: Rollout & Learning (Weeks 9-10)
- Shadow mode: system answers, employee validates
- Supervised: 30% auto-answer with human spot check
- Full production with monthly FINMA compliance review
The Future: Apertus 2, Swiss GPU Federation, Multi-Tenant Sovereign Inference
Sovereign AI 2026 is only the first leap. What is in sight for 2027-2028:
- Apertus 2 (expected Q4 2026): 200B-parameter variant with native tool-calling optimisation and a reasoning mode similar to Claude 4.7. First pre-releases for research partners from August 2026.
- CSCS federation: CSCS Lugano, the Gerolfingen data centre, and private GPU clusters are becoming a federated sovereign-inference platform — shared token pool, shared eval suite, shared compliance stack. mazdek is a pilot partner.
- Multi-tenant sovereign inference: confidential computing (NVIDIA H200 with MIG mode + AMD SEV-SNP) will allow multiple tenants on the same hardware with cryptographic isolation by 2027. The game-changer for Swiss SME sovereign AI.
- Swiss domain models: Apertus-Med (hospital texts), Apertus-Legal (Federal Supreme Court corpus), Apertus-Fin (banking regulations) are in preparation for 2026-2027. We are already training an Apertus-Fiduciary variant for a mid-market partner.
- Swiss AI governance standard: the Federal Council plans an AI ordinance for Q4 2026 that defines EU AI Act-compliant paths. Sovereign AI stacks will probably be favoured.
- Apertus on Mobile: Apertus 1B (edge variant) on Apple Foundation Models / Snapdragon X Elite — Swiss AI without a cloud round trip. Pilots in hospital mobile apps are running.
Conclusion: Sovereign AI Is a Deployable Obligation in 2026, Not a Marketing Slogan
- Default 2026: Apertus 70B on Swisscom Sovereign AI Platform. Apache-2.0 model, FINMA-certified SOC, MSA under Swiss law, multilingual with Swiss German — the most pragmatic path for 80% of Swiss mid-market engagements.
- High-risk domains: hybrid with Claude EU. Reasoning-intensive edge cases (credit review, legal research, claims assessment) via Vertex EMEA with DPA — the rest on Apertus.
- Air-gapped: only for tier-1 banks, pharma, defence. CapEx of CHF 380K-580K only pays off above 1 B tokens/month or under hard confidentiality requirements.
- No longer in 2026: OpenAI direct API without an EU DPA. FDPIC and FINMA audit risk is too high. Migration to Apertus, Swisscom, or Azure CH is unavoidable.
- Model diversification is mandatory: at least two independent model families (Apertus + Llama or Apertus + Mistral) against lock-in and FINMA risks.
- ROI in 4-7 months: 14 production mazdek sovereign AI engagements, average 5.4 months payback versus US hyperscaler setups.
- Compliance is feasible: revFADP, EU AI Act, FINMA, and Swissmedic are cleanly mapped using ARES guardrails, the WORM archive, and confidence thresholds.
At mazdek, 19 specialised AI agents orchestrate the entire sovereign AI lifecycle: PROMETHEUS for model selection and routing; HEPHAESTUS for the Swiss Kubernetes and GPU infrastructure; ORACLE for RAG, pgvector, and data provenance; HERACLES for ERP, banking, and SwissID integration via MCP; ARES for compliance, Lakera, Llama Guard, and WORM archive; ARGUS for 24/7 drift and cost observability; NABU for model and data cards and FDPIC/FINMA reporting; NINGIZZIDA for FHIR/MepV conformity in the hospital context. 14 production sovereign AI deployments since the Apertus release in September 2025 — FADP-, GDPR-, EU AI Act-, FINMA-, and Swissmedic-compliant from day one.
