What’s happening in the market and with providers. News for daily operations. For professional users. And for strategic direction.
Regulation
EU AI Act: high-risk obligations from 2 August 2026 – despite the ‘Digital Omnibus’ debate
From 2 August 2026 most remaining EU AI Act obligations apply, including enforcement powers and national sandboxes; a simplification via the ‘Digital Omnibus’ is under discussion but not yet law. If you serve EU customers, treat August 2026 as the operative deadline.
On 2 August 2026 the bulk of the AI Act starts to apply (except Article 6(1)); authorities can enforce from that date, and GPAI obligations have applied since August 2025 anyway. Member States must also operate at least one national AI sandbox. Fines reach up to EUR 15m or 3% of global annual turnover.
The ‘Digital Omnibus’ would tie parts of the high-risk obligations to the availability of standards and stretch deadlines to 2 December 2027 and 2 August 2028 (product-embedded systems). That delay is not yet enacted, however – relying on it would be risky.
For Swiss providers the place-of-market principle applies: if you place AI systems on the EU market or let them have effect there, you fall in scope regardless of where you are headquartered. In practice: set up risk classification, technical documentation, logging and human oversight now, rather than betting on a possible extension.
Data residency becomes a business criterion: 51% of Swiss firms require CH/EU-compliant AI
An EY survey shows AI is established in Swiss companies, yet 51% require AI systems to comply with Swiss or European data protection and keep data within Switzerland or the EU.
EY reports that AI is already widely embedded in Swiss companies – though many are still early in scaling. The clearest signal for mid-sized business: data residency and privacy compliance are treated as business-critical.
For fiduciary, advisory and law-firm contexts this means: anyone advising on contracts, outsourcing, AI procurement or client confidentiality should tighten internal AI governance and vendor review now. This is exactly where a sovereign, Switzerland/EU-hosted AI architecture comes in.
Mistral Medium 3.5: an open 128B model you can self-host on four GPUs
Mistral released Medium 3.5, a dense 128B model (256k context) as open weights under a modified MIT licence; the vendor says it runs on as few as four GPUs. That puts capable on-prem AI within realistic hardware reach for SMEs.
Medium 3.5 merges instruction-following, reasoning and coding into a single set of weights, with reasoning effort configurable per request. Mistral cites 77.6% on SWE-Bench Verified. The weights are on Hugging Face; the hosted API costs USD 1.5 / 7.5 per million input/output tokens.
The practical point for CH/EU SMEs is the licence and the hardware threshold. A modified MIT licence permits commercial use and self-hosting, and ‘four GPUs’ means a single server rather than a cluster. That makes it feasible to process confidential data (case files, HR records, contracts) locally instead of sending it to an external API.
We’d advise getting the licence wording (‘modified’) reviewed before production use, and reading ‘four GPUs’ as an optimistic floor – memory needs climb with quantisation choices and longer context. Even so, Medium 3.5 is one of the most serious open candidates for audit-ready on-prem operation this spring.
FDPIC: Swiss data-protection law applies directly to AI – no special rules needed
The FDPIC reaffirms that the revised FADP applies directly to AI-driven data processing: transparency, disclosure of automated decisions and DPIAs for high-risk use apply today. Swiss SMEs need not wait for AI-specific legislation – the duties already exist.
The FDPIC is explicit: manufacturers, providers and users must ensure transparency about purpose, functionality and data sources, users must be able to tell when they are talking to a machine, and tools for face, image or voice falsification must be disclosed. A data-protection impact assessment is mandatory for high-risk processing. In January 2026 the FDPIC also joined 60+ authorities in a joint statement against AI-generated images of real people without consent.
For the Mittelstand the message is clear: the revised FADP, in force since September 2023, is enough to govern AI use today. Waiting for a future Swiss AI law (consultation expected toward the end of 2026) misreads the current state of the law.
In concrete terms for fiduciaries and law firms: document AI processing of personal data, flag automated decisions, grant a right to human review, and run a DPIA for sensitive mandates. Running AI on-prem or in a Swiss/EU cloud makes these records substantially easier to produce than relying on a US API.
Prompt injection remains the most-exploited AI vulnerability in 2026 – with a bigger blast radius for agents
Security analyses for 2026 name prompt injection as the most-exploited AI vulnerability; for agentic systems (sending email, querying databases, executing code) the blast radius grows sharply. IBM reports that 97% of organisations hit by AI incidents lacked proper AI access controls.
Documented cases illustrate the risk: CVE-2025-53773 enabled remote code execution with GitHub Copilot via hidden prompt injection in pull-request descriptions (CVSS 9.6), and the ‘EchoLeak’ flaw in Microsoft 365 Copilot demonstrated zero-click data exfiltration. With RAG pipelines and agents, the threat shifts from ‘embarrassing’ to ‘catastrophic’.
For CH/EU SMEs the key lesson is that the model is rarely the problem – missing access controls, unchecked tool permissions and ‘shadow AI’ are. A self-hosted model without least-privilege, input filtering and audit logging is just as exposed as a cloud API.
Concrete steps: scope agent tool access strictly by least privilege, treat untrusted content (emails, documents, web pages) as untrusted and never feed it unfiltered into system prompts, segment RAG sources, and log every agent action. These same controls are the evidence the EU AI Act (logging, oversight) and the revised FADP (data security) require anyway.
Cohere acquires Aleph Alpha – European sovereign AI built on STACKIT cloud
Canadian AI vendor Cohere is merging with Heidelberg-based Aleph Alpha into a roughly USD 20bn entity, backed by Germany’s Schwarz Group (about EUR 500m) and running on its sovereign STACKIT cloud. For CH/EU SMEs this creates a credible alternative to US hyperscalers for regulated workloads.
The combined company explicitly targets heavily regulated sectors – defence, energy, finance, healthcare, telecoms. It runs on STACKIT, Schwarz Digits’ sovereign cloud, keeping data and operations inside Europe. Aleph Alpha’s strengths in small models, European languages and tokenizers complement Cohere’s enterprise stack.
What matters for the Mittelstand is that capital, model expertise and sovereign infrastructure are being bundled for the first time at this scale. That lowers the barrier to running AI without data leaving for the US – a recurring concern for fiduciaries, law firms and public bodies.
A word of caution on the framing: a Canadian-German entity is not automatically ‘sovereign’ in the strictest sense. If data residency and processing agreements genuinely matter to you, scrutinise the contract terms on location, sub-processors and access rights rather than trusting the label.
DeepSeek V4 ships open weights – strong benchmarks, but NIST measures a roughly 8-month gap
DeepSeek released V4 (1.6T-parameter MoE, 1M-token context) with open weights; a NIST/CAISI evaluation in May 2026 finds it lags the US frontier by about eight months. For SMEs that means cheap and open, but not always frontier-equivalent.
V4 comes in two variants (Pro: 1.6T / 49B active; Flash: 284B / 13B active), both with a 1M-token context and a novel sparse-attention design. Weights are on Hugging Face (V4-Pro under MIT). On coding and maths benchmarks V4 is strong.
The NIST/CAISI evaluation (1 May 2026) provides the sober counterweight: V4-Pro scored an IRT Elo of 800±28 versus 1260±28 for GPT-5.5, and just 46% on ARC-AGI-2 versus 79%. Its main edge is cost efficiency – cheaper than GPT-5.4 mini on five of seven benchmarks.
For CH/EU users the framing matters more than the headline. For RAG, document analysis and routine automation, a self-hosted open V4 can be sufficient and cut costs dramatically. For demanding reasoning, plan for the gap. Chinese-origin models also carry their own governance questions (training-data provenance, data compliance) – self-hosting air-gapped mitigates this technically but does not replace due diligence.