A German cloud provider cuts 800 jobs and names AI as a reason
United Internet said on 11 September it will cut around 800 jobs across two subsidiaries. Ionos Group goes from about 3,800 full-time employees to about 3,350, and 1&1's business-customer unit 1&1 Versatel from about 1,350 to about 1,000. The one-off charges are €35m at Ionos and €60m at 1&1, and 1&1 expects an annual earnings contribution of around €25m from fiscal 2028. On the Ionos cuts, DataCenterDynamics reports that “United stated that the implementation of AI within its processes is a factor for these cuts.” The reductions run through voluntary programmes and hit both domestic and international operations.
For a European buyer this is a capacity signal, not a headline about job losses. Ionos sits on a lot of EU procurement shortlists as the non-US option, and it is shrinking headcount while saying the savings go back into AI product development and cloud expansion. That is a bet, and the part a buyer should ask about is support and delivery coverage on the international side during the transition — the same question you would put to any supplier restructuring while promising to expand. It also tells you something about how thin the margins are under Europe's sovereign-cloud alternatives.
The Cyber Resilience Act's reporting clock started on Friday
As of 11 September, manufacturers of products with digital elements must report actively exploited vulnerabilities and severe incidents. The Commission's page sets out the structure: an early warning within 24 hours of becoming aware, a full notification within 72 hours, and a final report no later than 14 days after a corrective measure is available for an actively exploited vulnerability, or within a month of the 72-hour notification for a severe incident. Reports go through ENISA's Single Reporting Platform, established under Article 16 and operational the same day. One submission reaches the CSIRT where the manufacturer has its main establishment and, unless particularly exceptional circumstances apply, ENISA at the same time.
Two details matter for planning. The receiving CSIRT shares the notification without delay with every other CSIRT where the product is available, and the only brake is a delegated act adopted on 11 December 2025 setting out the cybersecurity grounds for delaying that dissemination. And open-source software stewards are not in scope until 11 December 2027, under Article 71(2). If you ship software into the EU, the 24-hour clock is now the binding constraint on your incident process — not the disclosure timeline you agreed with a researcher.
A vendor publishes open weights, then says the benchmarks are not for these weights
Agnes AI published Agnes-3.0-Flash on Hugging Face on 11 September: 33B parameters, a 262,144-token context window, multimodal, Apache-2.0. The next day it amended the card, adding “Preview” to the name and a clarification section. In its own words: “The production/API model uses a different checkpoint and configuration, with a 1M-token context window. Its benchmark results should not be attributed to the Preview weights released here.” The amendment also added a note above the benchmark table: the results there “belong to the Agnes-3.0-Flash Preview open-weight checkpoint released in this repository. They are not results for the production/API Agnes 3.0 Flash model listed on Artificial Analysis.” The card had already warned that its reference figures for other models “were compiled from different sources, harnesses, and model snapshots and do not constitute a controlled head-to-head comparison.”
The amendment is the story, and it is a good one. A third-party leaderboard row for “Agnes 3.0 Flash” now describes a model nobody can download, while the downloadable model carries the same name minus a suffix added after the fact. Anyone evaluating open weights by looking up the name on a benchmark site would have got the wrong artefact, and the only thing that prevented it was the vendor correcting itself in public. The general case is worth writing into a procurement checklist: the thing you can host and the thing that was measured have to be the same checkpoint, and the evidence for that is a model card, not a leaderboard.
Quick Hits
- A 397B open-weight science model lands under Apache-2.0. Shanghai AI Lab published Intern-S2-397B on 13 September — a mixture-of-experts model with 512 experts, 10 active per token, 60 layers and a 262,144-token maximum position. Read the card's evaluation note before comparing anything: text reasoning was measured at 256K inference length and multimodal at 64K, so the chart mixes two regimes.
- OpenAI gives a cyber model 20 days to live. The deprecations page added an entry on 11 September: gpt-5.4-cyber leaves the API on 1 October, replaced by gpt-5.6-cyber. The same page promises “At least 3 months” for specialized variants of generally available models — with an explicit carve-out when “safety or compliance concerns require a faster timeline”, which a cyber-capability model plausibly triggers. Either way, if you pinned that slug, you have about two and a half weeks.
- A PII masking model you can actually deploy. Perplexity published pplx-pii-masking-vllm on 13 September under MIT: a ~600M-parameter encoder with token classification over nine PII categories plus a conversation-level sensitivity score, 4,096-token sequences, shipped with a Docker Compose deployment pinned to vLLM 0.26.0. It is a repackaging of a July release rather than a new model, which is the point — it is now something an infrastructure team can run in front of a gateway.
- An open model with no licence at all is not an open model. Knowledgator's GLiFormer Large v1, published 11 September, is a 575.6M-parameter encoder for named-entity recognition, classification, relation extraction and structured records, with schemas supplied at inference time. It ships with no LICENSE file and no licence tag. Until terms are stated, the legal position is that you have no permission, which is worse than a restrictive licence because it cannot be assessed.
- A 120MW Italian data centre application was withdrawn over grid data. DataCenterDynamics reported on 11 September that East Gate pulled its building application in Travagliato, Brescia, ten months after filing. The mayor says the municipality never received the data it needed for a full environmental impact assessment or for the 120MW grid connection. EU-region capacity is increasingly a permitting and grid question rather than a construction one.