Deutsche Telekom hands each AI cloud customer a separate private cloud
vCluster Labs said on 21 October, at Nvidia's GTC Berlin, that Deutsche Telekom's Industrial AI Cloud uses vCluster automation and open-source components of Nvidia's DSX OS to provision a private AI cloud for each customer. A central vCluster Platform creates a dedicated platform instance per customer, each with its own management control plane, its own Deutsche Telekom-branded portal at a unique URL, and isolation at the network, storage and compute layers. Customers provision Kubernetes, Slurm, GPU virtual machines or bare metal through the portal or an API. The Munich AI factory underneath the cloud already runs about 10,000 Blackwell parts (DCD). The first customer is an unnamed German automotive manufacturer (Business Wire, SDxCentral).
Five hundred of the cloud's Nvidia GPUs are vCluster-supported today. Deutsche Telekom expects more than 10,000 Blackwell and Rubin GPUs on the platform; Rubin is a later generation than the parts in Munich. The release says new customers come online in hours instead of days, and that workloads and sensitive industrial data stay isolated within Germany, addressing Germany's C5 cloud security criteria. Those are the vendors' claims. Andreas Falkner, Deutsche Telekom's head of Industrial AI Cloud, is quoted: "With vCluster orchestrating our NVIDIA DSX stack, each customer gets its own private AI cloud, on one platform that spans every GPU generation we run." Ask what the isolation covers at the storage layer, and who holds the C5 attestation.
The AI Act's first harmonised standard is now cited in the Official Journal
Commission Implementing Decision (EU) 2026/2274, adopted on 12 October and published in the Official Journal on 14 October, cites the first harmonised standard for high-risk AI systems drafted in support of Regulation (EU) 2024/1689. The annex lists one entry: EN 18286:2026, "Artificial intelligence – Quality management system for EU AI Act regulatory purposes". It entered into force on the day of publication and invokes Article 40, the presumption-of-conformity mechanism. The Commission's standardisation page, which still describes the document as a draft, says it is designed to help providers of high-risk systems comply with Article 17, the quality-management obligation.
A provider that applies the standard is presumed to comply with the requirements the standard covers. The same Commission page says "the application of standards remains voluntary". CEN-CENELEC said in October 2025 that its technical boards had reorganised the work to make the six most delayed drafts available by the fourth quarter of 2026. Compliance teams mapping Article 17 against a management system they already run now have a cited document to check against.
Quick Hits
- Recursive published Flint's weights under Apache 2.0. The 39.8B-total, 2.5B-active mixture-of-experts coding model went up on Hugging Face on 21 October. Recursive's own evaluations put the BF16 checkpoint at 54.9% on Terminal-Bench 2.1; no third party has run it (MarkTechPost).
- OpenAI's frontier models are going into Google Cloud's Model Garden. They will be transacted through Google Cloud Marketplace, payable from existing cloud spending commitments. Axios, reporting the 21 October agreement, puts general availability "in the coming weeks" (Axios).
- Nine operators across EMEA are building or expanding Nvidia-powered infrastructure. Nvidia said on 21 October that Elisa has completed Europe's first AI-RAN field trial, and that BT and Deutsche Telekom plan field trials in 2027, with Nokia and Nvidia (Nvidia).
- Pigment raised $250m at a $2.3bn valuation. The EU-backed Scaleup Europe Fund, managed by EQT, led the Series E in the Paris-based enterprise planning company, announced on 21 October (Silicon Republic).