Two open-weight releases in three days, a European aerospace prime moving critical workloads off a US hyperscaler, and an AI Act deadline now two weeks out. The common thread for European teams is that capability and jurisdiction are becoming separate purchasing decisions, and the second one is getting harder to defer.
Moonshot announces Kimi K3 at 2.8T parameters, weights due 27 July
Moonshot AI introduced Kimi K3 on 16 July: 2.8 trillion total parameters in a mixture-of-experts layout across 896 experts, a one-million-token context window, and native image and video understanding. Only a fraction of the expert pool activates per token, so the serving cost is far below what the headline parameter count implies. Weights are not out yet — Moonshot has committed to publishing them on 27 July.
Pricing is the aggressive part: roughly $3 per million input tokens and $15 output, flat across the full context with no long-context surcharge. The caveat is equally clear. K3's hallucination rate rose from about 39% to 51% against its predecessor on the benchmark designed to measure it, and Moonshot's own documentation warns the model can act on a user's behalf during long autonomous runs without being asked.
For a European team, nothing is deployable until the 27th, and a 2.8T model is not a single-node proposition even then — self-hosting it on EU infrastructure means real cluster capacity, and the interim alternative is an API in a non-EU jurisdiction. The rising hallucination rate also puts it firmly outside unsupervised regulated decision paths regardless of where it runs.
Inkling gives US labs an open-weights answer that fits in fewer racks
Thinking Machines released Inkling on 15 July: 975B total parameters with 41B active, accepting text, image and audio input, with weights on Hugging Face at launch supporting a one-million-token context. Artificial Analysis puts it at 41 on its Intelligence Index, ahead of Nemotron 3 Ultra at 38 and the leading US open-weights release to date.
The number worth noting is token efficiency: Inkling averages 25K output tokens per Intelligence Index task, against 43K, 38K and 37K for GLM-5.2, Kimi K2.6 and DeepSeek v4 Pro. On agentic evaluations it scores above both Kimi K2.6 and DeepSeek v4 Flash. Accuracy on AA-Omniscience is 40% with a 63% hallucination rate — competitive for the class, not a substitute for retrieval and review.
41B active parameters is tractable on a modest H100 cluster, and the weights exist today. That makes Inkling one of the few frontier-adjacent models a European operator can run entirely inside its own jurisdiction this week, which removes the extraterritorial-access question rather than contracting around it.
Airbus moves critical workloads to Scaleway, citing extraterritorial law
Airbus has selected Scaleway as its sovereign cloud provider, with dozens of critical applications — aircraft design, engineering, manufacturing and enterprise operations — moving over the next two years. The award followed a competitive tender scored on technical capability, interoperability and AI services, alongside legal criteria: European jurisdiction, data protection, and explicit protection against non-European extraterritorial legislation.
This is the procurement pattern to watch, because it mirrors the Commission's own. The EUR 180 million sovereign cloud framework awarded in April scored providers against Sovereignty Effectiveness Assurance Levels, requiring SEAL-2 as a floor; Scaleway, STACKIT and the Post Telecom/OVHcloud/CleverCloud consortium reached SEAL-3, meaning immunity from supply-chain disruption by non-EU third parties. Sovereignty has stopped being a positioning statement and become a scored line item.
Article 50 transparency obligations apply from 2 August
From 2 August 2026, providers must inform people when they are interacting with an AI system and attach machine-readable marks enabling detection of AI-generated or manipulated content. Deployers carry their own duty: disclosure for deep fakes, AI-generated publications on matters of public interest, and emotion recognition or biometric categorisation. The Commission's draft guidelines on these obligations have been through consultation, complemented by a voluntary code of practice on marking and labelling.
The contrast with the high-risk regime is the planning point. Following the AI Omnibus agreement, high-risk obligations now apply from 2 December 2027 for biometrics, critical infrastructure, education, employment and border control, and from 2 August 2028 for AI embedded in products. Conformity assessments have room to breathe. Disclosure and content marking do not — that work is due in a fortnight, and it lands on deployers who may not consider themselves AI providers at all.
Two open-weight frontier models in one week, and the harder question for European buyers remains where the inference runs.
Correction (19 July): an earlier version dated the Kimi K3 introduction to 17 July; Moonshot began rolling it out on 16 July.