The proposal sets a goal of at least tripling the EU’s data centre capacity within the next 5-7 years. It would accelerate permitting and improve access to energy, land, water and financing. Research support would focus on frontier AI, industrial AI and physical AI, with adoption encouraged through national strategies and Experience and Acceleration Centres for AI.

A common EU-level procurement framework would help public administrations combine purchasing power. The initiative also promotes open source solutions and contributions to EU-based innovation and supply chain resilience. It seeks faster deployment in critical sectors while keeping the vast majority of the market open to international partners.

The proposed sovereignty framework comprises four assurance levels for public bodies to use according to risk assessments, with providers recognised by Member States following an audit. Level 1 requires EU-based data processing and storage. Level 2 adds independence from third countries and software supply chain transparency. Level 3 requires EU ownership and control, with additional criteria and scope for Commission recognition of third-country providers. Level 4 requires full software supply chain transparency and control, without third-country interference.

The proposed EU KIDS Act would introduce harmonised age restrictions, privacy-preserving age assurance and safety-by-design requirements. Its scope includes social platforms, AI companions, conversational chatbots, games and app stores, with safeguards against emotional dependency and mandatory child-safety testing. Adoption is not expected before 2028. Consumer authorities are separately challenging in-game currency practices that obscure prices or undermine consumer rights.

Cyber Resilience Act vulnerability and incident reporting became mandatory on 11 September 2026, including for legacy products. Connected-product data accessibility requirements followed on 12 September 2026. EU and UK financial supervisors are expanding third-party oversight while warning that AI-driven vulnerability discovery is outpacing remediation. Firms are being urged to test patching, monitoring and supplier readiness.

The UK has brought forward its subscription contracts regime to January 2027, requiring clearer information, renewal reminders and straightforward cancellation. Germany’s proposed Digital Media State Treaty would make AI information services responsible for generated content and impose additional transparency and journalistic duties on larger services. The proposal could also cover providers established outside Germany when their services target German users.

Longer context windows and open weights are expanding development options. Kimi K3 has published weights and a 1M-token context window, while the GPT-6 family supports a 1.05M-token context. DeepSeek’s V4.1 Flash targets low-cost coding and vision work. Media tools are also broadening: FLUX 3 handles image, video and audio within a single model.

Model selection should follow workload requirements rather than release headlines or a single benchmark. Balanced models provide a practical starting point for everyday coding and agents, with smaller models suited to high-volume tasks and flagships reserved for difficult reasoning. Testing on 20 to 50 real tasks can help compare correctness, speed, tool-call reliability and cost. Developers should also verify account access, plan for model retirements and remember that the same model can behave differently across products.

Data center revenue rose 97% in Q3 and accounted for 64% of AMD’s total revenue, compared with 47% a year ago. Su previously used the same description in 2024, when data centers represented just over half of revenue.

Gaming revenue dropped 40% and represented under 6% of total revenue. AMD attributed the decline mainly to lower semi-custom sales, including console chips for Sony and Microsoft, alongside weaker graphics demand. Gaming revenue was flat against Q2, while the previous year’s Q3 had benefited from a 181% console-driven increase.

At King’s College London, researchers used compact descriptions of objects and their relationships instead of periodically transmitting full images to guide a robot arm. In fruit-packing and cube-stacking experiments with deliberately introduced faults, the approach reduced task completion time by up to 52.6% and improved task success. The system also recovered from unsuccessful actions and, with a digital twin, checked planned movements for potential collisions.

SoftBank integrated a neural receiver into NVIDIA Aerial CUDA-Accelerated RAN and tested it with a commercial 5G smartphone. Under controlled interference, it delivered higher uplink throughput across all 15 tested settings, with a maximum gain of roughly 60%. Separately, ETH Zurich researchers found that adapting neural receivers using local measurements reduced block error rates compared with selected pretrained versions.

University of Leeds researchers forecast network demand 10 minutes ahead to guide base station sleep decisions, reporting a 45.6% reduction in RAN power consumption while maintaining connectivity. CTTC is also integrating GPU-accelerated radio channel models into its open source network simulator, shortening simulation runtime while preserving expected statistical behavior and network performance.

Home visits demonstrate both practical cleaning abilities and significant limitations. During an initial appointment, a robot struggled with stairs and stopped cleaning after heat disabled its operator’s monitoring equipment. Tau refunded the visit and offered a complimentary return. The subsequent cleaning went more smoothly, with the robot cleaning kitchen surfaces and a toilet, although it left taps dripping, knocked objects over and worked slowly.

The service remains in beta, with connectivity, coordination and reliability still presenting challenges. Cameras inside customers’ homes also raise privacy concerns, while inexpensive autonomous cleaning could eventually threaten livelihoods. Weinzierl points to cleaning’s injury risks and staffing difficulties as reasons to pursue automation. Tau is also considering future applications in mold remediation and home health care, but fully autonomous household cleaning remains a distant goal.

Samsung is building a vast semiconductor manufacturing campus in Pyeongtaek, South Korea, to meet soaring demand for the specialized memory chips that generative AI models require. At the complex south of Seoul, three fabs already operate around the clock, a fourth is coming online, and thousands of workers are racing to build a fifth. The expansion highlights the manufacturing infrastructure needed to support the AI boom.

Novagentica Architect™ supports centralized engineering, deployment, operation and governance while allowing agents to run on different platforms and models. A certified “Decision Contract” defines value limits, approval thresholds, escalation paths and prohibited decisions. After an action executes, its result is independently checked against data in the system of record. The resulting evidence chain is designed to support documentation requirements under the EU AI Act.

The partners will initially focus on controlled write access to systems of record, compliance-by-design aligned with the EU AI Act, and industrial and logistics applications. They are already implementing controlled agent write access in an initial joint customer project, aiming to establish a standardized, scalable approach for aviation, logistics, mobility, manufacturing and retail.

Greater Manchester faces exposure through pensions, investments and development plans. A GMCA-commissioned study identified 663 businesses developing or using AI and estimated the wider regional AI workforce at almost 11,000. A prolonged investment downturn could delay planned data centres, including projects in Stockport, and undermine anticipated jobs and business-rate revenue. Start-ups that compete mainly by replacing workers with AI could also weaken local employment and shift spending from wages to external technology suppliers.

Councils, businesses and residents could face higher charges or tighter usage limits if providers become less willing to absorb losses. Regional planning should account for reduced investment and potentially higher service costs. Public bodies should assess the impact if AI prices doubled, retain skills needed to operate essential services without these tools and secure workable supplier-switching or contract-exit terms. Start-up support should weigh service improvements, sustainable employment and locally retained income alongside growth.

The framework organizes assessments around purpose, risk, control, oversight, resilience and assurance. It highlights testing for hallucinations, adversarial inputs and changing model behavior, alongside safeguards such as restricted permissions, circuit breakers and fallback procedures. Human oversight must be substantive: reviewers need the expertise, authority and capacity to reject, modify or halt actions. Control intensity should reflect a system’s autonomy, potential consequences and reversibility.

Customer-facing applications receive particular attention because inaccurate statements about fees, interest rates or account status can cause harm and create legal liability. Banks remain responsible for compliance when AI generates an improper statement or action. Vendor oversight should address access to customer data, visibility into agents’ activities, responsibility for errors and notice of material changes. An illustrative risk-tiering methodology is not mandatory, and banks are not expected to use the framework to assess their programs. The OCC plans regular updates and invites stakeholder feedback.