Databricks announced the completion of a $4 billion funding round, achieving a post-money valuation of $134 billion. The company stated that the funds will be used to help enterprises build AI applications and agent systems using their proprietary data, while enhancing production-grade deployment, scalability, and operational capabilities. The report reflects a shift in enterprise AI investment from 'buying models' to 'building infrastructure,' where data platforms must support data governance, access control, and agent-oriented application development to enable longer-chain automated workflows.

According to The Information, Nvidia may unusually skip launching a new generation of gaming GPUs in 2026 due to ongoing memory supply constraints. The company is prioritizing limited memory resources for AI accelerators and data center product lines. Citing reports, its GAAP gross margin was approximately 75% in fiscal year 2025, with data center revenue accounting for about 90%. HBM capacity from SK Hynix and Samsung has been pre-secured, and shortages may persist beyond 2026, potentially affecting consumer GPU upgrade cycles and pricing.

The National Data Bureau, Ministry of Industry and Information Technology, Ministry of Public Security, and China Securities Regulatory Commission jointly released the 'Opinions on Cultivating Data Circulation Service Providers to Accelerate Data Element Marketization and Value Realization.' The document promotes the development of data exchanges, data circulation service platforms, and data vendors, emphasizing clear functional positioning, enhanced service effectiveness, and improved regulation. It notes existing challenges such as unclear roles, insufficient service capacity, and inadequate oversight, and sets goals for significant capability improvements by end-2029, diversified circulation models, richer data products and services, and higher societal levels of data circulation and utilization to support the 'AI+' initiative.

The State Administration for Market Regulation (SAMR) disclosed five typical cases of unfair competition in the AI sector, covering imitation, false advertising, and trade secret infringement. The通报 shows: unauthorized use of the 'DeepSeek' name and logo was fined 5,000 RMB; a fake 'ChatGPT Chinese Version' official account was fined 62,692.7 RMB; assisting loan intermediaries with AI-powered outbound calls for false promotion was fined 200,000 RMB; an engineer who downloaded 15.88GB of files containing algorithmic trade secrets was fined 360,000 RMB; and setting up a fake DeepSeek website to induce payments was fined 30,000 RMB. Regulators emphasized the need to standardize AI applications and protect innovation outcomes and consumer rights.

Researchers from Columbia University and Microsoft Research proposed Reinforced World Model Learning (RWML), enabling agents to self-supervisedly explore environments and learn intrinsic world models of 'action-consequence' relationships for multi-step planning and tool use, mitigating causal gaps from static text-only training. Reported results show: ALFWorld task success rate increased from 13.0% to 32.6%, reaching 87.9% when combined with task rewards; in the τ?Bench customer service scenario, tool usage error rate dropped from 24.9% to 8.84%. The paper has been published on arXiv (2602.05842v1).

Xiaomi's MiMo team proposed HySparse, a hybrid sparse attention architecture: a small number of Full Attention layers provide real attention distributions to guide subsequent Sparse layers in reusing token selection and KV Cache, enabling cross-layer cache sharing. The design reportedly reduces KV Cache overhead by about 80%, alleviating memory and bandwidth bottlenecks in long-context inference. In an experiment with a 49-layer, 80B-scale MoE model, retaining only 5 Full Attention layers maintained or improved performance across math, code, and long-text tasks, offering an engineering approach for cost control in ultra-long-context models.

DeepMind employed the AlphaEvolve system to automatically search for activation functions: a large model generates and rewrites function code within an 'infinite Python code space,' followed by screening candidates using small-scale synthetic tasks and out-of-distribution (OOD) tests for better generalization. The report noted that functions incorporating periodic perturbations, such as GELUSine (a variant of GELU), were found to enhance extrapolation ability. It also revealed that 'turbulent-type' functions relying on batch statistics may fail in real-world tasks. This approach advances traditional NAS from fixed operator combinations to programmable search paradigms.

StrongDM's AI team unveiled a 'Software Factory' model: humans neither write code nor perform code reviews, instead defining specifications, scenario sets, and monitoring systems to constrain and validate agent outputs. In implementation, the team built digital twins of third-party SaaS platforms (e.g., Okta, Jira) for high-frequency end-to-end testing and treated user stories as independent holdout scenarios to reduce the risk of agents generating 'fake tests.' This practice aims to replace manual review with scalable validation systems to improve iteration efficiency and consistency in secure software development.