Musk's xAI announced a new funding round raising $20 billion, exceeding its original $15 billion target, with NVIDIA's participation. The company stated the funds will be used to expand the Grok model and build its own AI infrastructure, with NVIDIA supporting its compute expansion via GPUs and software stacks. Some reports also mention that xAI had previously deployed ultra-large-scale data center clusters, and the funding will accelerate the iteration of its hardware and training/inference systems.

Consulting giant Accenture announced the acquisition of UK AI company Faculty for $1 billion. Faculty provides AI consulting and software for businesses and governments, and engages in services such as frontier model safety assessment; upon completion of the deal, its 400+ employees will join Accenture, with Faculty CEO Marc Warner becoming Accenture's CTO and joining the management team. This acquisition points to consulting firms expanding their delivery capabilities to implement 'model + data + governance' into clients' production systems.

At CES 2026, AMD emphasized that AI compute power is moving towards Yottaflops scale, with demand shifting from centralized training to long-term, high-frequency inference. The company released the MI455X data center GPU and the Helios platform for rack-scale deployment, and previewed the MI500 series launching in 2027 using 2nm process; it also set a goal of a 1000x increase in AI performance over the next four years, aiming to compete for large-scale inference infrastructure orders with open racks and software/hardware stacks.

Liquid AI released and open-sourced weights for the LFM2.5 series, targeting device and edge inference. The core model has 1.2B parameters, with pre-training data expanded from 10T to 28T tokens, and includes Instruct versions and variants optimized for Japanese, vision-language (1.6B), and native audio-language (1.5B). Official introduction states it outperforms some similar open-source models on benchmarks like GPQA, MMLU Pro, and IFEval/IFBench, and can be deployed on various inference stacks and hardware.

Model evaluation platform LMArena completed a $150 million Series A funding round, reaching a post-money valuation of $1.7 billion; it has raised approximately $250 million in total over about seven months. The platform is famous for its 'Chatbot Arena' blind test voting, and disclosed over 5 million monthly users and about 60 million monthly conversations, covering 150 countries. Reports indicate its commercial evaluation service has launched, generating roughly $30 million in annualized recurring revenue, and the funds will be used to expand the technical team and research capabilities.

Databricks proposed the Instructed Retriever architecture, which integrates system specifications (user instructions, examples, index descriptions, etc.) throughout the entire retrieval and generation pipeline to enhance the instruction-following and system-level reasoning of search agents. It achieved 35–50% higher recall on the StaRK-Instruct benchmark and reported up to about 70% better answer quality on enterprise QA datasets; it also demonstrated that an optimized 4B model performs close to larger closed-source systems on this task, emphasizing a cost-controllable path for enterprise deployment.

Zhejiang Lab and BGI jointly released the Genos gene foundation model, with Genos-10B designed for whole-genome ultra-long sequence modeling. It utilizes HMoE and Fractal Attention to achieve million-base-level context and single-base precision prediction, covering tasks like functional element identification, expression simulation, and clinical diagnosis. The team also provides 1.2B and 10B versions to adapt to different compute conditions, and mentioned deployment optimization for domestic hardware and RESTful API services, aiming to lower the barrier to using AI in life sciences.

The EU AI Office released a draft 'Code of Practice on Transparency of AI-generated Content,' providing an actionable framework for Article 50 transparency obligations under the AI Act. The draft distinguishes between 'provider/deployer' responsibilities: providers need to implement multi-layer labeling and detection technologies (including metadata and digital watermarks), and provide free detection interfaces; deployers must disclose and uniformly label deepfakes and AI-generated/manipulated text related to public interest, with exemptions set for artistic creation and content reviewed after human editing. The final version is expected in June 2026, taking effect for signatories from August 2, 2026.

Researchers disclosed a high-risk vulnerability CVE-2025-64496 in Open WebUI's 'direct connection' feature: when a user connects to a model server with a compatible OpenAI interface controlled by an attacker, malicious JavaScript can be injected and executed via insecure SSE events ({type: execute}), stealing JWTs from browser localStorage to take over accounts, access chats, and API keys. If an account has workspace.tools permissions, attackers can also use the Tools API to execute unsandboxed Python code, escalating the intrusion to backend RCE. The vulnerability affects v0.6.34 and earlier versions, fixed in v0.6.35 by blocking such events; exploitation requires users to manually enable the feature and add a malicious URL.

Netskope's Cloud and Threat Report stated that over the past year, the number of users accessing generative AI apps via the cloud tripled, with average monthly prompts rising from 3,000 to 18,000. Although the proportion of users via enterprise-managed accounts increased to 62%, 47% still use 'shadow AI' via personal accounts, creating visibility and data exfiltration risks. The report shows data policy violations related to generative AI doubled year-over-year, averaging 223 incidents per organization per month, mainly involving source code (42%), regulated data (32%), and intellectual property (16%), reflecting gaps in enterprise governance and DLP policies.