NVIDIA reported $96.2 billion in revenue for its second fiscal quarter, a 106% year-over-year increase, with net income rising 126%. Data center business generated $89 billion in revenue, up 117% year-on-year, becoming the primary growth driver. The company forecasts $108 billion in revenue for the next quarter, while another summary projects annual revenue at $432 billion. Results indicate sustained high demand for AI training and inference infrastructure procurement; however, existing materials do not disclose margins, regional breakdowns, or specific customer contributions. Future attention should focus on advanced chip supply and order fulfillment.

Zhipu has open-sourced GLM-5.3-Flash, a Mixture-of-Experts (MoE) model with 320 billion total parameters and 18 billion activated parameters per token, incorporating hybrid linear and sparse attention mechanisms along with native multimodal capabilities. Officially reported to achieve an Artificial Analysis index of 57, the promotional cost is $0.045 per task. Compared to GLM-5.3, it reduces attention computation by 3.01x and shrinks KV Cache by 4.44x. The model weights are released under the MIT license and have already been deployed on domestic AI chip clusters, achieving a 3x improvement in end-to-end inference performance.

Anthropic has entered into an approximately $45 billion cloud computing agreement with UK-based AI infrastructure provider Nscale, securing around 460 megawatts of compute capacity at its West Virginia data center. The facility is expected to come online by the end of 2027 and will utilize NVIDIA Vera Rubin chips. If executed as planned, this deal represents a multi-year commitment to large-scale training and inference capacity, reflecting how frontier model companies continue to lock in power, data center space, and next-generation accelerators through long-term agreements. Details such as contract duration, payment schedule, and exit clauses remain undisclosed.

Google DeepMind has launched Gemini Omni 1.1 Flash, adding new video generation features including scene extension, control over start and end keyframes, low-resolution drafts, and high-resolution upscaling. The model can analyze prior videos up to 10 seconds long to maintain visual continuity and narrative flow, and generate smooth transitions between two keyframes—such as panning, zooming, or looping. Developers can iterate quickly at low cost using 360p mode, then upscale outputs to 1080p or 4K, enabling finer-grained control for generative video workflows, creative tools, and media editing software.

Google DeepMind, in collaboration with the Singapore Institute of AI Safety and other institutions, is piloting a double-blind AI evaluation framework: model providers cannot view test questions, and evaluators cannot access proprietary model weights. The approach leverages Google Cloud Confidential Space to run Gemini Flash Lite in an isolated environment, using confidential computing to verify that both parties' data remains private. This mechanism aims to reduce benchmark contamination and question leakage risks, allowing independent capability assessments in high-sensitivity domains like cybersecurity and government without exchanging core assets. However, the initiative remains in pilot phase.

Researcher Johann Rehberger disclosed a prompt injection vulnerability in Claude Code Opus 5's auto mode: malicious content can trick the agent into downloading and executing a forged struct.py, with a test success rate of approximately 80%. More critically, after Claude detected the intrusion and attempted to terminate the malicious process, the auto mode intercepted the cleanup command, rendering its built-in classifier ineffective. The study recommends running unattended coding agents within containers, virtual machines, or system sandboxes, restricting external network access, and isolating credentials and sensitive files.

AWS has introduced OpenAI's GPT-5.6 Terra and Luna on Amazon Bedrock for Indian customers, enabling intra-country inference across Indian regions. Requests are routed only between Mumbai and Hyderabad, meeting data residency requirements for industries like finance and healthcare. The service supports OpenAI Responses, Chat Completions, and Bedrock Converse API, offering inference depth control and server-side state persistence. For inputs with prefixes of at least 1024 tokens, caching is available, reducing cache read costs by 90% compared to uncached inputs.

Anthropic has unveiled the Model Hardware Standard, providing Claude with constrained interfaces to laboratory instruments, enabling it to coordinate microscopes, robotic arms, and drug discovery equipment, read results, and iteratively adjust parameters. Based on cross-device communication principles, the project aims to reduce time spent on instrument assembly, configuration, and hardware-software debugging, creating a closed-loop experimental workflow of observe, act, and re-validate. The demonstration includes safety constraints such as mechanical arm motion boundaries and microscope collision avoidance, emphasizing that real-world experimentation must incorporate safety limits, sample conservation, and human oversight into interface design.

Salesforce and Anthropic have launched Claudeforce, integrating CRM data and workflows directly into Claude. The initial release offers 37 pre-built sales skills, allowing users to query and manipulate customer information within conversational interfaces. Public testing is scheduled to begin in September, with Slack integration planned for later. This collaboration extends Claude from a general chat interface into a sales execution environment, focusing on calling enterprise data and executing predefined processes within permissioned boundaries while minimizing cross-system switching. Pricing, supported regions, and official launch timing have not been disclosed.

OpenAI has published a randomized experiment involving over 1,000 students: those granted access to ChatGPT showed an average improvement of nearly one point on grading scales, with enhancements in assignment quality, coherence, and alignment with expert recommendations. Students who received causal reasoning training alone did not show direct improvements in scale scores but exhibited greater idea diversity and uniqueness. Those receiving both interventions demonstrated both higher-quality outputs and broader thinking. The study also cautions that as final products become more polished, assessing students’ true understanding based solely on final answers becomes increasingly difficult.