SpaceX has announced a $60 billion all-stock acquisition of Anysphere, the parent company of AI code editor Cursor. This marks SpaceX's first major acquisition post-IPO, rapidly entering the AI programming tools market. The Cursor team stated it will collaborate with SpaceX to advance useful AI and promised significant improvements to the editor soon, with CEO Michael Truell expressing enthusiasm for the partnership. Meanwhile, Cursor also announced the upcoming launch of Origin, a code storage and Git hosting service for teams and AI agents, set for release this fall. This transaction signifies SpaceX's expanded AI ambitions within the Musk ecosystem, integrating AI programming into its core technology stack.

OpenAI has released new research on 「Deployment Simulation」, a method that replays de-identified historical user conversations to new models in order to predict their real-world safety behavior before official deployment. In studies involving GPT-5.4, the method achieved 92% accuracy in predicting the direction of safety metric changes for categories where productivity varied by more than 1.5x, significantly outperforming the 54% accuracy of high-difficulty prompt benchmarks. The approach also supports evaluation of agent-like scenarios dependent on file systems and APIs by using auxiliary models to simulate external states. OpenAI emphasizes this serves as a complementary signal to traditional evaluations and red-teaming, reducing model 「evaluation awareness」 to near-production levels.

NVIDIA achieved a clean sweep in MLPerf Training 6.0, being the only platform to submit results across all tests—including newly added workloads for DeepSeek-V3 and GPT-OSS MoE—and setting new records for large-scale training times. Key technologies include full-iteration CUDA graphs eliminating CPU-GPU synchronization bottlenecks by offloading entire training iterations to GPU; and performance boosts exceeding 8% on DeepSeek-V3 and 93% on GPT-OSS via CuTe DSL kernel fusion and MXFP8 precision. Through software optimizations alone, DeepSeek-V3 training throughput increased 1.3x within three months, highlighting the value of full-stack hardware-software co-design.

Salesforce announced the acquisition of AI customer service platform Fin for $3.6 billion, aiming to enhance its agentic AI capabilities and expand Agentforce applications in customer service. This deal continues the trend of consolidation in enterprise AI customer support. On the same day, Meta launched a new 「AI Mode」 on Facebook, transforming the search bar into a conversational tool that answers user queries by accessing public information from its Groups, Reels, and Marketplace platforms. The feature is currently rolling out gradually in the U.S. These moves reflect accelerating commercialization of AI in both enterprise services and consumer-grade search.

Anthropic has released a new economic study introducing a framework to track the usage of Claude Code—examining who uses it, for what tasks, and how domain expertise affects task success rates. The study found that domain experts achieve higher success rates, but the gap between intermediate users and experts is 「moderate,」 suggesting that reaching proficiency is sufficient for effective use. Anthropic announced plans to incorporate Claude Code usage metrics into its Economic Index to monitor AI’s impact on work patterns and labor markets, providing quantitative insights into AI’s real-world effects on employment structures.

Google has been named a Leader in the 2026 IDC MarketScape Worldwide Security Information and Event Management (SIEM) Vendor Assessment. The report highlights Google's AI-native architecture, vertical integration, and Mandiant threat intelligence as key differentiators. Leveraging custom silicon, operational infrastructure, Gemini models developed through DeepMind, and internal security expertise embedded into an agent evaluation loop, Google achieves cost efficiency and iteration speed difficult for third-party APIs to replicate. AI agents are shifting SOC workflows from alert triage to proactive threat hunting, with some customers reporting up to 97% reduction in alerts. A unified data lake and continuous UDM search across time further strengthen its technical differentiation.

Current Robotics has released Curr-0, a full-body dexterous manipulation model that unifies robot mobility, posture adjustment, and hand-level fine manipulation through a single end-to-end trained policy, enabling real-time coordinated adjustments of body and hands during motion. Training data comes from its proprietary HumanEx exoskeleton system, totaling 21,000 hours—humans wearing the exoskeleton perform real-world tasks while full-body poses, joint movements, and hand motions are recorded, shifting data growth from 「robot deployment hours」 to 「human task hours.」 The company has built a full-stack closed loop covering data, modeling, evaluation, and deployment, reducing reliance on physical robot deployment.

Analysts have raised concerns about the financial sustainability of the AI industry, primarily due to the fundamentally 「unpredictable」 nature of token usage costs—enterprises cannot anticipate how many tokens a single prompt will return or how many attempts are needed to obtain satisfactory outputs, making budgeting akin to a 「slot machine.」 Some companies’ strategies linking employee performance metrics to AI usage—termed 「tokenmaxxing」—have backfired, causing surging token consumption without corresponding business value. Reductions in corporate AI spending directly threaten revenue pipelines and profitability paths for providers like OpenAI and Anthropic, while Apple’s entry into integrated, privacy-focused, free AI could further erode the paid consumer AI market.