Weekly Big Tech Official Announcement News (June 7–13, 2026 Issue)
Update Date: 2026/6/14
Executive Summary
In this week's Big Tech official announcements, the trend of AI agents shifting from experimental chats to long-running, production-grade operations has become clear. OpenAI's acquisition of Ona, the MCP/plugin foundations from Salesforce and xAI, and Adobe's CX Enterprise Coworker all signal the development of agent execution platforms connected to enterprise systems. Google, Databricks, and Microsoft have expanded enterprise data utilization and continuous workspaces, while Oracle, NVIDIA, Apple, Meta, and AMD have accelerated investments in AI infrastructure, sovereign AI, and data centers. Meanwhile, safety, governance, and transparency are increasingly becoming competitive requirements.
*The content of the created article was input into Gamma to automatically generate slides. Please take a look here if you find slides easier to view.
1️⃣ AI agents shift to long-running, production-grade operations
1-1. OpenAI strengthens Codex's sustainable execution foundation with Ona acquisition
Source URL: OpenAI Official Announcement
OpenAI announced an agreement to acquire Ona, signaling a plan to incorporate secure cloud execution and orchestration technology into Codex. The referenced page highlights that Codex's weekly active users have exceeded 5 million, a 400% increase since the beginning of the year. As AI agent tasks shift from minutes-long assistance to long-running tasks spanning hours or days, a foundation that manages operations while controlling customer-managed cloud environments, credential scope, logs, reviews, and data boundaries is becoming a critical requirement for enterprise adoption.
1-2. Salesforce and xAI expand agent connection foundations with MCP and plugins
Source URL: Salesforce Developers Blog / xAI Official Announcement
Salesforce launched Headless 360 as a major theme in its Summer ’26 Release, opening key features as APIs, MCP tools, and CLI. It is significant that MCP-compatible clients like Claude, ChatGPT, and Cursor can connect to Salesforce organizations and data with OAuth authentication. xAI also released a plugin marketplace for Grok Build, enabling the distribution of skills, slash commands, agents, hooks, and MCP servers. AI agents are evolving from standalone apps into execution foundations connected to enterprise systems.
1-3. IBM/ServiceNow and Databricks strengthen enterprise AI production deployment frameworks
Source URL: IBM Newsroom / Databricks Blog
IBM and ServiceNow have expanded their multi-year collaboration, aiming to address the lack of AI-ready data and the modernization of legacy application layers. The vision supports data governance and autonomous IT operations by combining Workflow Data Fabric, watsonx.data, Ansible, Instana, Terraform, and Vault. Databricks has also formalized 'Forward Deployed Engineering,' establishing a framework to support AI deployment outcomes closer to the customer's site. Support models that bridge the gap between enterprise AI—which often stalls at the PoC stage—and actual business results are beginning to be prioritized.
1-4. Adobe turns marketing operations into agentic AI with CX Enterprise Coworker
Source URL: Adobe Newsroom
Adobe has launched general availability for 'CX Enterprise Coworker,' an agentic AI solution for marketing and customer experience. It functions as a 'central intelligence layer' that integrates data from Adobe and third-party apps to coordinate across analysis, content generation, journey design, and campaign execution. By adopting a headless design based on open standards like MCP and A2A, and utilizing the Adobe Experience Platform as a context layer, it is positioned as a foundation for automating production workflows with agents while collaborating with existing stacks and external AI platforms.
2️⃣ Development of enterprise data, MCP, and developer ecosystems
2-1. Google to commercialize multi-step enterprise search with Agentic RAG
Source URL: Google Research Blog
Google Research and Google Cloud have released Agentic RAG for the Gemini Enterprise Agent Platform. For multi-source, multi-hop queries that are difficult for traditional single-step RAG, a Root Agent, Planner, Query Rewriter, RAG Agent, Search Fanout Agent, and Sufficient Context Agent collaborate to generate answers while iterating through search planning, query decomposition, context sufficiency judgment, and re-searching. It shows up to a 34% improvement in accuracy compared to standard RAG and maintains a high correct answer rate in cross-corpus environments like FramesQA, positioning it as a practical foundation for constructing grounded answers even from fragmented enterprise data.
2-2. Databricks announces OpenSharing for sharing data, models, and Agent Skills
Source URL: Databricks Press Release
Databricks has announced 'OpenSharing,' a new open standard under the Linux Foundation. This extends Delta Sharing to the agent era, providing a vendor-neutral protocol that allows for the secure, zero-copy sharing of Agent Skills, AI models, and unstructured data in addition to data. Furthermore, through Iceberg IRC client support and integration with on-premises/private cloud storage such as Everpure and MinIO, it enables the application of cloud-side AI and analytics without moving data. OpenSharing is positioned as an initiative to expand the AI utilization foundation of the Databricks platform, including Unity Catalog, Genie, and Agent Bricks, into a multi-platform and cross-border asset sharing layer.
2-3. Microsoft expands Copilot Notebooks into a continuous AI workspace
Source URL: Microsoft Tech Community
As part of the Microsoft 365 Copilot Notebooks update, Microsoft has expanded Copilot Notebooks to Copilot Chat users, not just traditional Microsoft 365 Copilot users, positioning them as knowledge workspaces that can be shared within teams. Within OneNote and the Microsoft 365 Copilot app, users can aggregate files such as Word, PowerPoint, Excel, and Outlook emails into a single notebook, enabling cross-file Q&A, mind mapping, and study support in Study Guide format. The feature is being rolled out to commercial and educational Copilot Chat users, and with capabilities like collaborative work on the same notebook even in mixed-license environments, it is being strengthened as a continuous knowledge-sharing foundation rather than just a one-off chat tool.
3️⃣ Acceleration of AI Infrastructure, Sovereign AI, and Data Center Competition
3-1. Oracle Announces Record Earnings and OCI Growth Driven by AI Demand
Source URL: Oracle Official Announcement
Oracle announced its fiscal 2026 fourth-quarter and full-year results, reporting Q4 total revenue of $19.2 billion, cloud revenue of $9.9 billion, and Cloud Infrastructure revenue (including OCI) of $5.8 billion. The reference page highlights that RPO has expanded to $638 billion (up 363% year-over-year), much of which consists of cloud contracts for AI training and inference. Additionally, AI-related services are driving growth, with the Oracle Multicloud AI Database growing by 404%. Meanwhile, the earnings release does not mention OpenAI's use of models or Codex via Oracle Universal Credits, so that point must be treated as supplementary information based on separate sources.
3-2. NVIDIA Expands AI Factory, Memory, and Physical AI Collaboration in South Korea
Source URL: NVIDIA Newsroom / NVIDIA Blog
NVIDIA is integrally expanding AI factories, next-generation memory, sovereign AI, and Physical AI in South Korea through collaborations with SK hynix, NAVER, SK Telecom, and LG Group. With SK hynix, it is advancing the joint development of next-generation memory for Vera Rubin-generation AI supercomputers, Vera CPUs, RTX Spark-equipped PCs, and Jetson Thor robots, as well as semiconductor design, manufacturing, and digital twin utilization using CUDA-X, PhysicsNeMo, Omniverse, OpenUSD, and cuOpt. With LG Group, it is utilizing Isaac Sim, Isaac Lab, Isaac GR00T, and Cosmos to build home robots, modular robots, autonomous robotics for logistics and manufacturing sites, and liquid-cooled AI factories and GPU cloud infrastructure compliant with DSX. AI competition is spreading from models to the co-design of entire industrial infrastructures, including memory, robots, and data centers.
3-3. Apple Expands Private Cloud Compute to Google Cloud and NVIDIA GPUs
Source URL: Apple Newsroom / NVIDIA Blog
In its WWDC26-related announcements, Apple presented updates to the Foundation Models framework, Core AI, App Intents, and agentic coding in Xcode 27. Apple Foundation Models are built in collaboration with Google and Gemini technology, and server-side inference is extended to NVIDIA Blackwell GPUs on Google Cloud via Private Cloud Compute. Private Cloud Compute is designed as a high-privacy inference foundation connecting on-device AI and cloud AI, ensuring that even on the server side, developers and operators cannot view user data through hardware-rooted trust, encrypted communication, and remote attestation utilizing NVIDIA Confidential Computing.
3-4. Meta and AMD Strengthen Regionally Distributed AI Infrastructure and Sovereign AI Investment
Source URL: Meta Newsroom / AMD Official Announcement
Meta announced a partnership with Reliance Industries to lease a 168MW AI-ready data center in Jamnagar, Gujarat, India. Meta will cover the facility's energy and cooling costs via seawater desalination, and is also promoting the introduction of approximately 1GW of renewable energy in total through contracts with CleanMax and Fourth Partner Energy. AMD announced plans to invest up to £2 billion in the UK over the next five years to strengthen research infrastructure and talent development supporting AI, quantum, next-generation computing, and photonic networks. AI foundations are becoming closely linked not only to cloud companies but also to sovereign AI infrastructure and industrial policies in each country.
3-5. NVIDIA RTX Spark and AWS SageMaker Expand Local AI and Long-Context Inference Foundations
Source URL: NVIDIA Blog / AWS Machine Learning Blog
NVIDIA showcased RTX Spark at a Korean PC bang event, indicating the direction of a high-performance AI PC platform for individuals that spans gaming, creation, and local AI. RTX Spark is positioned as a flagship for thin laptops and small desktops, enabling 1440p/100fps+ RTX gaming as well as local generative AI and agent execution. KRAFTON's PUBG Ally also introduced AI co-op characters using NVIDIA ACE technology. Meanwhile, AWS made NVIDIA's open LLM "Nemotron 3 Ultra" available in SageMaker JumpStart, enabling high-load agentic workloads such as long-duration agent inference, coding, and deep research to be executed at low cost on the cloud with a 1-million token context length and hybrid Transformer-Mamba MoE.
4️⃣ Frontier Models, Specialized AI, and Research Automation
4-1. Anthropic Enhances Long-Context and Long-Term Tasks with Claude Fable 5 and Mythos 5
Source URL: Anthropic Official Announcement
Anthropic announced Claude Fable 5 for general availability and Claude Mythos 5 for limited access. Both models handle 1M-token class long-context and demonstrate performance significantly exceeding previous models in long-term tasks such as software development, knowledge work, vision, and scientific research. On the other hand, safety classifiers are triggered for queries related to cybersecurity, biology/chemistry, and model distillation, with Fable 5 designed to automatically fall back to Claude Opus 4.8. Mythos 5 is provided to trusted users such as Glasswing partners under relaxed constraints, particularly for cyber defense and life science research. An approach that designs high performance and safety controls as one is becoming a key evaluation axis for corporate adoption of frontier models.
4-2. Anthropic Organizes Recursive Self-Improvement Where AI Accelerates AI Development
Source URL: Anthropic Institute
In "When AI builds itself," the Anthropic Institute organized the current state where AI is accelerating the AI development process itself. The reference page shows that the amount of code merged per Anthropic engineer has increased 8-fold compared to 2024, and as of May 2026, over 80% of the lines of code merged into production are derived from Claude. On the other hand, there is a note that line count is not a metric that can fully measure quality or true productivity, and that perceived productivity improvements in surveys tend to be overestimated. It concludes that as AI supports the development of successor models and research processes, the design of monitoring, verification, stop-ability, and human-led direction and review becomes increasingly important.
4-3. Google DeepMind Explores High-Speed Text Generation with DiffusionGemma
Source URL: Google Blog / Google DeepMind
Google DeepMind released "DiffusionGemma," an experimental open model that uses a text diffusion method. The reference page explains that it is a 26B Mixture of Experts model under the Apache 2.0 license, designed to activate only 3.8B parameters during inference and generate 256 tokens at once, aiming for up to 4x speedup on GPUs. While it may be inferior to standard Gemma 4 in terms of quality, it is positioned as a model suitable for developer workflows where latency is critical, such as inline editing, code completion, local conversational AI, and non-linear text structure generation, indicating that generation speed itself is becoming a key competitive axis for user experience.
4-4. Anthropic Verifies Practicality of Specialized AI with Chemistry and Biology Agents
Source URL: Anthropic Research / Anthropic Research
As part of its efforts to bring Claude closer to practical chemical applications, Anthropic has published an evaluation of structure prediction and determination using NMR spectra, indicating a direction where AI can interpret experimental data that researchers paste into the chat. Furthermore, in a separate study on biology research agents titled "Paving the way for agents in biology," they demonstrate that by combining a deterministic retrieval layer called "gget virus" for the task of acquiring sequence data from NCBI Virus, the accuracy of state-of-the-art agent groups improves significantly to nearly 100%. They conclude that for domain-specific AI, it is essential to design not only model performance but also auditable and reproducible data acquisition foundations and deterministic tool layers.
5️⃣ AI Safety, Governance, and Social Implementation Risks
5-1. Google DeepMind, up to $10 million for multi-agent safety research
Source URL: Google DeepMind Official Blog
Google DeepMind, in collaboration with Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org, has launched a call for proposals for multi-agent safety research with a scale of up to $10 million. The focus is on understanding the collective behaviors and risks that arise when multiple AI agents communicate, negotiate, trade, and cooperate in digital environments. Key areas include sandboxes/testbeds for multi-agent safety evaluation, the science of agent networks, infrastructure protocols supporting identity, reputation, and commitments, and the monitoring and control of deployed agent groups. The goal is to enhance the stability and transparency of the entire large-scale agent ecosystem, rather than just the safety of individual models.
5-2. IBM survey: Control gap widens in corporate AI adoption
Source URL: IBM Newsroom
Based on a survey of 2,000 technology CxOs across 33 regions and 19 industries, IBM has published findings on the control gap in corporate AI adoption. The referenced page highlights that 70% of respondents stated that technology adoption by business units is outpacing the IT department's visibility, and 77% answered that AI adoption is overtaking existing governance capabilities. Furthermore, while the number of AI agents is expected to increase by 38% by 2027, only 11% of respondents said they were "fully prepared" to scale. The focus of corporate AI is shifting from simply increasing model performance and adoption numbers to reducing incidents, managing costs, achieving real-time visibility, and strengthening accident response capabilities through embedded governance.
5-3. Microsoft and OpenAI sound alarm on AI brand abuse and influence operations
Source URL: Microsoft Security Blog / OpenAI Official Announcement
Microsoft has published an investigation into social engineering that impersonates AI brands such as ChatGPT, Microsoft Copilot, DeepSeek, and Claude, citing instances of phishing, malvertising, and credential theft, credit card fraud, and malware infection via SEO abuse. Meanwhile, OpenAI has reported the suspension of ChatGPT account clusters originating from China, identifying two influence operation clusters ("Data Center Bandwagon" and "Tech and Tariffs") that attempted to intervene in public opinion regarding AI data center power consumption, electricity costs, and US-China technology competition and tariffs. The proliferation of AI itself is becoming a new vector for attackers and foreign powers to manipulate public opinion.
5-4. MIT and Google DeepMind verify AI dependency in education and news verification
Source URL: MIT News / Google DeepMind Blog
Google DeepMind published the results of an RCT on Gemini Guided Learning targeting middle school students in Sierra Leone, evaluating math score improvements and conversation design for 1,763 students over 8 weeks. The study reported that students using guided learning showed a learning effect of +0.258 standard deviations compared to the control group, equivalent to 1.2 to 1.7 years of typical progress. Meanwhile, the MIT Media Lab investigated the impact of AI chatbots on news verification and human judgment, demonstrating an "AI dependency paradox" where accuracy in detecting fake news improved by 21% during support, but performance dropped by 15 points after the AI was removed. In the education and media sectors, it is necessary to design for both the effectiveness of AI and the risks of cognitive dependency.
5-5. US AI regulation draft and Dario Amodei's policy proposals accelerate transparency debate
Source URL: Representative Jay Obernolte / Dario Amodei Official Blog
In the US Congress, a discussion draft for frontier AI regulation, the "Great American AI Act," has been released, indicating a direction that requires AI developers above a certain scale to operate under a federal framework that includes safety evaluations, transparency, and auditability. Additionally, Anthropic CEO Dario Amodei, in his essay "Policy on the AI Exponential," argues that frontier models should be treated like aircraft, proposing mandatory third-party testing for models exceeding a compute threshold and risk assessment and reporting in four areas: (1) cybersecurity, (2) biological weapons, (3) loss of control, and (4) automated R&D that could accelerate these risks. AI regulation is shifting from a focus on voluntary guidelines to a framework that includes verifiability and government intervention and suspension authority.
Comprehensive Analysis
This week, overall, the main battlefield of AI competition shifted from "smarter models" to "operational capabilities that safely bundle data, tasks, and computing resources inside and outside the enterprise to continuously turn them into results." Codex, Copilot Notebooks, Agentic RAG, OpenSharing, and CX Enterprise Coworker are all moves to make AI not just a one-off generative tool, but a central layer that maintains business context, connects to external systems, and operates in an auditable manner. At the same time, the AI infrastructure race is expanding to GPUs, memory, power, cooling, and regional distribution, and safety and regulation are beginning to be redesigned based on the premise of uncontrollable risks after deployment.
Future Points of Interest
The value of AI agents is entering a phase where it will be evaluated not just by response accuracy, but by the maturity of the execution control foundation, including credentials, logs, reviews, and data boundaries.
With the spread of MCP and OpenSharing, the axis of competition for corporate AI is shifting from functional differences in individual apps to connectivity with external systems and data assets.
The moves by Apple, Meta, and AMD are signs that AI infrastructure is expanding from cloud-centric models to national industrial strategies that include privacy, sovereign AI, and regional distribution.
Research from Anthropic and Google DeepMind shows that for domain-specific AI, reproducible data acquisition layers and secure tool integration are becoming more important than the models themselves.
As cases of AI dependency, brand abuse, and influence operations increase, companies need to prepare user education, verification procedures, and monitoring systems in parallel with promoting adoption.
Draft US AI regulations and policy recommendations may strengthen the trend of requiring third-party verification, transparency, and kill switches for frontier model companies.
