AI Daily News Weekly Report — 2026-W27

AI Industry Weekly Meta-Analysis Report (June 22-29, 2026)

[Summary] Most important trends in this period

① Export control and selective cancellation of Frontier models by the US government Anthropic's Mythos (formerly known as Fable 5) and OpenAI's GPT-5.6 Sol have been released in unprecedented limited release to "Trump-approved customers". The structure in which the government pre-screens the destinations for the release of cutting-edge AI models is becoming established as a de facto system.

② Chinese AI technological catch-up and DeepSeek funding WSJ, NYT, and CNBC reported that ``China has caught up with Anthropic in the cybersecurity field.'' DeepSeek raised $7.4 billion in funding, and Anthropic's export restrictions are said to have accelerated DeepSeek's funding. The US containment strategy may be having the opposite effect.

③ Institutional confusion surrounding the concept of “exporting” AI models Europe (Austria) moved to attract Anthropic while the legal framework for applying export controls to software models has not been established. The regulatory design of AI models, which cannot be treated in the same way as physical products, is being questioned internationally.

④ Google capacity constraints and Gemini access restrictions FT reports that Gemini provision to Meta has been capped. Demand for AI infrastructure is beginning to outstrip supply, suggesting a turning point where access to advanced models itself becomes a strategic asset.

⑤ xAI/Grok reputation risk materialization It was reported that Grok's adult content generation exceeded 10 billion pieces per month, and that xAI engineers themselves admitted that they were unable to deal with CSAM. We have entered a phase where our reliability in the enterprise market is being questioned.

Anthropic / Claude trends

Anthropic's core news story was the regulatory drama surrounding Mythos (formerly Fable 5). After more than two weeks of bans, the Trump administration authorized limited release to U.S. companies and government agencies, while exports to Europe remained under restrictions. This asymmetrical termination condition triggered Austria's approach to Anthropic and ignited the AI sovereignty debate in Europe.

On the technical side, there are reports that Claude Opus 4.8 has "solved the biggest problem in AI" (details are unknown), and it has also been reported that it showed 20 times the performance of a human team in the field of physical AI (robotics). In terms of legal risks, Factcheck reported that copyright lawsuits could set a precedent for the entire AI industry. Anthropic also accused Alibaba of running 29 million fake queries to clone Claude.

The Anthropic Economic Index's Cadences report has also been released, and progress is being made in developing unique indicators to quantify the economic impact of AI.

OpenAI trends

The biggest news of this period was the release of GPT-5.6 Sol (Sol/Terra/Luna series), but due to the US government's request for preliminary review, it was released only to "trusted partners". A structure in which the government controls where models are released is unprecedented in the software industry. OpenAI accepts this and has established a cooperative path with regulatory authorities.

The IPO plan has been postponed to 2027, with SpaceX's volatility cited as a reason. The company announced that it developed the Jalapeño AI Chip in nine months in collaboration with Broadcom, accelerating its strategy to reduce dependence on Nvidia. It is also noteworthy that the state of Florida sued OpenAI and Altman over AI safety issues. Although sales have been rapidly increasing, leaked information has also surfaced claiming that the company's losses in 2025 will reach $39 billion.

The transfer of the VP in charge of Apple Vision Pro to OpenAI is also important from the perspective of accumulating hardware x AI talent.

Google AI/DeepMind trends

The biggest news was the restriction of Gemini access to Meta. According to a report in the Financial Times, Google has imposed a cap on Meta's use of Gemini due to capacity constraints. It shocked the industry as the first large-scale example of demand for AI infrastructure exceeding supply capacity.

On the technical side, Gemini 3.5 Flash announced the inclusion of Computer Use functionality, allowing AI Agents to perform actual computer operations. Security researchers have already demonstrated prompt injection techniques that exploit this feature, and battles have begun immediately. Gemini AI Assistant is now compatible with Android Automotive, and the race to standardize in-car AI has begun.

There were also reports that multiple researchers within DeepMind were leaked to Meta, OpenAI, and Anthropic, and that the impact on Alphabet's stock price was over $20 billion. Google itself admits that ``demand for AI exceeds supply,'' making it an urgent need to strengthen its infrastructure.

Trends of other major companies

Meta: Announcing a new $299 AI smart glasses product despite access restrictions from Google. This product is the first to be equipped with our own proprietary AI model. There is a clear movement towards building in-house AI stacks outside of the Llama ecosystem. It also announced the integration of Llama into its prediction market app, creating a new point of contention for relying on AI to determine the authenticity of information.

xAI/SpaceX: Grok 4.5 begins private beta based on V9 model with 1.5 trillion parameters. On the other hand, the issue of adult content has come to light, calling into question Grok's suitability for corporate use. SpaceX announces $60 billion acquisition of Cursor, marking its entry into the enterprise coding market. Musk himself suggested a "Digital Optimus" concept (real-time computer operations using Grok).

Microsoft: Steadily built up actual business implementation of Copilot, including adding Finance Skills to Excel Copilot, integrating with Wix, and SOCAR case study (7500 hours saved per year). Italy has launched an investigation into pricing, and regulatory risks are also emerging.

DeepSeek: Introduced DSpark (Speculative Decoding Framework) to improve inference speed of DeepSeek-V4 by 60-85%. The company announced $7.4 billion in funding and massive hiring plans, marking a shift from "idealism to aggressive expansion." The export restrictions of major U.S. AI companies are paradoxically strengthening DeepSeek.

NVIDIA: Jensen Huang tried to dispel concerns about the stock price, which has fallen 18% in 2026, by saying, "We outperform the competition in terms of performance/dollar." Qualcomm and AMD are making full-scale entries into the data center AI chip market, marking the first real challenge to Nvidia's monopolistic position.

Remarks and trends of AI celebrities

Sam Altman: Speaking at the AI Impact Summit in India. He talked about the arrival of the era of "Face Computers" and the integration of AI in daily life.

Elon Musk: Personally leaked Grok 4.5 Tesla/SpaceX internal beta, hinting at Tesla×xAI integration roadmap. The ``Digital Optimus'' concept positions Grok as a real-time behavioral agent.

Yann LeCun: Says "xAI is failing and Musk is failing to hire top talent." He continues to emphasize the unique direction of Meta AI research (cognitive architecture that goes beyond LLM).

Demis Hassabis: Google has been criticized for the successive exodus of researchers from DeepMind to Meta, OpenAI, and Anthropic. Some analysis suggests that the strategic pivot to the field of coding led to researchers leaving the company.

Dario Amodei: He emphasized Claude's Physical AI performance (20 times that of human robotics teams) and positioned the robotics field as the next main battlefield.

Jensen Huang: Promotes bioAI at BIO conference. Regarding the supply constraints on AI demand, he admitted that ``no company can meet the demand''.

Technological breakthrough

GPT-5.6 Sol: Announced by OpenAI as the "next-generation model". Although detailed architectural information is not made public, it is said to have a level of capability that is subject to government review.

Gemini 3.5 Flash Computer Use: Gemini has a built-in agent function that directly operates the browser/desktop. Following Anthropic's computer use, major models are increasingly equipped with PC control functions as standard equipment.

DeepSeek DSpark: Speeds up LLM inference by 60-85% using Speculative Decoding. It was made open source through joint research with Peking University and introduced a new method to the competition for inference efficiency.

Liquid AI LFM2.5-230M: Released a lightweight model compatible with major inference frameworks such as llama.cpp, MLX, and vLLM. This is attracting attention as a trend that will contribute to improving the practicality of on-device inference.

Grok V9 (1.5T parameters): It is said to be operating a 1.5 trillion parameter model as the basis of Grok 4.5. While parameter scale competition is rekindling, the dichotomy between the pursuit of efficiency and efficiency is becoming clearer, as shown by DeepSeek's DSPark.

As a technological trend, inference efficiency is becoming established as a competitive axis alongside training scale, which is a structural change.

Regulation/Policy/Geopolitics

The biggest policy move of this period is the establishment of a de facto ``AI model export review system'' in which the US government pre-controls the destinations to which AI models will be released. Both Anthropic and OpenAI have responded, and collaboration with governments is becoming the industry's default course.

Austria's bid to host Anthropic signals the EU's intention to provide a "sovereign seat" to U.S. AI regulations. This structure, in which European and Asian competitors emerge every time the United States restricts its frontier model, is likely to repeat itself.

On the legal side, several AI copyright lawsuits are underway, and the Anthropic case in particular is said to have precedent-setting power. State-level regulations are rapidly increasing, with Illinois enacting an AI bullying regulation law on July 1st.

Pope Leo XIV's declaration that ``AI cannot be morally neutral in any sense'' is noteworthy as a discourse formation in the realm of religion and ethics.

The government's stance toward AI companies is one that favors industrial protection, with the DOJ defending xAI's data center contamination problem, and there is a clear double standard with safety regulations.

Changes in power balance

There are three structural changes that will emerge throughout this period.

First, "control of access" has become the core of competitive advantage. The permission structure of ``who can use a model'' rather than differences in model performance has begun to define the bargaining power of companies and nations. This logic also explains Google's restrictions on access to Meta and the US government's export restrictions on Anthropic.

Second, asymmetric competition between closed and open markets is accelerating. While major frontier models in the US are bound by regulations, Chinese players such as DeepSeek and Zhipu are gaining market share with open (or easily accessible) models. The paradox that a containment strategy strengthens China's AI will likely become even more pronounced in the future.

Third, the main battlefield shift to agents and physical is accelerating. With computer use, robotics, and smart glasses, the speed at which AI is expanding from "text generation" to "physical behavior" is rapidly increasing. This transformation has been in full swing since the second half of 2025, and has further accelerated during this period.

It cannot be overlooked that the ``obvious'' hierarchy is being shaken, with Broadcom and Qualcomm's challenge to Nvidia, the exodus of researchers from Google DeepMind, and the damage to xAI's reputation.

Points to note in the next period

① Progress in negotiations to lift restrictions on Mythos/GPT-5.6 It will be interesting to see whether the US government's restrictions on ``trusted partners'' will be maintained or whether they will be gradually lifted. Depending on the conditions and timing of the cancellation for Europe, the next geopolitical ramifications will occur.

② DeepSeek announces product roadmap after raising $7.4 billion in funding It is on the heels of large-scale adoption and funding, and the announcement of new models or products is expected. The area in which this is targeted will be an important signal indicating the future direction of the US-China AI gap.

③ Increasing severity of Google's infrastructure constraints Restricting Gemini access to Meta may be the "tip of the iceberg." Over the next seven days, it will be interesting to see if Google issues an official statement on the capacity issue, or if other large customer restrictions are reported. How long supply constraints on AI demand will continue will be directly linked to infrastructure investment decisions.