Kill Switch Fallout, OpenAI's $39B Loss, and a $60B Cursor Deal

U.S. shuts down Anthropic models globally, triggering sovereign AI push

On Friday, the U.S. government cut off global access to Anthropic’s most powerful AI models, including Fable 5 and Mythos 5, effectively using a kill switch to restrict foreign access. The move triggered panic among politicians worldwide, especially in Europe, where it reignited calls for sovereign AI — national control over AI models, computing infrastructure, and data. The EU relies on non-EU countries for over 80% of its technology and 70% of cloud computing, while the U.S. and China together control roughly 90% of global AI computing infrastructure, according to Epoch AI. In the UK, MP Al Carns said hospitals and researchers were suddenly stripped of access to Fable 5, and former security minister Tom Tugendhat argued that national security is now more about code than cannons. The European Commission had released the European Technological Sovereignty Package on June 3, targeting €422 billion in investment over a decade, though critics called it very feeble. Competitors like Mistral are reportedly in talks to raise $3.5 billion, and politicians across Europe called for accelerated domestic AI development. Experts like Sandra Wachter noted that the shutdown made Europe feel the pain of dependency firsthand, and warned that the U.S. action, while perhaps justified, could undermine American soft power and drive allies to reduce reliance on American AI over the long term.

The shutdown of Anthropic’s Mythos model sparks a global scramble for sovereign AI →

Leaked OpenAI financials show $38.5 billion net loss in 2025

OpenAI’s 2025 audited financials were leaked and reported by blogger Ed Zitron and the Financial Times, revealing a net loss of approximately $38.5 billion, up from $5 billion in 2024. Excluding restructuring charges and non-cash items, the adjusted loss was $8 billion. Revenue reached $13.07 billion, up from $3.7 billion in 2024, and ChatGPT hit one billion monthly active users. Total spending hit $34 billion, including $19 billion for R&D and nearly $6 billion for sales and marketing. The gap between net loss and adjusted loss stemmed from a non-cash charge of roughly $30 billion related to how certain investor rights were accounted for before restructuring. OpenAI said in March it was generating $2 billion in monthly revenue. The company may pursue an IPO as early as September at a valuation of up to $1 trillion. By comparison, Anthropic has a $965 billion valuation and an annualized revenue run rate of approximately $47 billion, though its losses remain unknown.

OpenAI’s balance sheet remains the most mysterious—and consequential—in business →

Anthropic faces class action over Claude Max usage claims

Anthropic is facing a class action lawsuit filed June 14 in the U.S. District Court for the Northern District of California by customer Karl Kahn. The suit alleges that Anthropic misled customers about the amount of usage included with its Claude Max 5x and Max 20x subscription tiers, which cost $100 and $200 per month respectively. Kahn claims he overpaid for both subscriptions and received less usage than advertised, and that the company did not clearly explain how usage was measured, making it impossible for consumers to verify if they were getting what they paid for. The lawsuit seeks damages, restitution, injunctive relief, attorneys’ fees, and class certification for all U.S. residents who purchased or upgraded to those plans between April 9, 2025, and the present. Anthropic did not immediately respond to a request for comment. The case comes days after Anthropic launched Claude Fable 5 and Mythos 5, which drew criticism over usage limits and invisible safeguards, and the U.S. government later ordered suspension of access to those models for foreign nationals.

Anthropic Faces Lawsuit Over Allegedly Misleading Claude AI Pricing - Decrypt →

Microsoft shareholders sue over Azure slowdown and AI spending

A securities class action was filed on June 12 in Seattle federal court by the City of St. Clair Shores Police and Fire Retirement System against Microsoft, CEO Satya Nadella, and CFO Amy Hood. The lawsuit alleges that Microsoft defrauded investors by failing to disclose slowing growth in its Azure cloud business and the need for heavy AI infrastructure spending. On January 29, the day the lawsuit anchors, Microsoft shares fell about 10%, wiping out roughly $357 billion in market value after a quarterly earnings report showed Azure revenue growing 39%, a deceleration from 40%, and capital spending of $37.5 billion, up nearly 66% from a year earlier and above analyst expectations. The complaint argues that Microsoft attributed Azure’s cooling growth to capacity constraints caused by diverting computing resources to AI research and development and to its Copilot assistant. Microsoft has not yet filed a response, and the allegations remain untested. The plaintiffs must show the company knowingly misled investors about something material.

Microsoft sued by shareholders over Azure slowdown and AI spending →

SpaceX acquires Cursor maker Anysphere for $60 billion

SpaceX has agreed to acquire Anysphere, the maker of the AI coding assistant Cursor, in a deal valued at roughly $60 billion, according to Reuters. Cursor helps developers write, edit, understand, and debug software, functioning as a workflow-integrated tool rather than a general chatbot. The acquisition highlights a shift in AI opportunities toward systems that can handle large chunks of work with minimal supervision, as developers are early adopters willing to pay significant subscription fees for productivity improvements. OpenAI, Anthropic, and Google are similarly building tools that move beyond chat interfaces toward task-completion models. The deal signals that AI coding assistants may be among the most monetizable applications of the technology.

SpaceX just spent $60 billion on Cursor — and it proves AI chatbots aren’t the future anymore →

Alibaba launches first AI models for robots as agents take centre stage

Alibaba released its first suite of AI models for robots, led by RynnBrain, designed to help machines understand space, objects, and motion. In a demonstration from Alibaba’s DAMO Academy, a robot using the system identified a piece of fruit and placed it in a basket. Alongside RynnBrain, Alibaba announced Qwen3.7-Max, a large language model positioned as a foundation for AI agents, claimed to run autonomously for up to 35 hours without performance degradation. Alibaba describes itself as an AI factory operating all five layers of the stack: chips, agentic cloud, models, model-serving platforms, and applications. The strategic shift from chatbots to agents drives the launch, with robotics representing the most physical expression of that bet. Alibaba faces competition from other Chinese tech giants and American labs, but holds advantages in hardware and supply chain. Pricing and availability details have not been disclosed.

Alibaba unveils AI models for robots as China’s focus shifts to agents →

Kimi K2.7 code model rivals Opus 4.8 at a fraction of the cost

Moonshot AI released Kimi K2.7, an open-source model with a trillion total parameters and 32 billion active at any given time. It features a 256k context window, a thinking mode for pre-response deliberation, and a modified MIT license on Hugging Face. Compared to its predecessor, Kimi K2.7 requires 30% fewer tokens for reasoning tasks and improved its Kimiko Bench score from 51 to 62, and MLS Bench Light from 26 to 35. On MCP Atlas, it scored 76, approaching GPT-5.5 and Opus 4.8. The model excels in coding, voxel art generation, instruction-following, and agent task management, delivering comparable or superior results using fewer tokens. Limitations include not matching top-tier models on highly complex tasks and occasional UI glitches. It is suited for SMEs, independent developers, and educational institutions prioritizing local deployment and cost-effectiveness.

How Kimi K2.7 Code Rivals Opus 4.8 and is 5X Cheaper to Run →

Claude Voice mode gains multilingual support and push-to-talk

Anthropic has started rolling out multilingual support for Claude Voice Mode, with users spotting new language options including Spanish (Latin America). The updated settings also include a Push-to-talk mode that lets users press and hold a button while speaking and release to send, alongside the existing hands-free conversational mode. The rollout is gradual and inconsistent across users; some have access while others do not. One X user praised the update as 100 times better than ChatGPT. TestingCatalog previously uncovered references to the upgrades. A new voice-related icon resembling a phone call button was spotted in a recent iOS build, but Anthropic has not explained its purpose. The author’s iPhone Claude app now supports hands-free and push-to-talk modes with broader language selection including German, Portuguese, Chinese, Japanese, Russian, and Ukrainian.

Claude’s biggest Voice mode upgrade yet is finally rolling out →

Microsoft may use DeepSeek V4 as optional model in Copilot Cowork, shifts to usage-based billing

Microsoft is considering a self-hosted, fine-tuned version of DeepSeek V4 as a cheaper model option for Copilot Cowork, according to Axios. The company is also shifting Cowork to usage-based pricing, because Copilot EVP Charles Lamanna said flat-rate pricing is not sustainable due to users who do hundreds of tasks a week. Microsoft already made a similar move with GitHub Copilot. A Chinese AI model could draw criticism, but Microsoft stresses that DeepSeek would be optional and fully hosted on Azure with customized safeguards against bias. A final decision is expected in the coming weeks. CEO Satya Nadella recently argued for an ecosystem of AI models that companies can pick and tune for specific use cases and costs, calling AI a consumption business.

Microsoft’s Copilot Cowork moves to usage-based billing and may tap DeepSeek →

Florida Attorney General sues OpenAI over child safety, column argues for broader accountability

Florida Attorney General James Uthmeier filed a lawsuit against OpenAI alleging that AI systems can cause harm, particularly to children, citing links between AI tools and self-harm, suicide, and violence. Investigators cited cases where ChatGPT was allegedly used in planning a mass shooting at Florida State University. The state argues OpenAI pushed a powerful product to market while understating risks and prioritizing growth over safety. A column in the Tampa Bay Times argues that the lawsuit misses two deeper issues: AI companies train on human-created content without paying creators, and they resist being called publishers while generating content for millions. The column calls for requiring AI companies to pay for inputs and to be held accountable for harmful outputs, testing current legal protections from a 1996 law that shields platforms from liability.

Florida’s Attorney General should push even harder for AI safety | Column →