Governments just proved they can literally switch off top-tier AI models and hold companies liable for whatever those models say, while communities start blocking the data centers needed to run them. At the same time, capital is bidding AI infra and platforms up to trillion‑dollar valuations even as core players burn tens of billions a year.
The game is shifting from who has the cleverest model to who can survive the regulators, the power grid, and the next repricing of AI hype.
Key Events
/US government invoked an AI 'kill switch,' ordering Anthropic to disable its Fable 5 and Mythos 5 models globally under export controls.
/SpaceX IPO raised $75B at a roughly $2.1T valuation, making Elon Musk the first reported trillionaire.
/SpaceX agreed to acquire coding assistant Cursor for $60B in stock, one of the largest software acquisitions ever.
/China announced a $295B AI data center buildout while domestic AI providers jumped from under 2% to over 45% of global model traffic in a year.
/Fox is acquiring Roku in a $22B streaming deal to lock up distribution.
Report
Washington just proved it can turn a frontier model off like a light switch, and Berlin just said AI answers are legally the company's own speech.
Capital is meanwhile pricing AI infra like it’s a sure thing, with trillion‑dollar IPOs and double‑digit billion losses in the same breath.
the ai kill switch and liability shock
The US government ordered Anthropic to suspend access to its Fable 5 and Mythos 5 models for all foreign nationals just three days after launch, citing national security and export controls.
Officials described this as the first use of a global AI 'kill switch,' forcing the company to disable its most powerful models worldwide while leaving earlier Claude versions untouched.
Amazon security teams reported a jailbreak-style flaw in Fable 5 to the Commerce Department, triggering the directive even as cybersecurity experts argued the model posed no unique threat.
At the same time, a German court ruled that Google is legally liable for inaccuracies in its AI Overviews, treating generated answers as Google's own statements rather than neutral search results.
The EU Commission has already begun examining the Anthropic decision, while enterprise buyers see KPMG retracting an AI report over hallucinations and Microsoft blocking internal use of Fable 5.
china’s parallel ai stack
China is planning a $295B AI data center buildout that explicitly aims to exclude Nvidia hardware, signaling a push for a self‑reliant compute base.
Chinese AI providers increased their share of global model traffic from under 2% to over 45% in a single year, rapidly eroding US labs’ dominance.
Huawei launched its 92‑billion‑parameter openPangu 2.0 model, activating 6 billion parameters per use and integrating more than 2,000 AI agents into HarmonyOS 7 across devices.
Taiwan is considering criminal penalties for exporting AI chips or servers to China, while the US formally classified frontier AI as a controlled export and blocked foreign access to Anthropic’s newest models.
Commenters increasingly describe a two‑pole AI world, with Americans in walled gardens around US providers and the rest of the world potentially standardizing on a Chinese or open‑source stack.
the infrastructure squeeze: power, water, and politics
Amazon secured a $17.5B loan and committed $10B to a new AI data center campus in Missouri even as it disclosed using about 2.5 billion gallons of water in its data centers last year.
Oracle’s Project Jupiter AI data center in New Mexico is under scrutiny over water use, while Oracle’s stock fell as investors focused more on data center costs than AI growth.
Local backlash is mounting: protests and permitting fights have blocked an estimated $130B in data center projects this year, Seattle imposed a one‑year ban on new AI data centers, and Mississippi residents are suing xAI over constant noise.
The US EPA refused to set nationwide environmental rules for data centers, even as 77% of Americans told pollsters they fear AI‑driven data centers will raise electricity bills and grid operators warn that power demand is becoming a bottleneck.
Cloud economics are tightening too, with Oracle halving its free‑tier CPU and RAM limits and component suppliers prioritizing NAND for AI data centers over retail hardware.
open weights versus gated frontier models
MIT’s GLM‑5.2 open‑weights model became the first open system to clear 80% on Terminal‑Bench and now tops open‑model Elo rankings, edging out Claude Fable 5.
The city of Rio de Janeiro released a 397B‑parameter Rio‑3.5‑Open model on Hugging Face, while Kimi K2.7‑Code and MiMo Code offer open‑source coding systems with better token efficiency and up to 6x speedups over earlier versions.
Google’s DiffusionGemma, an Apache‑licensed text‑generation model, can generate text up to 4x faster on dedicated GPUs, trading some quality for speed in tasks like context compression and exploratory coding.
In contrast, Anthropic engineered Fable 5 and Mythos 5 to be less helpful for AI research, then had them placed under US export controls that block foreign access, while AWS Bedrock plans to require data sharing with Anthropic for Mythos‑class models.
Across forums, developers increasingly describe open‑source AI as something that 'must succeed,' both to counter US export bans and to avoid reliance on single US vendors.
frontier economics and the frothy capital window
OpenAI’s losses jumped nearly eightfold to about $34B in 2025, including $19B on R&D and $6B on sales and marketing, with internal forecasts of roughly $14B in losses this year.
Internal calculations show a $200/month ChatGPT subscription could cost OpenAI around $14,000 in compute if fully utilized, and the company is already contemplating major price cuts to compete with Anthropic.
Despite these unit‑economics worries, OpenAI has filed a confidential draft S‑1, while Databricks spent about $20M training its DBRX model and Jeff Bezos’s Prometheus raised $12B at a $41B valuation to build an 'Artificial General Engineer.' On the equity side, SpaceX’s IPO raised $75B at a roughly $1.77–$2.1T valuation—around 94x sales—amid reports of nearly $5B in annual losses, and it almost immediately agreed to buy Cursor for $60B, or 20–30x projected 2026 AI revenue.
Commenters increasingly describe AI‑linked stocks as propping up major indices, warning that a downturn in this sector could drag on broader market stability and leave retail investors carrying the bag after insiders exit.
What This Means
The AI stack is being carved up by governments, energy grids, and capital markets at the same time that open models are catching up, so platform risk now lives as much in law, power, and balance sheets as in benchmark scores. For anyone exposed to AI, the core decision is how much of your future to park on a few regulated US platforms versus spreading bets across open, non‑US, and infra‑adjacent plays that can survive a kill switch or a repricing.
On Watch
/Coinbase registering an AI agent with the SEC as an autonomous investment advisor hints at regulators treating AI systems as regulated financial actors rather than just tools.
/Taiwan’s proposal to criminalize AI chip and server exports to all of China could rapidly reshape GPU supply and force on‑shore or alternative compute strategies.
/Tensordyne’s Napier compute chips, claiming 17x tokens per watt and 13x throughput versus Nvidia Blackwell, are an early sign that serious challengers to Nvidia’s GPU dominance are emerging.
Interesting
/Canadian Prime Minister Mark Carney highlighted the dangers of over-reliance on American providers in light of U.S. restrictions on Anthropic's AI models.
/Amazon's stake in Anthropic, valued at around $200 billion, positions it as a significant player in the AI market, influencing both investment and regulatory dynamics.
/Tensordyne announced AI compute chips that are 17x more efficient in tokens per watt compared to NVIDIA Blackwell.
/Samsung is building floating data centers on ships, which have received regulatory approval.
/A closed-door simulation predicted AI could double U.S. GDP growth while increasing underemployment from 8% to 14% by 2030.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/US government invoked an AI 'kill switch,' ordering Anthropic to disable its Fable 5 and Mythos 5 models globally under export controls.
/SpaceX IPO raised $75B at a roughly $2.1T valuation, making Elon Musk the first reported trillionaire.
/SpaceX agreed to acquire coding assistant Cursor for $60B in stock, one of the largest software acquisitions ever.
/China announced a $295B AI data center buildout while domestic AI providers jumped from under 2% to over 45% of global model traffic in a year.
/Fox is acquiring Roku in a $22B streaming deal to lock up distribution.
On Watch
/Coinbase registering an AI agent with the SEC as an autonomous investment advisor hints at regulators treating AI systems as regulated financial actors rather than just tools.
/Taiwan’s proposal to criminalize AI chip and server exports to all of China could rapidly reshape GPU supply and force on‑shore or alternative compute strategies.
/Tensordyne’s Napier compute chips, claiming 17x tokens per watt and 13x throughput versus Nvidia Blackwell, are an early sign that serious challengers to Nvidia’s GPU dominance are emerging.
Interesting
/Canadian Prime Minister Mark Carney highlighted the dangers of over-reliance on American providers in light of U.S. restrictions on Anthropic's AI models.
/Amazon's stake in Anthropic, valued at around $200 billion, positions it as a significant player in the AI market, influencing both investment and regulatory dynamics.
/Tensordyne announced AI compute chips that are 17x more efficient in tokens per watt compared to NVIDIA Blackwell.
/Samsung is building floating data centers on ships, which have received regulatory approval.
/A closed-door simulation predicted AI could double U.S. GDP growth while increasing underemployment from 8% to 14% by 2030.