The U.S. just demonstrated a real AI kill-switch by shutting Anthropic’s new models off for most of the world, and big customers are already shifting toward open-weight and Chinese alternatives. Underneath that, the economics look frothy: frontier labs are burning tens of billions, infra and compute costs are spiking, and meme-priced platforms like SpaceX are using inflated stock to buy AI assets at 20–30x revenue.
The bets now are about which stacks can’t be turned off by regulators, and which vendors survive once the subsidies and hype compress back toward cash flow.
Key Events
/U.S. Commerce invoked export controls to cut off Anthropic’s Fable 5 and Mythos 5 models for all foreign nationals, forcing a global suspension.
/OSS models have overtaken proprietary models in market share over the last three months on OpenRouter.
/OpenAI generated $13B of revenue in 2025 but still lost about $21B that year.
/SpaceX completed a record IPO and then agreed to acquire coding startup Cursor for $60B in stock at roughly 20–30x revenue.
/Broadcom, Tensordyne and HPE announced the Napier AI chip platform, claiming around 13x higher token throughput than NVIDIA Blackwell.
Report
Frontier AI just got treated like a weapons system: the U.S. pulled a kill‑switch on Anthropic’s Fable 5/Mythos 5 models three days after launch, cutting off every foreign user overnight.
At the same time, open‑weights and Chinese models are gaining share while Western labs burn tens of billions a year and meme‑priced platforms like SpaceX lever up to buy AI assets at 20–30x revenue.
the model kill‑switch and politicized AI risk
The U.S. Commerce Department placed Anthropic’s Fable 5 and Mythos 5 under export control, forcing the company to suspend access for all foreign nationals, including its own non‑U.S. staff.
Access was cut globally just three days after launch, after Amazon flagged a jailbreak and Chinese access concerns to the White House.
This is the first clear use of a government 'kill switch' on a commercial frontier model, with criminal and civil penalties threatened for non‑compliance.
U.S. allies like the U.K. were denied carve‑outs, and over 150 cybersecurity leaders have publicly asked for the ban to be lifted, highlighting both geopolitical fracture and industry dependence on a handful of U.S. labs.
open‑weights, China, and the parallel AI stack
Open‑weights models have quietly overtaken proprietary ones in recent usage data, with OSS now ahead on OpenRouter’s share metrics. Z.ai’s GLM‑5.2 is the first open‑weights model to clear 80% on Terminal‑Bench, offers a 1M‑token context window, and ranks first in coding Elo, giving enterprises a credible non‑proprietary workhorse.
Regional models like Rio de Janeiro’s Rio3.5 and Kimi K2.7 Code are beating larger‑brand peers on targeted benchmarks and efficiency, reinforcing the idea that 'good enough' local stacks are now widely available.
In parallel, Chinese providers jumped from under 2% to over 45% of global model traffic in a year, capped by DeepSeek raising $7.4B with investors locked in for five years, signaling deep, patient capital behind a China‑aligned ecosystem.
A Fortune 500 has already committed to move half its coding to GLM‑5.2, explicitly walking away from Anthropic after the Fable disruption.
loss‑making foundations and meme‑priced platforms
Leaked financials suggest OpenAI generated $13B of revenue in 2025 but still lost about $21B, driven by $34B of spend including roughly $19B on R&D. Industry commentary pegs marketing at ~44% of revenue, reinforcing that current pricing is essentially a loss‑leader land grab rather than a sustainable SaaS margin story.
On the demand side, a $200‑per‑month ChatGPT plan could cost OpenAI up to $14,000 if a power user fully exercised the quota, while an enterprise CFO reportedly racked up a $500M Claude API bill in one month, underlining how agentic workflows can explode token usage.
Equity markets are leaning into this: SpaceX, still reportedly losing around $5B a year, briefly traded at 94x sales after its IPO, then used $60B of stock to buy Cursor at 20–30x its ~$2B ARR.
Commenters now openly describe SpaceX as a 'meme stock,' and AI‑heavy indices are seen as vulnerable if these valuations correct in line with underlying cash flows.
the compute and data‑center squeeze (and non‑NVIDIA options)
Compute and hosting costs are spiking across the stack: Hetzner has hiked dedicated server prices by 3–4x and cloud VMs by ~120% since May, while Broadcom’s VMware pricing has reportedly risen up to 10x, pushing Tesco to migrate 40,000 workloads off the platform.
Amazon just secured a $17.5B loan for AI data center build‑out and committed $10B to a Missouri campus expected to consume about 2.5B gallons of water in 2025, underscoring how capital‑ and resource‑intensive AI capacity has become.
At the same time, protests have blocked or delayed roughly $130B of U.S. data center projects this year, and Arizona has paused tax incentives, making siting and permitting a first‑order constraint.
On the silicon side, Broadcom, Tensordyne and HPE are touting a Napier AI chip with ~13x Blackwell token throughput, while Tensordyne separately claims 17x tokens per watt vs NVIDIA, signaling an emerging market for non‑CUDA accelerators.
Retail SSD and NAND supply is also being pulled toward AI data centers, with commentary linking this to rising consumer storage prices and shortages.
What This Means
AI is now behaving like a regulated, capital‑intensive utility layered on top of a speculative tech bubble, with control over models, data centers, and silicon increasingly concentrated in a few political jurisdictions and meme‑valued platforms. The live risk is that your AI exposure is simultaneously over‑dependent on a handful of vulnerable vendors and under‑priced relative to the regulatory, infrastructure, and macro shocks now clearly on the table.
On Watch
/A German court ruling that Google is liable for false statements made by its AI systems sets an early precedent for search and chatbot legal exposure.
/The EU AI Act’s high‑risk and transparency rules, including mandatory labeling and watermarking of AI outputs, become enforceable from August 2026 and will shape how agentic systems can be shipped in Europe.
/China is designing a futures market for AI tokens, hinting at financialization of inference capacity as a tradable commodity.
Interesting
/The Pentagon's use of generative AI tools by 1.5 million personnel underscores the military's reliance on AI for operational efficiency.
/China's universities have cut 12,000 'obsolete' degrees to adapt to the AI era, showcasing a significant educational shift.
/The US is perceived to be pushing researchers towards China due to restrictive policies, potentially strengthening China's AI sector.
/The U.S. may face a future where open-source AI is banned, potentially increasing reliance on foreign AI technologies.
/A simulation predicted AI could double U.S. GDP growth by 2030, but also increase underemployment from 8% to 14%.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/U.S. Commerce invoked export controls to cut off Anthropic’s Fable 5 and Mythos 5 models for all foreign nationals, forcing a global suspension.
/OSS models have overtaken proprietary models in market share over the last three months on OpenRouter.
/OpenAI generated $13B of revenue in 2025 but still lost about $21B that year.
/SpaceX completed a record IPO and then agreed to acquire coding startup Cursor for $60B in stock at roughly 20–30x revenue.
/Broadcom, Tensordyne and HPE announced the Napier AI chip platform, claiming around 13x higher token throughput than NVIDIA Blackwell.
On Watch
/A German court ruling that Google is liable for false statements made by its AI systems sets an early precedent for search and chatbot legal exposure.
/The EU AI Act’s high‑risk and transparency rules, including mandatory labeling and watermarking of AI outputs, become enforceable from August 2026 and will shape how agentic systems can be shipped in Europe.
/China is designing a futures market for AI tokens, hinting at financialization of inference capacity as a tradable commodity.
Interesting
/The Pentagon's use of generative AI tools by 1.5 million personnel underscores the military's reliance on AI for operational efficiency.
/China's universities have cut 12,000 'obsolete' degrees to adapt to the AI era, showcasing a significant educational shift.
/The US is perceived to be pushing researchers towards China due to restrictive policies, potentially strengthening China's AI sector.
/The U.S. may face a future where open-source AI is banned, potentially increasing reliance on foreign AI technologies.
/A simulation predicted AI could double U.S. GDP growth by 2030, but also increase underemployment from 8% to 14%.