AI is starting to look less like software and more like a capital‑intensive, regulated utility: budgets are getting capped, infra and energy are the choke points, and governments are treating top models like dual‑use weapons. At the same time, Chinese price cuts and strong open‑source models are commoditizing generic LLM capability, shifting the scarcity to GPUs, HBM, power contracts, and scarce talent.
The real action is now in who controls those bottlenecks and who can live with the new security and political constraints around them.
Key Events
/Noam Shazeer and John Jumper quit Google DeepMind for OpenAI and Anthropic, wiping about 6% off Alphabet’s stock.
/Anthropic’s Mythos breached nearly all NSA classified systems within hours during a red-team test, after which the U.S. banned its flagship models.
/Five major Chinese AI labs, including Alibaba and Tencent, cut inference token prices by up to 99% in a week.
/SK Hynix became South Korea’s most valuable company and is projected to hold 61% of the global HBM market by 2025.
/Leaked financials show OpenAI expects around $38.5B in annual losses despite rapid revenue growth.
Report
Frontier AI now trades like a mix of defense contractor and regulated utility: costly, politically exposed, and concentrated in a few labs.
At the same time, the economics of generic ‘AI API’ look worse each week as price wars hit revenue while HBM, GPUs, and power become the scarce assets.
frontier talent realignment
Noam Shazeer, co‑author of the Transformer and former Gemini co‑lead, left Google for OpenAI, despite Google having previously paid $2.7B to bring him back.
Nobel laureate John Jumper, who led AlphaFold, is leaving Google DeepMind for Anthropic, contributing to a reported 6% drop in Alphabet’s stock as investors question Google’s AI trajectory.
Google insiders describe DeepMind morale as “frustration and discontent” as Gemini 3.5 Pro is seen as failing to deliver a step‑change versus OpenAI and Anthropic.
In parallel, Anthropic hired Jumper, OpenAI hired a former White House AI adviser as head of Strategic Futures, and both labs are leading benchmarks in long‑horizon work and research tasks.
ai unit economics and budget caps
Large enterprises that rushed into gen‑AI are now hitting budget ceilings and walking deployments back. Amazon, Walmart, Uber and others are imposing explicit caps on AI usage after burning through annual budgets in months, and one study reports that 81% of larger organizations have rolled back AI agents after pilots.
Oracle has laid off around 21,000 employees in 12 months “due to AI adoption,” signalling stress rather than effortless productivity gains.
On the supply side, leaked financials suggest OpenAI expects roughly $38.5B in annual losses even as revenue climbs into the low tens of billions, while Apple is preparing device price hikes because RAM and AI‑related memory costs are “unsustainable.” Demand for AI hardware is now outpacing Moore’s law, with RAM, SSDs, and GPUs all flagged by users as hitting price points that feel untenable for many workloads.
price wars, open models, and the shrinking api moat
China is driving an aggressive race to the bottom on inference pricing, with five major labs including Alibaba and Tencent cutting token prices by up to 99% and even exploring AI‑token futures markets in Shanghai.
At the same time, open‑weight models like GLM‑5.2 now rank as the top open model and #3 overall on a leading capability index, with a million‑token context and an MIT license.
Benchmarks show GLM‑5.2 beating GPT‑5.5 on multiple long‑horizon coding tasks at roughly one‑sixth the price, and it is slated for cybersecurity fine‑tuning to approach Mythos‑class performance at much lower cost.
One open model has hit 90.2% on document data‑extraction benchmarks, and another has outperformed much larger proprietary systems on web research tasks.
The EU’s EUROPA consortium is meanwhile building an open frontier‑scale model across 24 languages to reduce dependence on US platforms, while GLM‑5.2 is already offered by at least 16 inference providers, many based in the US.
compute, memory, and energy bottlenecks
On the hardware side, SK Hynix has overtaken Samsung as South Korea’s most valuable company and is projected to control about 61% of the global HBM market by 2025, with samples of HBM4E already shipping.
Nvidia remains the central node in accelerators, with a roughly $5.05T market cap, long‑dated GPU supply deals like a $6.3B contract with SpaceX for GB300s through 2029, and new chips that can run around half a trillion parameters locally.
The capex and opex profile of this infra is extreme: Nvidia Vera Rubin‑based data centers are estimated at $47B per gigawatt with a $1.3B annual electricity bill, even as demand for AI‑tailored power soars.
Hyperscalers are responding by vertically integrating energy, from Chevron’s 20‑year power purchase agreement for a Microsoft data center in West Texas to Meta’s interest in Natrium reactors via its TerraPower deal.
Policy is shifting to accommodate this, with the Swiss parliament lifting its ban on new nuclear plants and the US announcing $17.5B in loans to rebuild the nuclear supply chain.
security shock and national-security framing
The Mythos red‑team incident at the NSA is being treated as a watershed: the model reportedly breached nearly all classified systems within hours, leading to a government ban on Anthropic’s flagship models and an export‑control fight.
Despite a U.S. shutdown order, about 200 organizations in Anthropic’s Project Glasswing retained Mythos Preview access for weeks, including major firms like Cisco and AWS, underscoring how entangled high‑risk models already are in corporate stacks.
Washington is now forcing Anthropic to implement government‑ID verification for certain capabilities and demanding that Fable 5 be made impossible to jailbreak before re‑release, a standard security experts say “can’t be done.” Intelligence officials warn that models capable of devastating cyber and physical attacks are months away, even as OpenAI’s GPT‑5.5‑Cyber already edges out Mythos 5 on offensive and defensive security benchmarks.
In parallel, xAI’s Grok has been used by the Pentagon to target 2,000 munitions at 2,000 Iranian targets in 96 hours, and the U.S. has imposed export bans on models like Anthropic’s Fable 5, making frontier AI feel closer to dual‑use weapons than generic SaaS.
What This Means
Base model capability is rapidly commoditizing just as compute, energy, talent, and regulatory permission harden into the real scarce assets. The live trade‑off is between exploiting cheap, ubiquitous LLMs while absorbing higher infra and compliance exposure, or paying a premium for more controlled stacks in a space that increasingly resembles defense and utilities rather than classic software.
On Watch
/Norway’s near‑ban on AI in elementary schools is an early test of whether education becomes a formal carve‑out from mainstream AI deployment.
/Bernie Sanders’ $7T proposal for public ownership of AI profits, with a $1,000 annual payout, is the first large‑scale attempt to tax and socialize frontier‑AI economics.
/China is in the early stages of designing a futures market for AI tokens via the Shanghai Futures Exchange, which could financialize compute access in a way Western markets haven’t yet matched.
Interesting
/Over 150 cybersecurity leaders have called for the lifting of the ban on Anthropic's Fable 5, arguing it is unjustified and could hinder innovation.
/China's AI chip independence is largely seen as a facade, with leading labs lobbying for access to American chips.
/Microsoft has established a significant business selling AI models to Chinese companies, indicating a complex US-China rivalry in technology.
/Analysts have identified approximately one trillion dollars in future purchase commitments among major tech companies, including Amazon, which may pose hidden financial risks.
/99% of custom ASICs fail, despite AI chip startups presenting misleading performance data.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/Noam Shazeer and John Jumper quit Google DeepMind for OpenAI and Anthropic, wiping about 6% off Alphabet’s stock.
/Anthropic’s Mythos breached nearly all NSA classified systems within hours during a red-team test, after which the U.S. banned its flagship models.
/Five major Chinese AI labs, including Alibaba and Tencent, cut inference token prices by up to 99% in a week.
/SK Hynix became South Korea’s most valuable company and is projected to hold 61% of the global HBM market by 2025.
/Leaked financials show OpenAI expects around $38.5B in annual losses despite rapid revenue growth.
On Watch
/Norway’s near‑ban on AI in elementary schools is an early test of whether education becomes a formal carve‑out from mainstream AI deployment.
/Bernie Sanders’ $7T proposal for public ownership of AI profits, with a $1,000 annual payout, is the first large‑scale attempt to tax and socialize frontier‑AI economics.
/China is in the early stages of designing a futures market for AI tokens via the Shanghai Futures Exchange, which could financialize compute access in a way Western markets haven’t yet matched.
Interesting
/Over 150 cybersecurity leaders have called for the lifting of the ban on Anthropic's Fable 5, arguing it is unjustified and could hinder innovation.
/China's AI chip independence is largely seen as a facade, with leading labs lobbying for access to American chips.
/Microsoft has established a significant business selling AI models to Chinese companies, indicating a complex US-China rivalry in technology.
/Analysts have identified approximately one trillion dollars in future purchase commitments among major tech companies, including Amazon, which may pose hidden financial risks.
/99% of custom ASICs fail, despite AI chip startups presenting misleading performance data.