Capital & Industrial Strategy

The Trump administration has blocked Anthropic from distributing its Mythos and Fable 5 models to foreign nationals, citing national security concerns — but without publishing a clear legal basis or regulatory framework for the decision. According to Wired, Anthropic itself cannot determine precisely what it did wrong, leaving the company and its enterprise customers in a legal vacuum. A Bloomberg report adds further complexity: select early testers have retained preview access despite the shutdown of all other versions, suggesting the enforcement is selective rather than technically absolute.

A Fortune investigation identifies a phone call from Amazon CEO Andy Jassy as a trigger point in the escalating chaos, hinting at the degree to which hyperscaler relationships with the White House are shaping regulatory outcomes in real time. JPMorgan has already responded by restricting Anthropic access for staff in Hong Kong, per Semafor. The practical consequence is that any global enterprise deploying Anthropic's frontier models now faces geographic compliance risk that cannot be managed with standard legal frameworks, because the rules themselves are being written ad hoc. This is not a narrow export control story — it is a signal that AI model access is being weaponised as a geopolitical instrument, and enterprises have no reliable mechanism for anticipating where the boundary will move next.

Baseten is reportedly finalising a $1.5 billion raise at a $13 billion valuation, having already completed a major round just months prior, according to TechCrunch and The Wall Street Journal. This is a reported intention, not a closed deal with confirmed terms. The strategic logic is clear: as open-weight models improve in capability, enterprises are increasingly motivated to run inference on dedicated infrastructure rather than pay per-token to OpenAI or Anthropic. Baseten sits at the intersection of that cost pressure and the growing ecosystem of deployable open models.

The speed of successive mega-rounds reflects a broader dynamic: capital is concentrating in the picks-and-shovels layer of AI rather than the model layer, where margin compression from open-source competition is most acute. General Intuition, which trains embodied AI and world models using a dataset of 2 billion videos annually from its Medal gaming platform, is separately in talks to raise $300 million at a $2 billion valuation per TechCrunch. Both rounds are reported intentions subject to finalisation.

AWS is in active talks to sell its Trainium AI chips externally to third-party data centres, a confirmed strategic intent announced by CEO Andy Jassy, who has characterised this as a $50 billion market opportunity per TechCrunch. Simultaneously, The Wall Street Journal reports that Google is using its financial firepower to win external data centre customers for its TPU silicon, explicitly replicating Nvidia's go-to-market playbook of subsidising adoption to build ecosystem lock-in.

These are parallel but distinct moves: Amazon is targeting the data centre operator market where Nvidia H100 and B200 clusters dominate; Google is using cloud credits and preferential pricing to pull workloads onto TPU infrastructure. Neither has disclosed closed deals with external customers at scale. The strategic logic for both is to reduce Nvidia's pricing leverage over their own internal AI compute costs while simultaneously opening a revenue line. The risk is that neither Trainium nor TPU has Nvidia's software ecosystem — CUDA compatibility remains the dominant switching barrier — and both hyperscalers are betting that model-level software portability will erode that moat faster than Nvidia can reinforce it.

OpenAI hired Noam Shazeer — co-inventor of the Transformer architecture and a foundational figure in modern LLM development — from Google DeepMind, and in the same week brought on Dean Ball, former lead author of the Trump White House's AI Action Plan, per TechCrunch and Politico. These hires serve different but complementary functions in an IPO context: Shazeer enhances the technical credibility narrative for institutional investors evaluating frontier model capability, while Ball provides direct intelligence on an administration whose regulatory posture — as the Anthropic crackdown demonstrates — can materially alter the operating environment for AI companies overnight.

The IPO timing is also being influenced by market dynamics beyond OpenAI's control. SpaceX's record $86 billion public debut is generating what Wellington Management's Head of Late-Stage Growth describes as a potential 'distribution event' for private markets, returning capital to LPs and creating fresh dry powder for deployment — per Bloomberg. California's tax authorities are separately anticipating a windfall from the wave of AI IPOs, though CNBC notes the revenue impact may be structurally blunted by deferred equity vesting schedules and interstate tax competition.

US regulators have taken their most significant step to date to accelerate power grid connections for AI data centres, per Bloomberg, while attempting to balance surging utility costs for consumers. This is confirmed regulatory action, not a proposal. Separately, the FT reports that the Trump administration is increasingly adopting sovereign AI investment postures — including direct government equity stakes in strategic AI companies — drawing explicit comparisons to China's state-led industrial strategy per FT.

China's government simultaneously announced measures to promote AI integration into domestic consumption sectors, per Reuters, signalling that both major AI powers are now deploying state capacity on the demand side as well as supply side. The US grid acceleration directly addresses one of the binding physical constraints on AI data centre expansion — interconnection queue backlogs have been a documented bottleneck. The policy signal is that infrastructure permitting is now being treated as a national competitiveness issue, not merely a utility regulation matter.