Capital & Industrial Strategy

The Trump administration has resolved — partially — a two-week standoff with Anthropic by allowing Mythos 5 to be accessed by more than 100 vetted US companies and government agencies, while a second advanced Anthropic model remains blocked. Simultaneously, the White House asked OpenAI to stagger the release of GPT-5.6, limiting it to a 'limited preview' for government-approved partners before broader rollout. Both companies complied while publicly objecting: OpenAI stated that 'this kind of government access process should not become the long-term default' because 'it keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.' Wired, TechCrunch, Bloomberg

The Bessent angle adds a critical dimension: Treasury Secretary Bessent's engagement was reportedly triggered by bank warnings about advanced AI models threatening Federal Reserve payment rails — meaning the security concerns driving these restrictions are concentrated in financial infrastructure, not just defence. Semafor This explains why the cleared access list includes both companies and agencies and why non-American employees of those entities are included — the regime is about controlling deployment context, not nationality per se. The FT notes that industry unease over Washington's ad hoc regulatory approach persists even as the immediate restrictions ease, which is the correct framing: what has emerged is an informal pre-release review process with no statutory basis, no defined criteria, and no predictable timeline. FT

OpenAI has disclosed plans for Jalapeño, a custom inference chip co-developed with Broadcom, joining Google (TPUs), Apple, Amazon (Trainium/Inferentia), and SpaceX in building proprietary silicon. The strategic logic is consistent across all these actors: inference workloads are volume-sensitive, margin-compressible, and increasingly predictable enough to optimise for in custom hardware. Nvidia's H100 and successor architectures are priced for a market where the buyer has no alternative; custom silicon eliminates that leverage at the inference layer even if training remains Nvidia-dependent for now. TechCrunch, Semafor

For OpenAI specifically, Jalapeño is not merely a cost play — it is a prerequisite for the unit economics required to serve a consumer and enterprise market at scale without perpetually widening losses. Sam Altman identified compute infrastructure as a strategic battlefront early, and the Broadcom partnership (rather than a pure in-house design) suggests OpenAI is optimising for speed to deployment over full vertical control. The contrast with Google's full-stack TPU approach is instructive: OpenAI is buying time and capability without the multi-year foundry investment Google absorbed.

OpenAI is weighing pushing its IPO to 2027, having confidentially filed its S-1 with the SEC earlier this month. The company has not held pre-IPO investor meetings or set a timeline, and sources characterise it as acknowledging it 'may be a while' before going public. The market reaction was immediate: Goldman Sachs and Morgan Stanley shares fell, reflecting that both banks had been positioned as lead underwriters and had priced in near-term fee revenue. SoftBank shares also declined. Bloomberg, CNBC

The delay rationale — tech stock volatility — is the proximate cause, but the structural issue is more significant: OpenAI's valuation in private markets has been set by strategic rounds (including Microsoft and SoftBank) at levels that require public market investors to accept a growth story that is simultaneously being pressured by enterprise budget tightening, Chinese model cost competition, and now government-imposed release constraints. A forced IPO into that environment would crystallise those uncertainties at the worst possible moment. The 2027 window implies OpenAI is betting on a clearer profitability narrative — likely tied to enterprise contract conversion — before facing public market scrutiny. CNBC

Zhipu's GLM 5.2 is benchmarking close to current US frontier models at materially lower cost, and the CNBC analysis frames the competitive dynamic accurately: the AI race is shifting from raw capability to intelligence-per-dollar, a metric on which open-source and cost-optimised Chinese models are increasingly competitive. CNBC The two-week Anthropic restriction and the staggered GPT-5.6 rollout have created a precise window in which US enterprise buyers seeking advanced AI capabilities face constrained access to the best US models while Chinese alternatives are unrestricted — at least for non-US-domiciled buyers.

This intersects with the Europe dimension: European buyers, frustrated by dependence on US AI infrastructure and now facing the additional uncertainty of US government-imposed access controls, have a strengthened incentive to develop domestic alternatives or engage with non-US providers. The Wired analysis of European AI ambitions Wired notes that Trump's unpredictability is an asset for European AI industrial strategy — access restrictions make the strategic case for sovereignty more concrete than any policy argument could. Italy joining the US-led Pax Silica initiative Reuters simultaneously shows European governments hedging — joining US-led frameworks while building domestic capacity.

Goldman Sachs bankers are internally flagging the physical economy as the next AI capital wave — encompassing robotics, industrial automation, and physical AI infrastructure. Axios This is consistent with ON Semiconductor's pivot: the company's acquisition of Synaptics is framed explicitly around a $30 billion addressable market expansion into physical AI — edge inference, sensor fusion, and embedded AI for industrial and automotive applications. The market punished the stock on deal announcement, but the strategic rationale reflects where capital allocators are positioning for the post-cloud AI deployment cycle. CNBC

The geothermal energy investment thesis reinforces this: I-Pulse's $250 million US government-backed investment in deep geothermal drilling technology is positioned explicitly as AI data centre power infrastructure. Bloomberg The pattern across Goldman's thesis, ON Semi's M&A, and the geothermal play is the same: capital is beginning to flow from software-layer AI into the physical infrastructure stack — power, networking, edge compute, and industrial deployment — where the next margin pool is expected to form.