The Gist — Capital & Industrial Strategy

Top Line

The White House's forced revocation of SK Telecom's access to Claude Mythos has transformed a bilateral export control dispute into a G7-level sovereignty crisis, with Macron and Modi openly seeking workarounds to U.S. kill-switch authority over frontier AI.

Google suffered a significant talent blow as Noam Shazeer, co-lead of Gemini, defected to OpenAI ahead of its IPO — one of the most consequential researcher moves in the current AI race and a direct hit to Alphabet's model development capability.

Enterprise AI's 'tokenmaxxing' era is ending: Uber blew through its annual AI budget in months, companies are cutting Claude licenses, and NEA's Tiffany Luck confirms enterprises are still struggling to demonstrate ROI — signalling a shift from uncritical adoption to disciplined deployment.

Goldman Sachs and Morgan Stanley are forming separate walled teams to manage the anticipated concurrent IPOs of OpenAI and Anthropic, confirming both listings are being treated as imminent and capital-markets-ready.

The U.S. government awarded $500 million to Nvidia-backed SandboxAQ for chipmaking materials research, marking a significant federal commitment to AI-adjacent deep-tech beyond software and model development.

Key Developments

Anthropic's Mythos Blackout Becomes a Global AI Sovereignty Flash Point

What began as a bilateral export control issue — the White House ordering Anthropic to revoke SK Telecom's access to Claude Mythos over alleged ties to China, as reported by Wired — has rapidly escalated into a foundational geopolitical dispute about who controls access to frontier AI. At the G7, French President Macron led talks exploring 'trusted partner' frameworks to route access to advanced models without direct U.S. government veto power, as reported by Bloomberg. Macron has since said he expects progress on broadening access, per Reuters. Trump told reporters negotiations with Anthropic were 'going fine', per Reuters, though the underlying regulatory structure remains unresolved. Semafor reports that Trump advisers are internally divided on how to structure any government equity stake in Anthropic as part of a broader deal, with Cabinet members favouring competing frameworks — suggesting the negotiation involves more than access controls and potentially touches on ownership and governance.

The financial contagion is already visible. JPMorgan Chase cut off Anthropic access for its Hong Kong staff, following Goldman Sachs, as reported by the Financial Times. Both moves reflect the acute compliance risk for global financial institutions caught between U.S. export control enforcement and operational continuity in Asia. The Cohere CEO used the G7 moment to frame the choice for allied nations as 'sovereign AI or digital serfdom', per Fortune, a framing that will accelerate European and emerging-market investment in domestic model capability. Dario Amodei and Sam Altman jointly called on G7 leaders to resist splintering the AI ecosystem, per the Financial Times, but the structural incentive for allies to diversify away from U.S.-controlled frontier models is now impossible to ignore.

Why it matters

The U.S. government has demonstrated it can unilaterally switch off access to the world's most capable AI models for foreign users, transforming export controls from a theoretical risk into a live operational threat that will reshape sovereign AI investment strategies globally.

What to watch

Whether the 'trusted partner' framework Macron is negotiating produces a durable multilateral access mechanism, or whether allied nations accelerate investment in domestic frontier model development as a strategic hedge against U.S. access controls.

Noam Shazeer Defects to OpenAI, Deepening Google's Talent Crisis Pre-IPO

Noam Shazeer, Google's vice president of engineering and co-lead of Gemini, has left for OpenAI, as confirmed by CNBC and Bloomberg. Shazeer is not a peripheral figure: he is a co-author of the original 'Attention Is All You Need' transformer paper and a foundational architect of modern LLM capability. His departure is among the most consequential individual talent moves in the current AI cycle, both for what it removes from Google DeepMind and what it adds to OpenAI immediately ahead of its IPO. The timing is strategically loaded — OpenAI's IPO roadshow will now carry Shazeer's credibility as a signal of technical depth to institutional investors.

For Google, this follows a pattern of senior model talent departures and compounds the challenge of maintaining Gemini's competitive position against GPT and Claude. Amazon's AI chief has separately acknowledged that Nova2 lags behind the latest OpenAI and Anthropic releases, per CNBC, suggesting the frontier is consolidating around a smaller number of leaders. The concentration of top research talent at OpenAI, combined with its forthcoming IPO, creates a self-reinforcing dynamic: capital access funds compute, which funds model quality, which attracts talent.

Why it matters

Shazeer's move to OpenAI is a talent-as-capital event — at this level of seniority, researcher moves directly affect model quality trajectories and carry significant investor signalling weight ahead of OpenAI's public listing.

What to watch

Whether Google responds with aggressive retention packages or structural changes to its research organisation, and how Shazeer's role at OpenAI is positioned in the IPO prospectus.

Enterprise AI ROI Reckoning: The Tokenmaxxing Hangover

The enterprise AI adoption cycle is hitting a disciplined spending phase. NEA partner Tiffany Luck, speaking to TechCrunch, confirmed that enterprises are still struggling to demonstrate clear ROI from AI deployments — a view borne out by operational data. Uber reportedly exhausted its annual AI budget within months under a 'tokenmaxxing' mandate encouraging maximum AI usage. Companies are cutting Claude licenses for non-essential business units, and Meta shut down its internal AI usage leaderboard. These are not isolated anecdotes; they represent a structural shift in how CFOs are treating AI spend — moving from growth-at-any-cost to cost-per-output accountability.

This matters for the AI vendor ecosystem in several ways. First, it creates pricing pressure on API-based model providers as enterprise buyers consolidate usage to high-value workflows rather than broad access. Second, it shifts the competitive advantage toward vendors who can demonstrate measurable productivity outcomes rather than raw capability. Third, it may suppress the near-term revenue growth rates that AI company IPO valuations are pricing in. The tension between frontier model investment cycles — which remain capital-intensive — and enterprise willingness to pay for deployed outputs is the central financial dynamic of the next 18 months.

Why it matters

The ROI reckoning represents a maturation of enterprise AI buying behaviour that will compress vendor pricing power and force a differentiation between AI tools that generate measurable value and those that remain productivity theatre.

What to watch

Whether AI vendors respond by shifting to outcome-based pricing models, and how OpenAI and Anthropic frame enterprise revenue quality in their IPO disclosures.

Europe's Industrial AI Push: Siemens, Schneider, and Mistral Lead Shop-Floor Deployment

Bloomberg's deep-dive into European industrial AI documents a concerted push by Siemens, Schneider Electric, and Mistral to deploy AI in manufacturing environments — positioning Europe's industrial base as a proving ground for applied AI rather than consumer or software applications. The strategic logic is clear: Europe's manufacturing sector faces acute efficiency pressure from energy costs and Asian competition, and industrial AI offers measurable ROI in predictive maintenance, energy optimisation, and production scheduling, areas where the ROI case is stronger than in knowledge-work applications.

OVHcloud separately announced plans to develop frontier AI models to become Europe's second LLM player after Mistral, per Reuters. The Anthropic access crisis will accelerate both moves: European enterprises now have a concrete operational risk case for sourcing AI from EU-domiciled providers, and the political will to fund sovereign model development — already present in France's Mistral backing — has been materially strengthened.

Why it matters

Europe's industrial AI deployment is transitioning from pilot to scale, and the Anthropic blackout has provided the political catalyst to accelerate investment in sovereign model infrastructure that complements rather than merely replicates U.S. frontier capability.

What to watch

Whether the EU's industrial AI push generates sufficient commercial traction to sustain domestic frontier model development, or whether OVHcloud and Mistral remain dependent on U.S. compute and model architecture at the infrastructure layer.

Capital Markets: IPO Pipeline, AI Credit Risk, and Pension Capital Flowing to Emerging Market Data Centers

Goldman Sachs and Morgan Stanley are constructing information barriers to manage simultaneous mandates for the OpenAI and Anthropic IPOs, per the Wall Street Journal. This is a confirmed structural preparation, not speculation — both banks are treating both listings as imminent enough to require compliance infrastructure today. Meta is separately exploring Wall Street financing structures for its $600 billion AI infrastructure push, with former Goldman Sachs executive Dina Powell McCormick leading discussions, per the Financial Times — a signal that Silicon Valley is increasingly accepting capital market instruments previously considered alien to tech self-funding models. Man Group has meanwhile warned that AI credit markets are at risk of a 'violent' correction, per CNBC, flagging that debt financing of AI infrastructure may be mispricing duration and demand risk.

In emerging markets, a Canadian pension fund has acquired an 8.2% stake in CtrlS, India's multi-data-center operator, per TechCrunch. This is a closed deal with confirmed terms, and represents the institutionalisation of the data center infrastructure trade — long-duration pension capital is now competing with private equity and hyperscalers for physical AI infrastructure in high-growth markets. The U.S. government's $500 million award to Nvidia-backed SandboxAQ for chipmaking materials research, per Reuters, adds a deep-tech federal spending layer that complements the data center and model investment cycle.

Why it matters

The concurrent IPO preparation for OpenAI and Anthropic, combined with Man Group's credit market warning, signals that AI capital markets are entering a phase where distribution of risk — not just accumulation of capital — becomes the dominant financial concern.

What to watch

Whether the Anthropic government negotiation and ongoing access uncertainty creates material disclosure risk for its IPO prospectus, and how underwriters price that regulatory overhang.

Signals & Trends

U.S. Export Controls Are Accelerating Sovereign AI Investment Outside America's Orbit

The Mythos blackout has done in days what years of European AI policy debate could not: it has provided a concrete, operational proof point that dependence on U.S.-controlled frontier AI is a sovereign risk, not merely a theoretical concern. OVHcloud's announcement of frontier model ambitions, Macron's active G7 diplomacy on trusted-partner access frameworks, the Cohere CEO's 'digital serfdom' framing, and the structurally motivated nature of allied governments to fund domestic alternatives are all directionally consistent signals. The investment implication is a medium-term increase in capital flows toward non-U.S. AI infrastructure and model development — not to match U.S. frontier capability immediately, but to establish optionality and reduce single-point dependency. For investors, this means European industrial AI, sovereign cloud infrastructure, and domestically anchored model providers in France, India, and the Gulf are likely to attract government-backed and institutional capital at above-market rates of return on a risk-adjusted basis.

China's AI Market Is Bifurcating Into a Liquid Pair Trade — With Microsoft in the Middle

Two distinct signals are emerging from China's AI capital markets. First, Bloomberg reports a clear winner-loser pair trade forming in Chinese AI equities, with investors concentrating positions in the perceived commercial leader and shorting the laggard — a sign that China's domestic AI market is reaching a consolidation inflection point where capital is making differentiated bets rather than sector-wide allocations. Second, Kingboard Laminates has surged 570% as mainland investors doubled their stake in the PCB laminate supplier — a materials play on AI hardware demand that mirrors the Nvidia-adjacent trade in Western markets. Against this backdrop, Bloomberg separately reports that Microsoft has built a substantial business selling OpenAI models to Chinese enterprises despite the U.S.-China AI rivalry. This creates a structural tension: U.S. export controls are tightening on frontier model access for allies like SK Telecom, while a major U.S. technology company is simultaneously growing AI model revenue in China. The regulatory consistency of this posture will be tested.

Physical AI Is Creating a New Data Labelling Infrastructure Investment Category

TechCrunch's reporting on XDOF — a startup being paid by AI labs to collect robot training data — points to an emerging infrastructure layer in the physical AI stack that is structurally analogous to the early data labelling and annotation market for LLMs. As robotics and world model development scale, the constraint shifts from model architecture to training data quality and diversity in physical environments. Odyssey's $1.45 billion valuation, backed by Amazon, reflects investor conviction that world models are the next foundational layer beyond LLMs. The capital implication is a nascent but fast-growing market for physical AI data infrastructure — collection hardware, annotation tooling, and simulation environments — that sits below the model layer and may prove as durable a business as the compute infrastructure layer has been. Investors who missed the LLM data infrastructure build should be watching this category now, while valuations and competitive dynamics are still forming.

Explore Other Categories

Read detailed analysis in other strategic domains

Compute & Infrastructure

Nvidia's $25 billion bond sale is its first in five years — and at current margins, it doesn't need the cash for operations. The real signal is strategic: a company preparing to absorb supply chain risk, secure packaging capacity, or vertically integrate in ways its historically fabless model never required. Watch where the money actually lands.

Frontier Capability Developments

The White House has demanded Anthropic guarantee zero jailbreaks before rereleasing its most advanced models — a condition security researchers confirm is technically unachievable for any large language model. This isn't a negotiating position; it's a category error baked into policy. The precedent it sets could structurally hobble US frontier labs while foreign competitors operate unconstrained.

Geopolitics & Sovereign Positioning

Majorities in 11 of 15 surveyed countries — including key US allies — believe China leads the global AI race, yet institutional trust remains low. That gap is the strategic terrain that matters: cost-sensitive developers, Global South governments, and SMEs are choosing Chinese models on price-performance grounds regardless. Washington's export controls may have inadvertently engineered the competitive model they most needed to prevent.

Public Policy & Governance

France's intelligence service is replacing Palantir with a domestic AI vendor — the first concrete government divestment from a US provider on strategic autonomy grounds. Meanwhile, the EU AI Act's newly adopted Omnibus amendments have diluted fundamental rights protections before the ink is dry. The compliance floor just shifted for every multinational operating in Europe.