Capital & Industrial Strategy

The combined AI infrastructure spend of Google, Amazon, Microsoft, and Meta has crossed $1 trillion since 2023, per Financial Times. Amazon's upward revision to $220 billion in 2026 capex — from $200 billion flagged in April — is the most concrete signal that demand signals from enterprise customers are strong enough to accelerate commitment rather than pause it. AWS's Q2 growth, its fastest since 2021, was anchored by new contracts with Meta and OpenAI, per CNBC, suggesting the hyperscalers are not just spending speculatively but converting pipeline into contracted revenue.

The critical divergence is Meta. Where Microsoft can point to Azure and Copilot revenue expansion to justify capex, Meta's AI spending is producing an $8 billion free cash flow hit with no clear enterprise monetisation path, per Wall Street Journal and Reuters. This split is now the central analytical question for AI equity investors: which companies have closed the loop between infrastructure spend and billable output, and which are still in the faith-based phase of the buildout.

The $15 billion debt facility being arranged by Morgan Stanley for Nexus Data Centers — with Google providing power and lease guarantees for a 1.6-gigawatt Texas project — represents a new archetype in AI infrastructure finance: sovereign-scale project financing backstopped by hyperscaler off-take commitments rather than traditional asset collateral, per Wall Street Journal and CNBC. This deal is announced but not closed — subject to finalisation of terms and lender syndication.

The parallel story at CoreWeave is instructive about the limits of that model. The Financial Times reports that CoreWeave was forced to sweeten debt terms linked to Anthropic contracts after investors pushed back, reflecting lender concern about single-counterparty concentration. Separately, a federal judge has ruled the Trump administration lacks sufficient evidence to label Anthropic a supply-chain risk, per TechCrunch — a ruling that matters because any successful designation would disrupt the government contract revenue underpinning Anthropic's balance sheet and, by extension, the debt structures built around it.

South Korea's plan to inject 20 trillion won ($13.9 billion) into its Korea Investment Corporation for AI, data centers, and infrastructure — with the fund's mandate expanded to domestic assets for the first time — is a direct policy response to the week's tech stock rout, per Bloomberg. This is an announced intention requiring legislative and administrative implementation, not yet committed capital. The strategic logic is dual: stabilise equity markets where Samsung and SK Hynix represent outsized index weight, and ensure Korea captures a share of the AI data center buildout domestically rather than ceding it to US or Chinese operators.

In Europe, the EU has opened a formal call for proposals to support up to seven AI gigafactories under a €10 billion plan, per Wall Street Journal and Reuters. The EU tender is a procurement mechanism, not a grant programme — the gigafactory operators will be commercial entities, but with state-backed demand signals and infrastructure support. Germany's digital minister simultaneously urged accelerated AI self-sufficiency following an incident involving an OpenAI model, per Reuters, adding political urgency to what was already an economic priority.

British neocloud Nscale's acquisition of Anyscale — the Ray framework commercialisation company — is a vertical integration play aimed at owning both raw compute and the orchestration layer that enterprise customers use to distribute AI workloads, per TechCrunch and Reuters. The strategic intent is differentiation from commodity GPU rental: a neocloud that can offer managed distributed training and inference at scale has a defensible moat that pure-play compute providers lack. Terms were not disclosed.

Okta's $200 million acquisition of Permiso targets a structurally different problem: as enterprises deploy AI agents that operate autonomously across cloud environments, traditional identity security — designed for human users — breaks down. Permiso's identity threat detection for non-human identities gives Okta a product wedge into the AI agent security category before it becomes a standard procurement requirement, per TechCrunch. The $200 million figure is sourced from a single unnamed source and has not been confirmed by either party.

Leopold Aschenbrenner's Situational Awareness fund — built on concentrated, leveraged long positions in AI equities — was forced to sell most of its public equity portfolio to Citadel after AI stock losses triggered margin calls, per Reuters and Semafor. The fund retains its Anthropic private equity stake, which cannot be force-sold in the same way. Per FT, the unwind ultimately required a call to Ken Griffin at Citadel, echoing the Archegos dynamic where a single fund's distressed liquidation became a price signal across correlated positions.

The episode illustrates a specific structural vulnerability: funds that use leverage to amplify AI equity exposure face asymmetric margin dynamics when crowded trades reverse simultaneously. Citadel's acquisition of the portfolio at distressed prices is the other side of that trade — a liquidity provider extracting a discount that reflects both market stress and forced-seller dynamics. Jim Cramer's characterisation of it as a 'clearing event' on CNBC may prove correct if no other similarly structured funds surface, but the Semafor comparison to Archegos warrants monitoring of prime brokerage exposure data.