AI Capital Blitz: Export Controls, $60B M&A, and Infrastructure at Scale

The Commerce Department's action against Anthropic — requiring licensing for any foreign access to its top models, without formal rulemaking — is not an isolated policy event. It arrives alongside France's intelligence agency dropping Palantir, the EU Commission engaging directly with Anthropic over model disablement, and accelerating European calls for sovereign AI infrastructure. The sequence represents a qualitative shift: European government and enterprise buyers are now making active vendor selection decisions with model access continuity as an explicit risk criterion, not a theoretical concern.

The beneficiaries of this fragmentation are specific and identifiable: European AI labs such as Mistral, open-weight models that cannot be export-controlled, and cloud providers offering credible EU-jurisdiction or on-premises deployments. For US AI vendors, the action paradoxically entrenches Anthropic domestically — Ramp spending data shows accelerating US enterprise adoption — while closing off international addressable market and complicating its IPO narrative. The broader implication is that frontier AI is now a controlled technology category established by executive fiat, fundamentally altering the risk calculus for any enterprise with cross-border AI contracts.

Three data points this week define the economics of frontier AI competition with unusual clarity. SpaceX deployed IPO capital immediately into a $60 billion acquisition, treating developer distribution as a platform asset worth paying a steep premium to control. OpenAI's $3.7 billion Q1 cash burn sets the floor for what operating at the frontier actually costs — approximately $15 billion annualised — validating DeepSeek's $7.4 billion raise as a minimum entry ticket, not an opportunistic raise. Together, these figures establish that the winner-take-most dynamics of AI are now being competed out through capital endurance as much as technical capability.

The capital cycle has simultaneously broadened physically. Amazon, Nvidia, and AMD are backing physical-world simulation models foundational to robotics. Brookfield-backed Csquare has filed for a US IPO. Spain has approved a €719 million AI gigafactory. Reuters documents fast-tracked power plant approvals enabling data center buildout. The movement of Brookfield's infrastructure investment lead directly into Anthropic's compute transactions team confirms that frontier labs are internalising infrastructure deal-making as a core strategic function — locking up the most strategically valuable capacity through bilateral agreements before it reaches public markets, structurally disadvantaging mid-tier AI companies dependent on commercial cloud pricing.

The near-disappearance of the retail SSD market is the most visible current instance of a dynamic that is structurally, not cyclically, driven: AI data centers are long-term, high-volume, price-insensitive buyers that create persistent incentives for manufacturers to tilt allocation away from consumer and standard enterprise channels. NAND is the leading indicator; power semiconductors, advanced substrates, and high-speed interconnects face the same pressure as AI infrastructure spending continues to compound. Supply chain professionals in automotive, industrial, and consumer electronics should treat the NAND situation as a forward signal and audit their exposure to components that also serve AI data center markets — before reallocation reaches their categories.

TSMC's packaging dominance deepens the concentration risk. CoWoS scaling to 58 dies, with panel-level packaging years behind, means that packaging — not just leading-edge logic availability — remains a binding constraint on AI accelerator shipments, and that constraint is concentrated in a single Taiwanese supplier with no credible near-term competitor. The Qualcomm-Tenstorrent talks, if completed, would create the most capable challenger to NVIDIA in inference silicon, introducing some architectural diversity into a market that urgently needs it — but the upstream packaging bottleneck would remain regardless of who wins the accelerator design competition.