Pharma Bets Big on AI Drugs, Pentagon-Anthropic Battle Escalates

The Pentagon's use of Anthropic's Claude model during Iran hostilities—and subsequent designation of Anthropic as a supply-chain risk—marks the first direct confrontation between a foundation model provider's ethical boundaries and government operational requirements. Anthropic's court challenge will determine whether commercial AI companies can maintain meaningful restrictions on military applications or whether national security imperatives render such policies unenforceable. The Pentagon's position, articulated by AI architect Drew Cukor, treats commercial foundation models as critical infrastructure that cannot be subject to developer-imposed limitations. This directly contradicts venture-backed AI companies' attempts to differentiate on ethical positioning while remaining commercially viable.

The dispute exposes a structural tension in AI market dynamics: companies seeking defence contracts must accept operational requirements that conflict with public safety commitments, while those refusing military work face both revenue loss and potential regulatory retaliation. The regulatory gap allowing third-party resellers to provide banned models through cloud infrastructure demonstrates that ethical restrictions are trivially circumvented, making them commercially costly without achieving stated safety objectives. For investors, the outcome establishes whether AI defence plays carry execution risk beyond technical capability—if providers cannot reliably contract with government customers, valuations predicated on dual-use deployment collapse.

OpenAI's decision to shut down Sora after six months represents the first major product retreat for a foundation model leader and signals that consumer AI monetisation remains fundamentally unsolved. Video generation's compute intensity—orders of magnitude higher than text—created unit economics that could not be justified by revenue or strategic positioning, leading OpenAI to kill the product rather than iterate. The failure contrasts sharply with enterprise adoption, where AI tools are embedding into workflows with measurable productivity gains and customers demonstrating sustained willingness to pay for API access. The divergence suggests foundation model companies' multi-billion-dollar valuations, predicated on platform ubiquity across consumer and enterprise segments, may be concentrated in enterprise sales with limited consumer revenue materialising.

Eli Lilly's $2.75 billion commitment to Insilico Medicine for AI-discovered drugs illustrates the opposite dynamic in enterprise deployments—pharma is willing to pay enormous sums for AI capabilities that demonstrably de-risk R&D pipelines, even without clinical validation. Financial services firms are embedding AI into trading infrastructure to meet investor expectations, as LSEG's David Schwimmer describes, indicating enterprise deployment follows measurable business logic rather than experimental novelty. The bifurcation reveals that AI adoption follows a power law: industries with quantifiable risk-reward trade-offs and regulatory tolerance are deploying at scale, while consumer applications requiring habitual engagement or subjective value judgments remain stuck in pilot purgatory.

Schneider Electric's Olivier Blum describes the industrial giant's challenge in supplying power and cooling systems fast enough to match Nvidia's chip production pace, revealing that physical infrastructure has become the binding constraint on AI deployment rather than semiconductor availability. The shift is forcing hyperscalers to site data centres based on grid capacity rather than optimal network topology, with clean energy commitments colliding with AI infrastructure energy demands. Cambridge researchers' breakthrough on memristor devices operating at million-fold lower switching currents offers a potential pathway around these constraints, but commercial viability remains years away. In the interim, energy efficiency gains must come from industrial systems—power delivery, cooling, thermal management—rather than silicon improvements alone.

The emergence of sophisticated Samsung SSD counterfeits with near-identical performance to authentic units demonstrates how AI demand pressure is driving supply chain integrity risks beyond headline semiconductor components. When performance testing no longer reliably identifies fakes, enterprises face reliability exposure that may not surface until deployed systems fail under production workloads. The Supermicro securities fraud lawsuit, alleging concealed dependence on illegal AI chip exports, further exposes governance gaps across the hardware supply chain as geopolitical compliance becomes a material investment risk. Together, these developments indicate capital is shifting from pure-play AI software toward industrial companies solving power density, thermal management, and supply chain provenance—the physical infrastructure enabling deployment.