The Pentagon cut Anthropic from its AI contract portfolio this week. It signed new deals with other frontier-model providers simultaneously. The criteria separating the selected from the excluded are not published. Anthropic's Constitutional AI framing and its safety documentation are prominent parts of its public-facing posture. Whether that framing worked against it in defense procurement is not stated. The same week, Anthropic approached a $900 billion valuation on the enterprise side. Defense exclusion and commercial dominance running in parallel is not a contradiction. It is a bifurcation. The defense procurement process evaluates deployment velocity, interface compliance, and operational reliability. Safety documentation registers differently there than it does in San Francisco. Beijing's defense and commercial AI analysts track US DoD vendor selections with precision. A frontier lab that clears DoD procurement is categorized differently in PRC assessments than one that does not. Anthropic's exclusion will be entered into that ledger. So will whatever labs replaced it. The capability map and the defense map are beginning to diverge from the same cohort of labs. That divergence, not the valuation number, is the signal worth watching.
Drone strikes on major tech data centers, reported this week, change the physical assumption of this beat. Compute geography has always mattered on these pages. Where the GPUs are sited. What jurisdiction holds the power contracts. Which latency envelope serves which inference workload. That calculus was economic and regulatory. It is now also kinetic. Data centers are large, fixed, and visible. They were built for redundancy against software faults and power disruptions. Physical interdiction by aerial means was not the design baseline. The PRC military has gamed this scenario for longer than most. Taiwan Strait contingency planning has included compute infrastructure targeting for years: undersea cables, terrestrial data centers, border-adjacent GPU clusters. The drone incidents, whoever is behind them, show that the analysis has spread beyond state actors. Hyperscaler physical security teams are recalculating now. Distributed compute architectures, sovereign hosting mandates, and hardened facility standards will move from compliance discussions to procurement requirements. How quickly that repricing happens will shape where the next generation of compute concentration lands. The geography question is no longer abstract.
The question the Pentagon's reshuffle leaves open is not which lab wins the next contract cycle. A $900 billion valuation does not purchase protection from procurement exclusion, and it does not harden a data center against a drone. The infrastructure layer and the capability layer are converging on the same set of physical constraints. Beijing has priced this in for years. The open question is not whether the industry eventually does the same. It is whether it does so before the next facility goes dark.