A retired general’s warning: America can’t fight the AI arms race on tech it doesn’t control

The article argues that the Pentagon’s split with Anthropic highlights a strategic risk: the U.S. buys access to AI capabilities it does not control, while China and partners may deploy more flexible open-source systems at scale. It cites Anthropic’s Mythos model as potentially capable of autonomously finding and weaponizing cybersecurity vulnerabilities, raising concerns about defense and cyber misuse. The piece calls for government-controlled, auditable open-source AI for defense applications.

Analysis

The key market implication is not that defense AI demand disappears, but that procurement is shifting from a software-licensing model to a control-plane model. That favors vendors that can deliver deployable, auditable, on-prem or air-gapped systems with government-owned weights, and it pressures closed-model leaders whose moat depends on restricting downstream use. In other words, the winner set is less about the best benchmark scores and more about who can survive federal security review, export controls, and sovereignty requirements across allied militaries.

Second-order effects likely show up in the infrastructure stack before they show up in headline model vendors. If defense agencies and allies demand controllability, the beneficiaries are secure cloud, edge inference, cyber tooling, and systems integrators that can harden models into mission workflows; the losers are pure-play API providers, especially those with concentrated revenue exposed to policy veto risk. The biggest underappreciated risk is a bifurcation of the AI ecosystem: a fast-moving open-weight defense stack on one side, and a slower, more fragile commercial closed-model stack on the other, which could compress premium valuations for companies whose pricing power depends on model scarcity.

The catalyst window is months, not days: budget reallocation, pilot programs, and framework contracts could move after the next procurement cycle, while the real regime shift takes 12-24 months as allies mirror U.S. requirements. Tail risk cuts both ways: a major cyber incident attributed to uncontrolled model behavior would accelerate sovereign-AI spending, while a high-profile policy accommodation between government and a leading model lab could reverse the immediate negative read-through. Consensus may be overestimating the durability of current private-sector AI oligopolies; once national security becomes the gating function, the market may start valuing adaptability, auditability, and control higher than raw model capability.

Key Decisions for Investors

• Long PLTR / short a basket of closed-model AI beneficiaries on the theme of sovereign control: 3-6 month horizon, looking for procurement share gains as defense buyers prioritize deployability and audit trails over frontier performance.
• Add to defense IT / systems integrators exposure via GD, LMT, or RTX on pullbacks; these names are better positioned to capture the integration and hardening spend if AI budgets shift from model subscriptions to mission systems.
• Consider a relative long CRWD / short high-multiple AI software basket for a 6-12 month horizon: if defense and critical infrastructure demand on-prem, secure, and monitored AI, cybersecurity budget intensity should rise faster than pure model spend.
• For event-driven alpha, buy 6-9 month put spreads on names with concentrated exposure to government AI adoption but weak control/hosting flexibility; the thesis is not model obsolescence, but procurement repricing if policy favors sovereign or open-weight alternatives.
• Monitor for a catalyst trade into any federal AI procurement or defense modernization announcement: pair long MSFT/AMZN cloud infrastructure against short a closed-model pure play if the policy language emphasizes compliant deployment rather than frontier leadership.