US moves to loosen AI curbs—while OpenAI gates GPT access and China’s Zhipu closes in

Anthropic PBC and the Trump administration are nearing an agreement that would suspend U.S. restrictions on the company’s two leading AI models after weeks of security-focused talks. Bloomberg reports that the negotiations have centered on how to ensure the systems’ safety and security, with both sides moving closer to a deal. In parallel, OpenAI is restricting the initial release of its newest ChatGPT model, GPT-5.6 Sol, to a Trump-approved group during a testing window, effectively turning access into a political and compliance gate. Separate reporting also indicates that the Trump government is deciding who can use a new ChatGPT update, reinforcing that model deployment is being treated as a controlled national capability rather than a purely commercial product. Strategically, the cluster points to a U.S. approach that is simultaneously liberalizing and securitizing: easing certain regulatory constraints for top domestic labs while tightening distribution through government-approved channels. That creates a leverage point for Washington in the global AI race, because model availability, evaluation, and deployment timelines can be shaped by security review rather than market demand alone. The competitive pressure is explicit in the reporting that China’s Zhipu is closing in on top U.S. AI models, with Anthropic and OpenAI allegedly held back—suggesting the U.S. is trying to manage both technological advantage and risk. The likely winners are U.S.-aligned labs that can clear security negotiations faster, while the losers are any competitors—foreign or domestic—that face slower approvals, narrower access, or higher compliance friction. Market implications are likely to ripple through cloud infrastructure, data center capex, and AI compute supply chains as labs adjust release schedules and customer access. UtilityDive’s note that data centers are ready to negotiate flexibility for speed signals that operators may seek faster interconnection, power delivery, and operational levers to meet surging AI workloads tied to new model rollouts. In equities and instruments, the most direct sensitivity is to AI infrastructure and power-related themes, where expectations for near-term demand can move quickly on regulatory headlines. While the articles do not provide numeric price moves, the direction is constructive for AI compute and grid-adjacent services in the U.S., but it also raises volatility for consumer-facing AI platforms due to gated releases that can dampen adoption curves. The next watch items are the specific terms of the Anthropic deal—especially what security conditions, monitoring requirements, and scope of “lifted curbs” apply to the two models. For OpenAI, the key trigger is how the Trump-approved group is defined and whether the testing cohort expands on a fixed timeline or remains discretionary. On the competitive front, investors should monitor credible benchmarks and product capability signals from Zhipu, because any demonstrated leap could pressure the U.S. to accelerate approvals or, conversely, to tighten controls again. Finally, data center operators’ negotiations with utilities and grid operators should be tracked for concrete milestones—such as power availability, latency/throughput commitments, and permitting timelines—that would indicate whether the U.S. can sustain faster deployment without triggering new bottlenecks.

Geopolitical Implications

• 01

  Washington is using security review and access approvals to shape the global AI race, balancing competitiveness with risk management.

• 02

  The U.S. model-release gate can create de facto export-control-like effects without formal export actions, influencing international competitive dynamics.

• 03

  If China’s Zhipu demonstrates rapid capability gains, the U.S. may tighten controls again or accelerate approvals to preserve strategic advantage.

Key Signals

• —Published details of the Anthropic agreement: scope, monitoring, and which model endpoints are covered.
• —Definition and expansion criteria for OpenAI’s Trump-approved testing group and whether access broadens on schedule.
• —Independent benchmark movement for Zhipu relative to Anthropic/OpenAI, especially on safety and reasoning metrics.
• —Concrete data-center power/interconnection milestones that indicate whether regulatory timelines translate into faster compute deployment.