My view is that blockading frontier AI models is economically unsustainable.

The US government’s rationale is to restrict frontier AI access exclusively to domestic companies to preserve American competitiveness. Economically, however, this makes little sense. Put simply, such a policy compresses the margins of OpenAI and Anthropic while inadvertently incentivizing companies worldwide to adopt open models originating from China.

With each new model release, OpenAI and Anthropic have historically raised effective pricing and token consumption. That dynamic is now disrupted by government regulation. Facing pressure from open-weight competitors, both labs must commit enormous capital to maintain their edge. Yet their pricing power is bounded by open models: price tokens too high, and users migrate to open alternatives. To justify premium pricing, OpenAI and Anthropic must deliver capabilities substantially ahead of open competitors to demonstrate clear superiority.

Today, those margins cannot be assured; they hinge on regulatory headwinds. Meanwhile, open-weight models continue to close the gap. A year ago, losing access to ChatGPT or Claude Code would have been deeply concerning. Today, using DeepSeek-V4 or GLM-5.2 yields remarkably close results—open models have firmly crossed the threshold of everyday viability. If open models keep improving under the blockade, more organizations will shift toward them, further eroding OpenAI and Anthropic’s revenue.

Considering both OpenAI and Anthropic target IPOs in 2026–2027, my assessment is that strict AI restrictions will only be enforced in the short term. Regulations will likely ease subsequently to ensure both companies can expand profit margins.

Over the long term, token pricing faces a hard economic ceiling: beyond a certain threshold, users simply refuse to pay, regardless of model capability. The only scenario sustaining monopolistic rents is if either OpenAI or Anthropic reaches AGI first—and deploys a sophon to lock down all open-source technological progress.