Chinese AI developers are racing to scale foundation models beyond one trillion parameters, even as Washington moves to cut off foreign access to the most powerful US models through sweeping export controls — a dynamic that is reshaping the global AI landscape as of June 2026.

The trillion-parameter threshold

Parameters are a primary measure of an AI model's capability, and the shift to trillion-parameter architectures marks a decisive break from the billion-parameter general-purpose models that dominated 2023 and 2024. According to a report by Donghai Securities, Chinese developers are rapidly migrating to trillion-parameter designs featuring million-token context windows and full compatibility with domestic chip stacks — a direct response to restrictions on US-made semiconductors.

In late April, DeepSeek — widely regarded as China's leading AI champion — launched its first trillion-parameter model, V4, marking a significant milestone for the country's homegrown AI ecosystem.

The competitive backdrop

US rivals OpenAI and Anthropic continue to aggressively expand the scale of their top models, both of which are estimated to have trillions of parameters. Late last week, Anthropic suspended global access to its top-tier models Mythos and Fable following export controls imposed by the Trump administration — a move that underscored the depth of the US-China technology divide.

Other major Chinese tech companies have also crossed the trillion-parameter threshold in recent months. Alibaba Group Holding was among the earliest, releasing Qwen-3-Max-Preview in September, while Xiaomi has similarly unveiled its own trillion-parameter offering.

Why it matters

The convergence of US export controls and China's domestic scaling push is forcing Chinese AI firms to build sovereign model stacks — from chips to context windows — at an unprecedented pace. The Donghai Securities report highlights that adaptation to domestic chip stacks is now a core architectural requirement, not an afterthought, reflecting the long-term impact of semiconductor restrictions.

The suspension of Anthropic's Mythos and Fable models for global users signals that export controls are no longer limited to hardware — frontier AI software is now also in scope, raising the stakes for every non-US developer and enterprise relying on American AI platforms.

What's next

With DeepSeek, Alibaba, Xiaomi, and a growing cohort of Chinese AI firms now operating at the trillion-parameter scale, the gap with US frontier models is narrowing — though benchmark parity remains contested. The critical variable to watch is whether Chinese domestic chip stacks can sustain the training and inference demands of next-generation models as export controls tighten further.

Companies most exposed to further policy escalation include Chinese enterprises still dependent on US cloud AI services, as well as global firms caught between competing regulatory regimes.