Chinese AI models reach 3 trillion parameters as domestic chip production looms

At the World Artificial Intelligence Conference in July, six Chinese AI labs unveiled models with 2 trillion to 3 trillion parameters, matching the scale of Claude Opus 4.8 while charging a fraction of the price.

"The US moat in building frontier AI software is not as durable as many of us had hoped," said Ryan Fedasiuk, a fellow at the American Enterprise Institute. "We should expect Chinese AI labs to continue distilling and freely releasing a version of the American frontier at a pace just weeks behind US labs."

Moonshot AI's Kimi K3, the most prominent of the new releases, packs 2.8 trillion parameters and a 1 million-token context window — enough to process entire books in a single query. Alibaba's Qwen3.8Max preview, at 2.4 trillion parameters, and DeepSeek's forthcoming V4 official version joined a lineup that also included Z.AI's GLM-5.2, MINIMAX's M3 and H3, and Tencent's HY3. The UK AI Safety Institute found that open-weight models now trail the frontier by four to seven months, down from six to 10 months through most of 2025.

The rapid iteration comes as Huawei prepares mass production of its Ascend 950 series in the second half of 2026, a development CMSI said will become a "significant catalyst" for large-model efficiency by driving down inference costs and improving gross margins for Chinese AI companies. CMSI maintained a "Strongly Recommend" rating on the sector.

Kimi K3's public debut illustrated both the progress and the constraints. Social media users demonstrated the model generating a browser-based version of macOS in minutes, drawing comparisons to frontier US models. Yet Moonshot CEO Yang Zhilin added a line to the release notes acknowledging a "noticeable gap" between K3 and leading US models. On launch day, more than 60 percent of users who logged into Kimi's web portal encountered errors, according to Fedasiuk, because Moonshot lacked sufficient access to high-end computing infrastructure.

"K3 is an extremely large and computationally demanding model," Fedasiuk said. "At nearly 3 trillion parameters, it requires nearly an entire server rack of high-end AI chips to run on."

Domestic Chip Production Targets the Infrastructure Bottleneck

The computing gap is where Huawei's Ascend 950 series enters the picture. Chinese chipmakers including Huawei and Alibaba's T-Head have shifted from competing on single-chip specifications to building "supernodes" that interconnect hundreds to thousands of chips into unified computing pools, using architectural innovation to offset the performance gap of individual processors. CMSI expects mass production of the Ascend 950 in the second half of 2026 to meaningfully reduce inference costs and improve gross margins for model companies.

The competitive field extends beyond infrastructure. CMSI identified Alibaba's full-stack AI capabilities, Tencent's consumer ecosystem advantages, Kuaishou's Kling AI video generation model, Z.AI's leading coding model, and MINIMAX's native full-modality capabilities as key investment themes. API pricing for Chinese models remains markedly below US counterparts, a gap that could widen as domestic chip production scales.

The political backdrop adds another dimension. The Trump administration has loosened chip export controls — allowing Nvidia to sell more chips to China in exchange for a government cut — while simultaneously announcing efforts in April to curb distillation, the practice of training AI models on outputs from US frontier models. The weekend before WAIC, David Sacks, the former AI and crypto czar, criticized top US AI companies that "want the government to eliminate their open source competition," while Emil Michael, a top Pentagon official, called OpenAI's new head of strategic futures a "supreme village idiot" in a dispute over how to respond to the Chinese open-source challenge.

For investors, the calculus is shifting. Alibaba trades at a discount to US cloud peers despite its full-stack AI positioning. Tencent's consumer ecosystem gives it a distribution advantage for AI products that US social platforms lack. And the mass production of domestic chips could compress inference costs by a magnitude that makes Chinese AI services structurally cheaper than US alternatives — a dynamic that threatens the pricing power of OpenAI and Anthropic in global markets.

This article is for informational purposes only and does not constitute investment advice.