Meituan Trains Trillion-Param Model on 50K Domestic Chips, Bypassing US Ban
Key Takeaways
- Meituan's LongCat-2.0, the first trillion-parameter LLM trained entirely on a 50,000-chip domestic cluster, signals a breakthrough for Chinese cloud and SaaS providers.
- The achievement reduces dependence on sanctioned Nvidia chips and enables sovereign AI infrastructure, potentially accelerating AI-powered services across China's digital economy.
Mentioned
Key Intelligence
Key Facts
- 1Meituan launched LongCat-2.0, a trillion-parameter large language model, on June 30, 2026.
- 2It is the first model of its size to complete end-to-end training and inference on a 50,000-chip domestic compute cluster.
- 3Meituan claims performance comparable to Google's Gemini 3.1 pro, released in February 2026.
- 4The company's AI research team began exploring domestic chips in 2023 amid US export restrictions on advanced Nvidia GPUs.
- 5The specific Chinese chipmaker is undisclosed, but the achievement proves domestic silicon can support trillion-parameter training.
- 6Other Chinese AI labs like DeepSeek and Zhipu have used domestic chips for inference but have relied on Nvidia for training.
First trillion-parameter model fully trained and inferenced using only domestic chips
Who's Affected
Analysis
For the SaaS and cloud industry, hardware independence is the holy grail of infrastructure resilience. Today, Meituan has turned that vision into reality by training a cutting-edge trillion-parameter AI model on a massive cluster of 50,000 homegrown chips. This milestone proves that Chinese cloud platforms can now build and deploy advanced AI services without exposure to Western export controls, fundamentally reshaping the competitive landscape for enterprise SaaS and AI-as-a-Service offerings.
On June 30, 2026, Chinese tech giant Meituan unveiled LongCat-2.0, a trillion-parameter large language model that it claims is the first of its scale to be trained and inferenced entirely on a 50,000-chip cluster made of domestically developed silicon. The announcement represents a watershed moment in China's push for semiconductor self-sufficiency and marks a direct challenge to the US-led export controls designed to limit Beijing's access to cutting-edge AI hardware.
Today, Meituan has turned that vision into reality by training a cutting-edge trillion-parameter AI model on a massive cluster of 50,000 homegrown chips.
The achievement comes against a backdrop of intensifying technological rivalry. Since 2023, Washington has reinforced restrictions on exports of advanced Nvidia GPUs—the workhorse of AI training globally—to China, citing national security concerns. In response, China has poured resources into domestic chip development, with companies like Huawei making strides in inference chips, but training trillion-parameter models was widely believed to require Western hardware. Meituan's disclosure that it began exploring domestic chips in 2023 underscores a multi-year strategic pivot that has now yielded a competitive model.
LongCat-2.0's performance is reportedly comparable to Google's Gemini 3.1 pro, released in February 2026. While Meituan did not name the chipmaker behind its 50,000-chip cluster, the feat demonstrates that Chinese fabricators have advanced to a level capable of supporting the most computationally demanding AI workloads. This move could accelerate the decoupling of China's AI ecosystem from US supply chains, reducing both dependency and vulnerability to sanctions.
For the US and its allies, the milestone signals that export controls may be losing efficacy as China's indigenous capabilities mature. For Nvidia, the dominant supplier of AI accelerators, it raises the specter of a permanently shrinking market in China—a region that historically contributed a significant share of its data-center revenue. The company may face increased pressure to navigate compliance while retaining Chinese customers who now have a proven alternative.
Within China, the demonstration of domestic training capacity could catalyze a new wave of AI model development. Other labs, such as DeepSeek and Zhipu, have used Chinese chips for inference but not for full-scale training. Meituan's success may encourage them to follow suit, potentially leading to a cluster of homegrown models that rival Western counterparts. It also boosts the prospects of Chinese cloud providers and AI-as-a-service platforms that can now offer sovereign AI infrastructure free of foreign hardware dependencies.
The lack of disclosure on the chipmaker introduces some uncertainty—the specific chips may be from a well-known player like Huawei or a rising startup—but the sheer scale of the deployment suggests a mature manufacturing ecosystem. The 50,000-chip cluster points to a massive investment in computing power, roughly equivalent to some of the largest AI supercomputers elsewhere, and hints at a broader buildup of China's domestic compute capacity.
What to Watch
Looking ahead, the global AI hardware market could see a bifurcation: Western models still advance on Nvidia's roadmap, while China's closed-loop system grows on its own chip architecture. This may spur accelerated innovation on both sides, but it also risks fragmentation in AI standards and safety protocols. Regulators in Europe and other regions may need to consider whether such technological separation impacts the future interoperability of AI systems.
For the immediate future, all eyes will be on independent benchmarks comparing LongCat-2.0 against Gemini and other models, as well as any evidence of real-world performance. Meituan's stock reaction and any follow-up announcements about cloud services built on the model will further signal the market's confidence in this domestic breakthrough. One thing is clear: the assumption that cutting-edge AI training requires American chips has been decisively challenged.
Sources
Sources
Based on 2 source articles- The Star Online (my)China's Meituan says new AI model trained on domestic chipsJun 30, 2026
- Afplast Updated (in)China's Meituan says new AI model trained on domestic chipsJun 30, 2026
From the Network
VCs eye $12B opportunity as China’s AI chip upstarts crack training
As China’s top AI labs begin training models on domestic silicon instead of Nvidia, the shift creates a huge opening for chip startups, model optimization platforms, and training-as-a-service ventures
AIChinese silicon tackles AI training: 5 models run on local chips, but LLMs lag
The first multimodal model trained entirely on Chinese chips — Zhipu’s GLM-Image — proves domestic hardware can handle serious AI workloads, but large language model pre-training remains elusive. The
Space & DefenseChina's DeepSeek Bypasses US Sanctions to Train AI on Nvidia's Blackwell Chips
How we covered this story
Every story in our saas coverage is assembled from multiple primary sources, cross-referenced for factual consistency, and scored along three independent dimensions: sentiment, operational impact, and source-cluster confidence. Single-source rumors and unverifiable claims do not pass our editorial gate. When a story shows "Verified by N sources" with N≥2, the development is independently corroborated; when N=1, we mark it explicitly so readers can weigh the signal accordingly.
Impact scoring uses a 1-10 scale weighted toward regulatory, financial, and operational consequence rather than coverage volume. A topic that runs in every outlet but moves no real decisions ranks lower than a niche regulatory filing that reshapes how operators in the saas space have to behave. Read our full methodology for the scoring rubric, our glossary for term definitions, and our trends index for the longitudinal view across the beat.
| Signal on this page | What it tells you |
|---|---|
| Verified by N sources | Independent corroboration count. N≥2 is our confidence floor; N=1 is marked explicitly. |
| Impact score (1-10) | Regulatory + financial + operational weight. 8+ signals an experienced-operator action item. |
| Sentiment | Five-tier classification trained on labeled saas-specific corpora. |
| Timeline | Where applicable, the related-events sequence that contextualizes today's development. |