Has Chinese AI sensation DeepSeek lost its mojo to rivals? Delays in R2 launch raise big questions amid fierce rivalry

DeepSeek stalls on Huawei chips after months of hype, exposing the fragile state of China’s push for AI self-reliance. The company’s attempt to train its R2 model on Huawei’s Ascend processors collapsed under technical problems, from unstable chip performance to weak software support.

What was meant to showcase independence from U.S. Nvidia hardware instead caused delays and gave rivals like Alibaba’s Qwen3 and Baidu’s Ernie 5.0 space to surge ahead.

What began as a bold showcase of China’s AI independence has instead delayed DeepSeek’s roadmap and handed rivals like Alibaba’s Qwen3 and Baidu’s Ernie 5.0 a golden opportunity to seize market share.


The stumble underscores a larger truth: despite political pressure to rely on homegrown chips, China’s semiconductor ecosystem is not yet ready to carry the weight of its AI ambitions.

DeepSeek rise that shook Silicon Valley

When DeepSeek burst onto the scene in January 2025 with its R1 chatbot, the mood in Beijing was electric. Within 72 hours of launch, the app topped Apple’s App Store charts in both China and the U.S. Analysts hailed it as a “Sputnik moment” for AI, suggesting China had finally built a model powerful enough to rival OpenAI and Anthropic.

ALSO READ: Seven months after stunning the world, China’s DeepSeek AI leans on US technology for critical upgrade

Unlike its Western counterparts, R1 didn’t rely on billion-dollar server farms or the latest Nvidia GPUs. It was lean, resource-efficient, and open source. That formula not only made it cheaper to run but also gave China’s tech giants — from Tencent to Baidu — a shot at scaling AI without bowing to U.S. chip restrictions.

Beijing’s push for hardware independence

Flush with political backing, DeepSeek was nudged to do something bold: train its next-generation R2 model entirely on domestic hardware. Huawei’s Ascend 910B chips were the chosen workhorse, a gamble that played directly into China’s self-reliance agenda.

Officials wanted proof that Chinese AI could thrive without Nvidia’s A100 and H100 GPUs, which remain blacklisted under U.S. export controls. Huawei promised its chips could deliver comparable performance — at least on paper.

Where the promise unraveled

By May 2025, engineers inside DeepSeek’s Hangzhou headquarters were running full-scale training on Ascend clusters. The results were dismal. According to two engineers, interconnect bandwidth between the chips collapsed under heavy workloads.

Training jobs that would normally converge in three weeks on Nvidia hardware stretched beyond two months — before failing altogether.

The software ecosystem proved another weak link. Huawei’s CANN toolkit, meant to rival Nvidia’s CUDA, lacked maturity. Debugging tools crashed mid-run, leaving teams to restart jobs from scratch. “It felt like trying to fly a rocket with half the control panel missing,” one engineer said.

Even after Huawei dispatched its own specialists to assist, the problems persisted. By July, DeepSeek quietly shifted strategy: Nvidia GPUs would resume training duties, while Ascend hardware would be relegated to inference tasks, where speed and precision matter less.

Delays hand rivals an opening

The immediate casualty was timing. DeepSeek’s R2 launch, once slated for May, has now slipped indefinitely. That vacuum has allowed competitors to accelerate.

Alibaba’s Qwen3, launched in June, now dominates the Chinese enterprise market with tighter integration into DingTalk and AliCloud. Baidu’s Ernie 5.0 has rolled out multimodal features that DeepSeek had promised but couldn’t deliver. And in the U.S., OpenAI’s GPT-5 continues to pull ahead with enterprise partnerships.

According to research firm Bernstein, DeepSeek’s market share in China’s enterprise AI sector slipped from 28% in March to 19% by August 2025, while Alibaba’s climbed above 30%. Investors who once treated DeepSeek as Beijing’s flagship AI bet are now hedging across multiple platforms.

What this means for China’s AI ambitions

DeepSeek’s struggles are more than a corporate hiccup. They expose the technical gulf that still exists between China’s semiconductor industry and Nvidia’s dominance.

• Performance gap: Benchmarks suggest Huawei’s Ascend chips lag 30–40% behind Nvidia’s H100s in training throughput.
  
  
• Ecosystem gap: CUDA has 15 years of developer tooling behind it; CANN is barely five years old.
  
  
• Strategic gap: Inference tasks can be localized, but without control of training hardware, the crown jewels of AI — the large models themselves — remain dependent on U.S. technology.
  
  

For Beijing, this raises uncomfortable questions. Can China truly achieve technological sovereignty in AI without solving its chip bottleneck? And how long can companies like DeepSeek absorb the cost of repeated failures before investors demand results?

The road ahead: survival or sidelining?

DeepSeek insists it is not abandoning domestic hardware. In an August statement, the company said it would “continue close collaboration with Huawei and other domestic partners to enhance inference capabilities at scale.” Yet insiders admit the R2 launch is unlikely before early 2026, a delay that could prove fatal in a field moving at breakneck speed.

Meanwhile, rivals are pressing their advantage. Alibaba and Baidu are already pitching overseas clients. If they succeed, DeepSeek risks becoming what it swore it would never be: a one-hit wonder, remembered for the shock of its debut rather than the strength of its legacy.

FAQs:

Q1. Why did DeepSeek stall on Huawei chips?
DeepSeek faced repeated failures training its R2 model due to unstable Ascend chips and poor software support.

Q2. Which rivals are gaining as DeepSeek struggles?
Alibaba’s Qwen3 and Baidu’s Ernie 5.0 are quickly capturing market share during DeepSeek’s delay.