When AI stumbles over its own circuits | A compact guide to large language models pdf 2021 | How was chatgpt created | Llm training dataset github | Turtles AI
DeepSeek has postponed the release of its R2 model after encountering technical hurdles while attempting to train it on Huawei’s Ascend chips. The limitations arose from unstable performance, immature software, and slow inter-chip connections, forcing the company to turn to Nvidia.
Key Points:
- R2 model originally scheduled for May delayed
- Training on Ascend chips failed despite Huawei’s intervention
- Training performed with Nvidia GPUs, inference maintained on Ascend
- Structural issues: instability, slowness, and immature software
Stirring like a sleepy sloth after a run, the launch of DeepSeek’s R2 language model stalled, betrayed by domestic chips that promised punch but showed weaknesses: shaky stability, data chains as slow as coffee shop chatter, and software toolkits less sharp than Nvidia’s CUDA. The young Chinese startup, which had caused a stir in January with its R1 processor trained on the power of Nvidia GPUs, was encouraged—or rather, pressured—by Beijing authorities to turn around and switch to Huawei’s Ascend processors; a kind of catwalk-imposed makeover, which ultimately led to a crisis. Despite a team of Huawei engineers being installed on-site with cables and ready-to-use scripts, the long-awaited Ascend training never materialized into a successful run. Consequently, DeepSeek had to return to Nvidia’s good old green mantle to train R2, relegating the Chinese chips to the role of inference, essentially fashion launches after the runway. These inconveniences, heavier than a backpack full of bricks, pushed the release timeline from May to a still-unclear date, causing it to lose precious ground in the race among competitors. CEO Liang Wenfeng, visibly dissatisfied with the progress, has imposed a creative reset: more time and greater ambition to sculpt a model worthy of maintaining leadership in the AI landscape. To make matters worse, labeling data has proven to be a more time-consuming task than interconnecting transistors, further delaying its release. In a market where Alibaba is showcasing Qwen3, more flexible and efficient than DeepSeek’s founding principles, technological tensions between China and the United States are escalating. Nvidia, not yet banished from the scene, continues to negotiate, donating a portion of its Chinese proceeds to the US Treasury to resume sales of H20. An ironic paradox: while the government pushes for independence, the beating heart of Chinese AI still beats to Nvidia’s beat.
The delay of R2 is still a work in progress, and its release remains a puzzle between hope and engineering.


