Fact-Checking the DeepSeek Leak: Can Domestic Chips Handle Trillion-Parameter Models?
pstAsiatech · x · 2026-08-02
In response to the widely circulated alleged DeepSeek investor meeting notes, semiconductor experts conducted a deep analysis to assess the true state of China's AI hardware ecosystem.
The article investigates three core claims from the notes:
- Compute Allocation: What does DeepSeek's reported allocation of 16,000 Huawei accelerators actually represent in practical training?
- Compute Myth: Could Meituan really train a 1.6 trillion-parameter model using only 50,000 domestic chips?
- Chip Comparison: How do Huawei’s latest Ascend chips actually compare against NVIDIA’s GB30?
More from Infra
- Baidu open-sources FlashMask: up to 1.55x faster than FlashAttention-4 for long-context training — BaiduResearch · 2026-10-16
- GYRO quant of Qwen3.8-Flash-Next tested: near Q3 quality at much smaller size — PlasticRevenue4601 · 2026-10-16
- RL train-sample systems are frontier AI's hardest infra problem — and a trade secret — brianryhuang · 2026-10-16
- Local tuning of GLM-5.3-Flash on 2×GB10 lifts decode from 49 to 58 tok/s — EAccelerate_42 · 2026-10-16
- 3x DGX Spark cluster runs GLM-5.3-Flash locally behind a proxy, 1.75M tokens/day at 119W — Teknium · 2026-10-16
- Optuna creator Akiba: when 2 hyperparameters cost $20M, what's left for BO/HPO — Stefania_druga · 2026-10-16