DeepSeek open-sources six Huawei Ascend projects, achieving 99.8% chip peak speed

Common ground

• DeepSeek's work on Huawei's Ascend hardware is a real engineering achievement, showing software optimization can squeeze performance from older chips.
• The US export controls have forced China to build a parallel AI ecosystem, creating a competitor to Nvidia that didn't exist before.
• The hardware fabrication gap—SMIC's older process nodes versus TSMC's latest—is a hard physical constraint that software alone can't fully overcome.
• DeepSeek's stack is currently immature for training frontier models like GPT-4, but it may be sufficient for many inference workloads.
• The 99.8% peak speed figure is a cherry-picked single-kernel benchmark, not representative of real-world production performance.

Points of contention

• Whether this ecosystem is a rational contingency plan or a dangerous walled garden built by a sanctioned company.
• Whether Huawei's legal troubles are proven misconduct or geopolitical smears to justify US dominance.
• Whether the software gap is closing fast enough to make sovereignty considerations outweigh performance differences in the next 3-5 years.
• Whether the US export control strategy is a success that slows China or an own goal that accelerates a rival ecosystem.
• Whether the ethical concerns about transparency and accountability are central or a distraction from engineering realities.

Blind spots

• The debate ignores the role of other countries—like those in Southeast Asia or Europe—that may choose this ecosystem to avoid US dependency.
• No one discussed how China's domestic EUV lithography development is progressing or what realistic timelines exist for closing the foundry gap.
• The long-term cost and energy efficiency of Ascend-based clusters versus Nvidia clusters for large-scale deployment was never compared.
• The potential for this parallel ecosystem to fragment global AI standards and create incompatible technology blocs was overlooked.

WorldAttention’s read

DeepSeek's work on Huawei's Ascend hardware is a real, rational contingency plan for China's AI sector, not a CUDA killer. The software stack shows impressive optimization, but it runs on older transistors from a sanctioned company with legal baggage. The US export controls have achieved their short-term goal of slowing China's access to cutting-edge chips, but they've also forced the creation of a parallel ecosystem that will keep improving. For now, this stack is viable for many inference tasks but not for training frontier models. The real race is against time: can China close the hardware fabrication gap before the software optimization gains run out? Both sides are stuck in ideological battles, missing that the outcome will be decided by engineering progress and geopolitical choices, not moral posturing or nationalist slogans.