China Is Quietly Winning the AI Race — Key Takeaways
Evaluate Chinese open-weight models (DeepSeek, Qwen) for cost-sensitive inference workloads — they run 36x cheaper than GPT-4.5 and now power the majority of global developer traffic, but route API calls through China-based servers subject to state intelligence law.
Key takeaways
DeepSeek V4 Flash costs $0.14/M tokens vs GPT-4.5's $5 — 36x cheaper via 5.5% parameter activation
DeepSeek V4 Flash costs $0.14/M tokens vs GPT-4.5's $5 — 36x cheaper via 5.5% parameter activation
- MoE architecture: 671B total params, only 37B activated per query, cutting compute >90% vs dense models.
- V4 Flash pushes further: 284B total, 13B activated — trained on a $5.6M compute budget for a single run.
Chinese open-weight models are free to self-host, but direct API calls route through China's jurisdiction
Chinese open-weight models are free to self-host, but direct API calls route through China's jurisdiction
- China's 2017 National Intelligence Law requires companies to cooperate with state intelligence on request.
- Lindy's workaround: host DeepSeek on US soil via Atlas Cloud so data never leaves US territory — adds cost but removes legal exposure.
Coinbase switching to Chinese open-weight models cut AI spend 50% and lifted cache hit rate from 5% to 60%
Coinbase switching to Chinese open-weight models cut AI spend 50% and lifted cache hit rate from 5% to 60%
- Used Zhipu GLM-5.2 and Moonshot Kimi; 91% of engineers stopped hitting usage caps entirely.
- Pinterest fine-tuned Alibaba's Qwen on proprietary visual data: 90% cost cut plus 30% accuracy gain on task-specific benchmarks.
This Dig holds the full set of insights, 4 flashcards, and 3 quotes — free in Homestake.
Unlock this Dig freeFree forever · No credit card required
In this video
- 1mThe 12-Month Collapse: Chinese Models Overtake American AI Traffic
- 2mThe Crossover Moment and DeepSeek's Rise
- 4mReal Companies Switching: Lindy, Coinbase, and Pinterest Case Studies
- 7mWhy Chinese Models Are So Cheap: Mixture of Experts Architecture
- 9mThe Export Control Backfire: How US Sanctions Turbocharged Chinese AI
- 10mVolume vs. Value: America's Remaining Lead and the Data Risk
- 12mThe Trillion-Dollar Bet and What It Means for Investors
“By forcing Chinese firms to engineer around hardware limits, the US may have accidentally turbocharged the very competition it was trying to contain.”
— Perplexity CEO
This page is a partial, transformative summary produced by Homestake. All rights to the original content remain with its creator — please support them at the source link above.
