Key Takeaways
- Accusations that Chinese AI labs win purely by scraping Western frontier models ignore architecture: distillation alone cannot bridge major capability gaps.
- US-based Reflection AI possessed massive clusters of Nvidia Blackwell GPUs, yet its flagship model trailed Chinese open-source models by three to six months.
- Chinese open-source releases sit roughly three months behind Western frontier models, driven by algorithmic choices rather than raw compute.
- Western infrastructure providers like Fireworks and Modal now run DeepSeek's open-sourced inference framework to cut GPU serving costs.
- Sparse mixture-of-experts architectures allow Chinese teams to match dense model quality while bypassing Western compute restrictions.
The Disagreement
The standard Silicon Valley narrative treats Chinese model progress as intellectual theft. In this telling, engineers in Shenzhen run scripts against Western APIs, siphon synthetic training data, and copy frontier weights without consequence.
As the Odd Lots hosts put it: “A lot of people will say they're distilling the Western models and cheating, basically, but they're still getting there.” (play from 2:44) The hosts added: “You're like building something out in like Shenzhen, brilliant engineers, I'm sure, but I don't think they care about terms of service violations. I don't think anything is going to bad is going to happen to them if they like find a way to do 100,000, you know, Claude queries surreptitiously.” (play from 32:35)
Daniel Saeedi counters that this story collapses under basic scrutiny. If synthetic distillation from frontier models were enough to create elite models, well-capitalized American startups with domestic API access and top-tier silicon would lead the open-source pack. They do not.
Saeedi points directly to the launch of Reflection AI: “They released their flagship model about a month ago and it was probably about 3 to 6 months behind the Chinese open-source models, which were about at the time three months behind the publicly-released frontier.” (play from 33:10)
Reflection AI had access to massive clusters of Nvidia Blackwell GPUs. Despite having the best hardware on earth, their model lagged behind Chinese teams that operate under strict chip export bans.
Who's Right (and When They're Wrong)
Both sides describe part of reality, but the distillation-only argument mistakes a data source for an engineering system.
Chinese teams do distill. Every serious AI lab on earth uses synthetic data from frontier models for fine-tuning and post-training alignment. Running Claude queries to train smaller task-specific weights is common practice in both San Francisco and Hangzhou. But distillation is a ceiling, not an engine. When you distill a model, you inherit its reasoning paths and its hallucinations. You cannot distill your way into new architectural efficiencies.
Saeedi notes that Silicon Valley researchers privately admit Chinese labs drive real mathematical and systems breakthroughs: “And so, what that basically tells you is that distillation is not the pure reason that the Chinese labs are where they are. They have lots of evidence of coming up with very novel techniques that are actually being used by the frontier labs.” (play from 33:47)
The evidence sits in production code. Constrained by chip restrictions, Chinese labs turned to sparse mixture-of-experts architectures to avoid activating full parameter sets on every token, cutting raw compute requirements. DeepSeek went further by publishing its inference framework openly. That code was promptly adopted by American infrastructure startups like Fireworks and Modal to optimize their own serving stacks.
If you assume your competitors across the Pacific are merely copycats, you will miscalculate your infrastructure strategy. They are solving hardware constraints with better software engineering.
What to Do With This
Audit your model pipeline tomorrow morning. If you spend five figures monthly on proprietary frontier APIs for structured extraction or classification, run a parallel benchmark against Qwen or an open-weight DeepSeek variant hosted on Fireworks. You will likely cut your token bill by 70% without sacrificing output accuracy.