//
Discussion · 2026-09-14

Do China's open models still need the frontier labs?

Daniel Newman argues the narrative that Chinese and open models won't slow if frontier labs ‘pace’ themselves…

The Debate

For vs against
Essence
Daniel Newman argues the narrative that Chinese and open models won't slow if frontier labs ‘pace’ themselves makes no sense: today's best open models depend on Anthropic and OpenAI for distillation, so without access to large-scale distillation their training costs would balloon and their ability to ship cheaper leading-edge models would stall. Hardik Desai, Ian Waring and others push back: Chinese labs now have real user data, synthetic-data pipelines and commodity model recipes, so the constraint no longer binds. What's at stake is whether a US-led pacing regime actually slows China at all, or merely cedes the frontier while models commoditize.
For
  • Hardik Desai (verified): 'Your basic premise is technically wrong. The open source labs especially Chinese labs aren't really dependent on Anthropic or OpenAI anymore. ... They have been collecting real data now that their models are being used extensively. They have the data equation solved.'
  • Desai: models are becoming a commodity — 'What will matter is harness and products, that's where you can build the moat.'
  • Ian Waring: DeepSeek and other Chinese labs are 'out innovating the two main Frontier vendors,' closing the gap from below rather than inheriting from above.
  • Moot Point: 'China labs have good enough frontier models now to train and take the lead themselves.'
  • David Drake (partially agreeing with Newman): don't underestimate the talent of the researchers in China — the human capital is there regardless of distillation access.
Against
  • Daniel Newman: 'All the best open models have large dependencies on Anthropic and OpenAI. The pace of all AI development is set by what the frontier does.'
  • Newman: without access to large-scale distillation, 'its training costs would balloon and its ability to deliver more cost efficient versions of leading edge models would stall.'
  • S (@NinBoss727): full agreement — 'Completely agree are you right 100%' — open models ride on frontier distillation.
  • Paperviewr: frontier-adjacent companies 'have no major cash flow compared to Google, meta,' and rising funding costs hit them first — pacing isn't costless.
  • Not Sure (@_PiR2_): partial agreement — it 'will slow a bit, but not stop'; pulling frontier capabilities further from the public reduces the distillation pipeline.

View the thread on X

The Tweets

2 posts
DNDaniel Newman@danielnewmanUV · 3:46 PM · Sep 13, 2026X Post

The narrative that China and Open models will not slow down even if the frontier labs were to slow makes no sense to me. All the best open models have large dependencies on Anthropic and OpenAI. The pace of all AI development is set by what the frontier does and while open models have done interesting work to deliver performance, its training costs would balloon and its ability to deliver more cost efficient versions of leading edge models would stall without access to large scale distillation. So what am I missing here? 🧐

4 Reposts · 49 Likes · View on X →
HDHardik Desai@hardiksdX Post

Your basic premise is technically wrong. The open source labs especially Chinese labs aren't really dependent on Anthropic or OpenAI anymore. The models do basic math, nothing fancy, all that matter is data. They have been collecting real data now that their models are being used extensively. They have the data equation solved, they also now that good models for generating synthetic data. IMO, models will become commodity. What will matter is harness and products, that's where you can build the moat.

2 Likes · View on X →
On this page
01 / The Debate02 / The Tweets