Open weights just flipped the script

Most people assume OpenAI and Anthropic quietly own the entire AI economy. One chart just showed the ground moving under them, and it’s not moving the way you’d guess. I came across this breakdown from the creator who dug into the Vercel data, and I couldn’t stop thinking about it.

Here’s the pattern break. The old story: closed models like ChatGPT and Claude eat everything. The new reality the original poster spotted: open weights models (the free-to-download, run-anywhere kind, mostly out of China) are now grabbing the majority of token volume. DeepSeek even passed Anthropic in token share, 25.2% versus 24.5%.

But here’s the twist that makes this so interesting.

🔑 The split nobody expected

Volume and money are going in opposite directions. Even though DeepSeek moves more tokens, the creator points out that Anthropic captures 64.6% of model spend versus DeepSeek’s 2.8%. That’s roughly 23x more dollars. Cheap models win usage. Frontier models win revenue.

Why? Because on the hardest problems, the right answer is worth a fortune. Think high-frequency trading, where a tiny edge is worth billions. The expert Gavin Baker frames it cleanly: frontier tokens are maybe 10 to 25% of volume but 60 to 90% of the economic value.

💡 Old way vs new way

  • Old way: rent intelligence from one or two labs, hand over your data, hope they never compete with you.
  • New way: grab an open weights model, run it yourself or through cheap inference providers, fine-tune it on your own data, and own the output end to end.

The price gap is wild. The creator compares Claude Fable 5 at $50 per million output tokens against DeepSeek V4 Flash at 18 cents. For most everyday tasks, why pay premium?

One sharp caution he flags from Bindu Reddy: judge cost per completed task, not per token. A cheaper model that burns more tokens can cost the same. His example: Kimmy K3 finished a task for 84 cents versus 96 cents for a pricier rival. Closer than the sticker price suggests.

🛠️ Why open weights matter (per the post)

  • Ownership: your data, your context, your output.
  • Bargaining power: dozens of hosting options, not two.
  • Customization: feed it your business knowledge and squeeze more intelligence for the same price.
  • Real companies already do this. Harvey fine-tuned Kimmy K3 into a legal expert that tops benchmarks. Airbnb uses Qwen. Perplexity uses DeepSeek.

The CTO of Thomson Reuters gave a line the creator loved: renting a model is like renting a house. Roof over your head, but zero equity. Build on open weights and you’re compounding real assets.

How to start, based on his advice:

  1. Download an open weights model (Qwen, GLM, or Kimmy are near frontier now).
  2. Run it locally or through a cheap inference provider.
  3. Feed it your own data and fine-tune for your use case.
  4. Benchmark it against your real tasks, not generic tests.

One worry the original poster leaves us with: most top open models come from China. If US companies build on them, and those models get co-designed around Chinese chips, we could drift into a real dependency down the road.

Want the full chart-by-chart walkthrough and the numbers behind the flip? Watch the full video, it’s worth your time.

Scroll to Top