(Last updated September 15th, 2026)

Most are conflating politics with economics in regard to open weight models.

Primer (Open source vs open weight vs local models):

  1. Open source in the traditionally defined sense (publicly available data, code, weights, training pipelines) doesn't really exist in any meaningful form.
  2. Open weights means you can download it, typically under licenses usually acceded via a cloud provider.
  3. Local is a deployment choice versus relying on cloud providers. Most open-source usage is served the same way closed labs serve their models via API.

The gap closes and reopens:

  1. Open models catch up to closed models faster than the frontier improves. True through instruction tuning, RLHF, now reasoning. This is likely, in part, due to 1) distillation of frontier models being a widely used mechanism and 2) frontier research gets harder over time, sustainable only through large amounts of R&D spend with massive compute budgets.
  2. Open weights models are best viewed as a static snapshot of a historical capabilities. I think its best viewed long term as oscillation versus convergence (and from my experience as an open-source fan for many years, this maps to most consumer and enterprise open source software generally).
  3. Subagent coordination is probably next, and worse for open models than reasoning was because 1) quality gaps compound under fan-out and 2) labs are currently training this behavior in, which requires tons of compute, with proprietary and closed harnesses

Token prices are falling, task costs are not:

  1. Price per token keeps collapsing for fixed performance, seen more closer to the frontier. Operating cost per task at the frontier is rising sharply over the same period.
  2. I think looking at blended-per-token price is close to a legacy measure. Output tokens are most of the cost regardless while models a few points apart can be orders of magnitude apart on cost per task.

Open models win tokens, lose dollars:

  1. Open weights are approaching half of routed token volume, yet a small fraction of the dollars.
  2. Frontier revenue compounded straight through it this pricing dynamic, Anthropic's run-rate went up an order of magnitude over the same time period.
  3. They should be seen as complimentary products versus direct substitutes. Cutting price barely moves usage (aligns with my personal use cases of local models).

For local deployment, it is free weights w/ invisible inference:

  1. Third parties often serve open weights at parity cost wise with the lab's own API.
  2. Every token-level dataset comes from a gateway or router - definitionally the not-self-hosted channel. Every self-hosting dataset is a survey counting companies, not tokens.
  3. So open source is more important than the dollar figures suggest, and nobody can say by how much.

You are paying for control:

  1. The durable advantage is half ownership (fine-tune and hold versions), and half private usage of models against your own proprietary data. Labs recently being accused of training on specific frontier math research make the later important for a very small set o frontier research companies.
  2. "It's free" is not a benefit - and it is usually a cost: Average enterprise GPU utilization is measured in the single digits, and cost scales inversely with utilization.
  3. You're not buying cheaper compute. You're buying the provider's gross margin, and taking their utilization risk with one demand curve instead of thousands.

"American open source" a hardware marketing tactic:

  1. Top open models are currently Chinese. DeepSeek, Qwen, Kimi and GLM sit at the top of the open leaderboards. The strongest American entrant is Nvidia's Nemotron, just behind GLM on the intelligence index.
  2. Nvidia's open-source models are best modeled as a marketing cost, providing tooling to enable less efficient, locally hosted models to push GPU demand. Evidenced further by the summer open-weights letter which was only signed by chipmakers, server vendors and cloud providers.
  3. Chinese labs are beginning to report the economics issues - cloud margins going negative, MiniMax's is collapsing, Moonshot reports margins are lower because the weights are free.

So the us-vs-them framing is an attempt to frame "China is winning open source" as an argument to try to stop any sense of pacing frontier research, and to continue pushing infrastructure spend over than anything else.