Open Weights Just Beat America's Trillion-Dollar AI Bet

The stat that should scare every AI lab in San Francisco: a16z's Martin Casado dropped this on The Economist and nobody in the replies disputed it. There's an 80% chance that any given startup is using Chinese models. Not "experimenting with." Using. In production.

While OpenAI and Anthropic were busy raising the next round at a number with too many zeroes, Moonshot, Alibaba, and Z.ai gave engineers something more valuable than a better eval score: permission.

This week HN put two related posts on the front page within hours of each other - "China's open-weights strategy is winning" racked up 1,100+ points and 800+ comments, and Ben Thompson's "Who's afraid of Chinese models?" followed close behind. That's not noise. That's an industry admitting it misjudged the game.

The moat nobody built

Here's the thing US labs got wrong: they treated the model as the product. Lock it behind an API, rate-limit it, gate it behind eval-passing safety reviews, and charge premium tokens for the privilege.

China's labs treated the model as a loss leader. Xi Jinping said the quiet part out loud in a policy speech Thompson quotes: push "open source, openness, collaboration and sharing" because the real prize isn't the digital chatbot - it's AI "moving from the digital world into the physical world," where China already has the robotics and manufacturing stack to capture the value.

That's a page straight out of the old Joel Spolsky playbook: commoditize your complements. If the model is free, the thing it enables - hardware, ecosystem lock-in, distribution - becomes the actual business.

Meta tried this domestically with Llama and mostly lost the plot. Moonshot and Alibaba just... did it, at frontier quality, for less money, with no ToS lawyer in the loop.

No, the tokens aren't actually cheaper - read the fine print

Everyone's favorite lazy take is "Chinese models are cheaper, therefore better economics." Thompson's rebuttal is the most important paragraph in the whole discourse and almost nobody is repeating it:

• Kimi K3: $3/million input tokens, $15/million output
• Sol (OpenAI's new flagship): $5/million input, $30/million output

Looks like a 40-50% discount. Except Kimi K3 burns significantly more tokens to do the same job. Tokens aren't a commodity - intelligence-per-dollar is what matters, and nobody's publishing that comparison because it doesn't flatter either side cleanly.

If you're doing procurement based on the sticker price per million tokens, you're doing it wrong. Benchmark on task completion cost, not list price.

# what everyone compares
cost_per_million_tokens(model_a) vs cost_per_million_tokens(model_b)

# what actually matters
(tokens_burned_to_solve_task * price_per_token) for each model

The asymmetry that should actually worry you: distillation

This is the part that's genuinely underrated. Chinese labs train on frontier model outputs - GPT and Claude completions scraped and distilled into their own checkpoints - without much regard for terms of service. Western open-weight players (the Llamas, the Mistrals of the world) mostly do respect those ToS restrictions, because their legal departments make them.

Net effect: one side is compounding on the other's R&D for free, and the side being compounded on is the one enforcing the rules against its own allies. If you're building an open model in the US or EU right now, you are playing with a hand tied behind your back that your Chinese competitor simply doesn't have.

The cybersecurity story that should be a bigger deal than it is

Buried in Thompson's piece is the anecdote that actually matters more than any benchmark chart: during an incident response, Hugging Face's own defenders reached for GLM 5.2 - Z.ai's Chinese open model - because US model guardrails (baked in partly due to Trump-administration restrictions on using frontier US models for offensive/defensive security work) got in the way.

Read that again. The people defending one of the most important open-source AI infrastructure companies on the planet had to route around American models to do their jobs, because American policy made the American tool less useful for defense. That's not a China problem. That's a self-inflicted wound, and it's the strongest part of Thompson's argument for loosening restrictions on Sol/Fable for legitimate security work.

What this means if you're actually shipping product

Stop treating "which foundation model" as a loyalty decision and start treating it like any other vendor call:

• If you need permissionless fine-tuning, on-prem deployment, or to avoid a $/token bill that scales with your traffic - open weights win, full stop, regardless of the flag on the lab's homepage.
• If you're comparing sticker prices between Kimi K3 and Sol/Claude, normalize for tokens-per-task first. The headline price is marketing, not economics.
• If your incident response runbook depends on an LLM and that LLM refuses to help because of a safety policy written for a different threat model - you have a single point of failure. Test that before you need it, not during.
• If you're building an open model yourself, know that you're competing against labs with fewer restrictions on their own training data pipeline. That's not fair, and it's not going to become fair - plan around it.

The US AI industry spent three years arguing about whether open-sourcing frontier weights was reckless. China spent three years shipping them and letting the rest of the world build the distribution network for free. Both things can be true - and only one of them results in 80% of startups running your model.

Sources: werd.io - China's open-weights AI strategy is winning, Stratechery - Who's afraid of Chinese models?