Frontier AI models cost money to train. Someone has to pay for the compute, the researchers, the data. For the past year and a half, a puzzle has hung over the industry: why were Chinese labs releasing frontier-tier open weights for free? I’ve been one of the beneficiaries, running Qwen and DeepSeek locally on my Mac Studio for work where I want privacy or throughput.

This month the puzzle started resolving. The short answer is that they aren’t anymore.

Alibaba released Qwen3.6-Plus as a closed hosted offering on Alibaba Cloud, keeping only its smaller models (35B and below) as open weights, now positioned as a developer acquisition funnel rather than a frontier release. The Information, ‘Alibaba Becomes Selective with Open-Source Models, New Release Shows,’ April 2026. See also SCMP, ‘Chinese AI giants pivot toward proprietary models.’ Z.ai rolled out GLM-5-Turbo closed. VentureBeat, ‘Z.ai debuts faster, cheaper GLM-5-Turbo model for agents.’ DeepSeek, famously self-funded by the hedge fund High-Flyer Capital and known for turning down outside money, is now seeking \$300M or more at a \$10B valuation, citing researcher departures. The Information, ‘China’s DeepSeek is Raising Money for First Time,’ April 17, 2026. ByteDance’s Seedance 2.0 and Kuaishou’s Kling 3.0 are both proprietary from the start.

A recent ChinaTalk analysis lays out the underlying economics. Chinese labs need revenue, and the DeepSeek shock got open source “a moment” rather than a sustainable business model. The funding environment for Chinese AI is orders of magnitude smaller than America’s. Gulf capital put roughly \$100M into Chinese labs while pouring roughly $15B into Anthropic and OpenAI, and Western venture money is almost exclusively American. Figures from the ChinaTalk analysis cited above. For corroboration on the scale of Gulf AI investment in US labs, see Bloomberg, ‘Abu Dhabi’s MGX Targets \$100 Billion in AI With OpenAI, Anthropic Investments.’ US export controls have tightened steadily, narrowing access to the chips Chinese labs need to train at frontier scale, which means every training run costs more than it does for a US competitor. The Chinese government has been willing to subsidize domestic hardware but not open model development, and domestic chips still trail NVIDIA by a meaningful gap.

Several other pressures compound the squeeze. American labs have documented large-scale adversarial distillation from Chinese labs going back to 2024 and are actively closing that channel through the Frontier Model Forum. Every Chinese model still has to pass a post-training alignment layer for CCP compliance, a tax Western labs don’t pay. And Alibaba’s own Qwen technical lead recently put the odds of a Chinese firm surpassing US tech giants in three to five years at under 20 percent. SCMP, ‘China AI has less than 20% chance to exceed US over next 3 to 5 years: Alibaba scientist,’ January 2026. Lin Junyang has since stepped down from the Qwen team.

Not every Chinese lab is making the same bet. This week Moonshot AI open-sourced Kimi K2.6. Moonshot AI, Kimi K2.6 release, April 20, 2026. See also MarkTechPost, ‘Moonshot AI Releases Kimi K2.6 with Long-Horizon Coding, Agent Swarm Scaling to 300 Sub-Agents and 4,000 Coordinated Steps.’ Moonshot is well capitalized, with significant backing from Alibaba and Tencent, but what makes it different is business model. Kimi runs as a consumer subscription and enterprise product, with a cloud service hosting its OpenClaw agent on top. Open weights are marketing for a business that sells something else. Whether that survives sustained margin pressure from closed Chinese competitors is a separate question, but as of this week it is a genuine third path, not a niche.

The open-weight ecosystem itself will survive. NVIDIA has committed \$26 billion over five years to the Nemotron family. Disclosed in an NVIDIA SEC filing March 11, 2026, first reported by WIRED. See also Nemotron, Wikipedia and Decrypt, ‘Nvidia Drops Nemotron 3 Super Amid \$26 Billion Open-Model AI Bet.’ Meta continues Llama. Google has Gemma. The Allen Institute keeps pushing OLMo and Tülu forward. But the providers are changing, and the incentive is changing with them. The generous era of Chinese frontier labs releasing weights for mindshare is ending. What replaces it is American hardware and platform companies using open models to sell chips and cloud services, and Chinese labs like Moonshot using open weights as the foundation for a wrapped business. The models I run locally today will keep working. The next generation of frontier-tier open weights will increasingly come from a different set of addresses.

This article was developed with AI assistance for research, outlining, drafting, and editing. All ideas, experiences, and perspectives are my own.