Werdmuller argues the American AI industry has trapped itself in a proprietary, hosted-only posture while Chinese labs ship downloadable, permissively-licensed frontier models. He frames the situation bluntly — 'American AI is locked down and proprietary. It's losing.' — pointing to the collapse of Western open-weights options (Llama going quiet, Mistral moving behind APIs, Gemma staying a tier below Gemini) as leaving Qwen, DeepSeek, Kimi, and GLM as the only frontier options developers can act
The editorial highlights that in the twelve months since DeepSeek-V3 shipped, Chinese labs released at least fourteen frontier-tier open-weights models with permissive licenses while US frontier labs released zero. This asymmetry — combined with senior engineers describing production workloads quietly migrating off OpenAI and Anthropic APIs onto self-hosted Qwen3, DeepSeek V4, GLM-5, and Kimi — is presented as evidence that the developer-layer battle is effectively over.
The editorial acknowledges the standard counter-argument: GPT-5 and Claude Opus 4.6 remain above any open checkpoint on the messiest agentic and long-horizon tasks, so for the absolute frontier the proprietary labs still lead. However, the piece frames this as missing the point — because most production LLM work doesn't need frontier-tier capability, the quality gap doesn't translate into developer-layer wins.
Ben Werdmuller's post — currently at 1,178 points on Hacker News — makes a blunt argument: the American AI industry has locked itself into a proprietary, hosted-only posture, and it's losing the developer layer to Chinese open-weights releases. The comment thread is unusually one-sided for HN on a geopolitics-adjacent topic, with senior engineers describing production workloads that quietly migrated off OpenAI and Anthropic APIs onto self-hosted Qwen3, DeepSeek V4, GLM-5, and Kimi variants over the last nine months.
The raw scoreboard is hard to argue with. In the twelve months since DeepSeek-V3 shipped in December 2024, Chinese labs have released at least fourteen frontier-tier open-weights models with permissive licenses; US frontier labs have released zero. Meta's Llama line, once the anchor of Western open weights, has effectively gone quiet after the Llama 4 reception, and its license was never fully OSI-compatible anyway. Mistral has moved most of its serious work behind a commercial API. Google's Gemma models remain a tier below the flagship Gemini. That leaves Qwen (Alibaba), DeepSeek, Moonshot's Kimi, Zhipu's GLM, and a rotating cast of smaller Chinese labs as the only source of downloadable, fine-tunable, on-prem-deployable models at the frontier.
Werdmuller's framing — "American AI is locked down and proprietary. It's losing." — landed because it matched what developers were already seeing in their own stacks. The comment section reads less like a debate and more like a group therapy session for engineers who kept trying to justify the API bill.
The usual response to this argument is that hosted API quality still beats open weights on the hardest benchmarks, and for the absolute frontier that remains true — GPT-5 and Claude Opus 4.6 sit above any open checkpoint on the messiest agentic and long-horizon tasks. But that framing misses what actually drives adoption at the developer layer. Most production LLM workloads don't need the frontier — they need something good enough that you can fine-tune, run inside your VPC, and not have rug-pulled by a pricing change or a policy update.
Open weights collapse an entire class of procurement and compliance headaches. You can run them air-gapped for defense and healthcare customers. You can fine-tune them on proprietary data without shipping that data to a third party. You can quantize them down to fit on hardware you already own. You can fork them if the upstream lab pivots. None of that is possible with a hosted API, no matter how good the model is. The Chinese labs figured out — or were forced by export controls into figuring out — that giving away the weights was the fastest way to become the default substrate.
The license terms matter here too. Qwen3 ships under Apache 2.0. DeepSeek V4 is MIT. GLM-5 is Apache 2.0. These are the licenses that let a startup ship a product on Monday without a lawyer meeting on Friday. Llama's community license, by contrast, has a 700M MAU cap and an acceptable-use policy that expands unilaterally. The difference between "you can use this" and "you can use this without reading anything" is enormous at the adoption margin.
The counter-argument that Chinese models come with hidden censorship or backdoors keeps getting weaker as the weights get audited. Independent evals from EleutherAI, HuggingFace, and academic groups have found the political guardrails in Qwen and DeepSeek are shallow and easily fine-tuned out — and importantly, they're inspectable, because the weights are in your possession. A closed API that quietly refuses to discuss a topic is functionally the same problem, except you can't see it happening and can't fix it. "Trust us" is not a security model.
What's actually happening under the hood is a divergence of business models. US frontier labs raised on the premise that model quality is a moat, so they treat weights as trade secrets and monetize inference. Chinese labs — locked out of the top-end Nvidia chips and locked out of Western capital markets — had no path to competing on hosted inference at scale, so they competed on developer distribution instead. Export controls intended to slow Chinese AI have instead pushed Chinese AI into the strategy that actually wins the developer stack.
If you're greenfield in Q3 2026 and picking a base model, the default has genuinely shifted. For coding assistants, Qwen3-Coder-480B is within striking distance of Claude and can run on a single H200 node. For general reasoning at the API layer, DeepSeek V4 is roughly 15x cheaper than GPT-5 per million tokens and lands within 3-5 points on most public benchmarks. For agentic workflows with tool use, GLM-5 has become the quiet favorite of the LangChain and LlamaIndex crowds because its function-calling schema is more forgiving than the OpenAI-compatible endpoints.
The practical migration path most teams are taking: keep a hosted frontier model (Claude or GPT-5) for the top 5% of hardest calls, and route the other 95% to a self-hosted or vLLM-served open checkpoint. This isn't ideology — it's just what the cost math looks like once your API bill crosses about $8k/month. The tooling has finally caught up too; vLLM, SGLang, and TensorRT-LLM all treat Qwen and DeepSeek as first-class citizens, and the Hugging Face ecosystem has normalized quantization down to 4-bit with minimal quality loss.
The risk you're taking on is different, not necessarily larger. You lose OpenAI's uptime SLA and gain the ability to actually see what your model is doing. You lose Anthropic's constitutional AI guarantees and gain the ability to modify them. For regulated industries, defense contractors, and anyone selling to European enterprises with data-residency requirements, that's an unambiguous win.
The most likely US response is not a new open-weights push from OpenAI or Anthropic — those companies have raised too much money on the closed-model thesis to reverse course. It'll come from a second tier: a well-funded Meta reset, a serious Apple entry, or a coalition of hyperscalers underwriting an open frontier lab as an anti-lock-in play. Until that happens, the honest advice for anyone starting a new AI-native product in 2026 is to build on open weights first and treat hosted APIs as a fallback for the hard cases. That's the inverse of the advice from two years ago, and it's the direct consequence of a strategy the US chose and China exploited.
I’m suspicious of some quotes here, “80% of startups using Chinese models,” doesn’t seem quite right to me. I just interviewed at several startups and they were all using the US models. Maybe they have some minor use of Chinese models but the bread-and-butter of most of these businesses model use is
This is a very strange article considering that Llama, the mother of all open-weight models, has led to anything but success for Meta.Also, enterprises don't give a rip if models are open. They care about zero data retention (and sticking with whatever vendor they're already using).This bl
I do think open-weights models are going to "win" in the sense that they're probably going to be dominant when the hardware to run them becomes affordable. (which might be a while). Although I guess you could probably rent the GPU's yourself to hypothetically save on costs. (I&#x
I do not understand the logic going into these companies. Flagrantly violate all IP in Human history, essentially claiming domain over the heritage of Humanity... And... Try to privatize it? When the technology -- and data -- are both public domain to begin with?It is ming-boggling stupidity. If the
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
The lesson of the last 50 years of the computer and software marketplace is that free and low-end eventually wins.- PCs destroyed minicomputers. Mainframes survive, but serving a much tinier portion of the market than they used to.- PC office productivity software destroyed expensive professional pr