Argues that frontier American labs like OpenAI and Anthropic have stopped competing on capability per dollar and instead engineer scarcity through rate limits and tiered subscriptions. Frames the $200/month 'unlimited' Pro tier as a transactional creator-economy relationship — an awkward landing point for an industry that promised AGI.
Points to Qwen3.7-Max, DeepSeek, Moonshot, and Zhipu as evidence that permissively-licensed open weights are landing within single-digit percentage points of GPT and Claude on benchmarks. The decisive differentiator isn't winning MMLU — it's zero download cost, zero inference cost beyond your own hardware, and no rate limits.
Argue Anthropic and OpenAI absorb the enormous R&D cost of frontier capability research, and that Chinese labs benefit from that groundwork while shipping derivatives at zero marginal cost. Without the closed labs funding the research treadmill, the open-weight cadence wouldn't exist.
Counter that 'free-riding' has always been how open source works — building on prior art is the point, not a moral failing. The complaint sounds less like a principled argument about research economics and more like incumbents rationalizing why their paywall is necessary.
A blog post titled *The OnlyFans Economy of American AI* hit the Hacker News front page this week (score: 106 and climbing), pegged to the release of Alibaba's Qwen3.7-Max. The premise is blunt: American frontier labs have stopped competing on capability per dollar and started competing on subscription-tier design. Meanwhile, Chinese labs — Alibaba's Qwen team, DeepSeek, Moonshot, Zhipu — keep shipping competitive open weights with permissive licenses, and they do it on a release cadence that makes the closed labs look slow.
The Qwen3.7-Max release follows the now-familiar pattern: weights on Hugging Face, paper on arXiv, an Apache-style license, and benchmark numbers that land within a single-digit percentage of GPT and Claude on most evals. The differentiator isn't that it wins on MMLU. The differentiator is that it costs zero dollars to download, zero dollars to run on your own H100s, and carries zero rate limits beyond physics.
The HN thread, predictably, fractured. One camp argued the comparison is unfair — Anthropic and OpenAI are absorbing the cost of capability research that Chinese labs free-ride on. The other camp pointed out that 'free-riding' is what open source has always been, and that the moralizing sounds suspiciously like a business model defending itself.
The OnlyFans metaphor is sharper than it sounds. Strip away the spice and what it describes is a creator-economy structure: a tiered subscription wall, a posted price for premium access, scarcity engineered through rate limits rather than capability gaps, and a relationship with the customer that is fundamentally transactional per interaction. That is now the dominant American AI business model, and it is a strange place to land for an industry that spent three years insisting it was building general intelligence.
Compare the unit economics. A senior engineer using Claude Code or ChatGPT Pro pays roughly $200/month for 'unlimited' usage that is, in practice, rate-limited. A team of ten pays $2,000/month. That same team could rent a single 8xH100 node for around $20-25/hour on-demand, or considerably less reserved, and serve Qwen3.7-Max at throughput numbers that make the per-token math collapse in their favor — assuming they have one engineer willing to babysit vLLM. The break-even point where self-hosting an open model beats a closed-API subscription has moved from 'never' in 2023 to 'a small team' in 2026.
The HN commenters who insist Chinese labs are 'just catching up' are reading the wrong scoreboard. Catching up was 2024. The current dynamic is commoditization — the deliberate destruction of margin in a layer of the stack you don't intend to capture. Alibaba doesn't need Qwen to generate API revenue. It needs Qwen to make AWS Bedrock and Azure OpenAI look expensive, while Alibaba Cloud sells the inference capacity underneath. DeepSeek doesn't need to monetize the model; it needs the model to exist as proof its quant fund's infra investments are sane. This is the playbook Google ran with Android against iOS, and Meta ran with PyTorch against TensorFlow. The American labs are now on the receiving end of it.
The community reaction reveals the discomfort. One top HN comment, paraphrased: 'We told ourselves the moat was capability. The moat was distribution.' Another: 'OpenAI's real product isn't GPT, it's the ChatGPT consumer brand. Everything else is a wrapper.' Both are partially right and both miss the same thing — distribution and brand are durable advantages only if the underlying commodity stays scarce. Open weights at frontier-adjacent quality is a structural problem for any business whose pricing assumes the model is the moat.
The practical implication isn't 'switch everything to Qwen tomorrow.' It's that your architecture should treat the model provider as a swappable component, the same way you'd treat a database driver. If your code calls `anthropic.messages.create()` in 400 places, you have a vendor lock-in problem the same shape as 2014's `mongodb.find()` problem. The fix is the same: a thin internal abstraction, a feature flag, and the discipline to actually test the alternate path quarterly.
For procurement and finance: the negotiating leverage you have with Anthropic, OpenAI, and Google in 2026 is materially different than in 2024, and the people writing the contracts may not realize it yet. A credible self-hosting threat — even one you don't intend to execute — is now a real BATNA. Qwen-class open weights mean 'we'll move to a Chinese open model on our own infra' is no longer a fantasy line in a vendor call. Some procurement teams are already using it. Others are about to discover their three-year enterprise commit was priced for a world that ended last quarter.
For anyone shipping AI features into products: the bifurcation matters less than you'd think. Most production workloads — classification, extraction, structured output, RAG — work fine on a model two notches below frontier. The interesting engineering question of 2026 is not 'which model is smartest' but 'what's the smallest model that holds my eval suite,' because that answer determines whether your COGS line is sustainable.
The end state isn't 'closed labs lose.' It's that the closed labs migrate up the stack — into agents, into vertical products, into the consumer subscription bundle — while the foundation-model layer becomes a low-margin commodity served by whoever has the cheapest GPUs and the most patient capital. Anthropic and OpenAI seem to have already figured this out; Claude Code and ChatGPT's agentic features are the visible bet. The lab that still thinks it's selling intelligence-by-the-token in 2027 will be the one that didn't read the room.
> and will judge, like any sane person, that US frontier models have stopped earning their multiplierI think that this is on the money, although I'd place the bar even lower - DeepSeek v4 Flash is sufficient for basically all day-to-day coding tasks.You might want something beefier for a com
I don't get the point. That Anthropic or OpenAI have more expensive products than Alibaba? So does Apple, AWS/GCP, and pretty much any other large western company vs its Chinese counterparts.
There's a lot of assumptions in here and reductivism of the paid plans to just the models. If that's your idea of how you want to use the API sure that's a reasonable mental financial model but if you want automatic integration with third party systems the cost of the "premium&qu
How many words did I read in this article before I realized it wasn't written by AI? 10? 20? A paragraph or two?It's quite strange that it's very easy to detect AI in writing.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Most American companies (regulated ones, definitely) can't dare to touch any Chinese models, though they knew that it makes perfect economic sense. Until the taboo prevails, the cartel get's their flood of profit. That's a cartel protected by regulations.