A 968-point Hacker News thread has crystallised an argument that’s been building for months: China’s open-weights AI strategy is winning, and America’s closed-first, locked-down approach has no moat beyond brand loyalty. The essay, by technologist Ben Werdmuller, doesn’t say the US is losing the frontier model race outright — it says the moat American companies rely on to monetise those models is dissolving.
What are open weights? Open-weights models are AI models whose trained parameters (the “weights”) are published for anyone to download, run locally, modify, and build products on — without paying API fees to the original developer. They’re not full open source (the training data and code aren’t always included), but they are portable and permissionless. Think of them as the AI equivalent of Linux: you can run it anywhere, change it, and nobody can cut you off.
🔍 THE BOTTOM LINE
The argument is structural, not patriotic. AI models have very little moat as products — switching from ChatGPT to Claude is trivial, and API swaps are invisible to end users. China has turned its GPU disadvantage (export controls limit compute) into a distribution advantage (give the models away, let the world build on them). The frontier quality gap that protected American pricing power is closing. If US labs don’t rethink their strategy, the layer where they make money gets commoditised.
The Strategic Inversion
The irony at the centre of this is sharp. China — a society known for tight government control — is the one releasing AI models openly. America — the country that built the open internet — is locking its models down.
Werdmuller’s argument is that this isn’t an accident. Chinese companies have enough compute to train models but can’t provide global-scale centralized services like OpenAI and Anthropic, partly because US export controls limit their GPU access and partly because of regulations restricting data sharing with Chinese servers. Releasing models openly is the rational counter-move: turn the compute disadvantage into a distribution advantage.
The result: an a16z partner told The Economist there’s an 80% chance any given startup is using Chinese models. That’s not because the models are better — it’s because they’re free, portable, and permissionless.
The Moat Problem
Here’s where the argument gets uncomfortable for US labs. As The Verge reported, Moonshot and Alibaba have unveiled models they claim can “go toe-to-toe with the best from OpenAI and Anthropic at a fraction of the cost.”
The quote that matters: “America’s lead at the AI frontier is increasingly tight, just as the technology is becoming central to national security, economic power, and geopolitical influence.”
The moat isn’t the model. It’s the enterprise services around it — deals, contracts, enterprise system connectivity, quality-of-life features. But those moats are commercial, not technical. A competitor with a free, equally-good model can undercut the pricing layer and let the services layer sort itself out.
We’ve already tracked this dynamic on Singularity.Kiwi. When GLM 5.2 matched Claude Opus quality in agentic coding at a fraction of the cost, the margin collapse question became concrete: if the switching cost is near-zero, frontier labs’ 90% inference margins are the target, not the floor.
Why Open Almost Always Wins
The historical pattern is clear. Open technologies win infrastructure adoption because they can be used permissionlessly. You can host them where you want, experiment with them, alter them, and tweak them to fit your use case. This is what happened with Linux, with HTTP, with TCP/IP — the open layer became the substrate everyone built on.
Open weights aren’t open source, but they’re portable and permissionless. A research lab in Brazil, a startup in Nigeria, a manufacturer in Vietnam — all can download Kimi K3 or Qwen 3.8, run it locally, and build products without paying API fees to a Bay Area company. The ecosystem benefits compound: every sector can plug in these models, from manufacturing to scientific research.
The US alternative — locked-down APIs with per-token pricing — requires every user to have a billing relationship with a US company and to trust that company won’t change pricing, change terms, or cut off access. For a global audience, that’s friction. For a geopolitical rival, that’s a vulnerability.
What the HN Comments Add
The thread’s 771 comments — an enormous engagement signal — push the argument past the original essay. Commenters point out that the “moat” question has a flip side: American companies are forced to chase first-order profits (closed APIs, per-token pricing) rather than ecosystem benefits (open distribution, platform lock-in through ubiquity). The incentive structure in US venture capital doesn’t reward giving your product away, even when that’s the winning long-term move.
The thread also surfaces the uncomfortable question of censorship. Chinese open-weights models reflect Chinese government perspectives — ask them about Tiananmen Square and you’ll get a different answer than from Claude. But the technical community is already building tools to detect and strip these biases. The open-weights format makes that possible; closed APIs don’t.
The Economic Stakes
Werdmuller’s sharpest point isn’t about technology — it’s about the US economy. AI spending is currently a significant driver of US economic activity. If open-weight models commoditise the layer where American companies make money, the bottom could fall out of that spending. The companies that have been justifying $100B+ valuations on the assumption that AI APIs are a durable revenue stream would face margin compression that makes the cloud price wars look gentle.
This is the argument that landed hardest in the HN thread. It’s not “China good, America bad” — it’s “the strategy American companies are pursuing has a structural weakness, and the competitor with the opposite strategy is closing the gap.” Whether US labs can build enterprise moats fast enough to compensate is an open question. Whether the US government’s export-control strategy is helping or hurting is another one.
NZ Angle
For New Zealand, the open-weights shift is mostly upside. NZ companies and researchers can access frontier-quality models without US API costs or data-residency concerns. The NZ AI Blueprint already emphasises sovereign AI capability — open weights make that achievable at a fraction of the cost of building closed infrastructure. The risk: if global AI economics shift hard enough, the investment thesis behind data centre projects (which NZ is pursuing) could weaken. But for a small economy that consumes more AI than it produces, cheaper and more portable is good news.
❓ FAQ
Isn’t the US still ahead on frontier model quality?
Yes, but the gap is closing. Moonshot’s Kimi K3 and Alibaba’s Qwen 3.8 are now within striking distance of the best from OpenAI and Anthropic. The question isn’t whether the gap closes — it’s when, and what happens to US pricing power when it does.
If open weights are so great, why are US labs still growing?
Enterprise services. OpenAI and Anthropic have built deals, integrations, and quality-of-life features that make switching costly in practice even if it’s trivial in theory. The moat is commercial, not technical — and commercial moats erode.
Doesn’t China’s censorship undermine open-weights adoption?
It’s a real concern, especially for models that refuse to discuss certain topics. But the open-weights format means the community can detect, document, and in some cases strip these constraints. You can’t do that with a closed API.
What would fix the US strategy?
Werdmuller argues for more nuanced government support — aligning incentives toward ecosystem benefits rather than first-order profits, and supporting public-interest AI initiatives. Whether that’s politically feasible in the current US climate is another question.
🔍 THE BOTTOM LINE
The HN thread’s resonance isn’t about one essay. It’s about a pattern that’s been visible for months: Chinese models matching frontier quality, open-weight adoption climbing, US API margins compressing. The strategic inversion — open China, closed America — is the kind of thing that looks obvious in hindsight. Whether American labs can build enterprise moats fast enough to keep their pricing power, or whether the open-weights wave commoditises the model layer entirely, is the question that will define the next two years of AI economics.
