explainx / blog
Ben Werdmuller argues China's open-weights strategy is beating America's closed, proprietary approach. The essay hit Hacker News with 775+ comments — here's the argument, the pushback, and what the data actually shows.
Jul 14, 2026
Claude’s 90-second “There’s hope in hard questions” spot landed Jul 9. Five days later Sam Altman asked if the account was fake — then jabbed silent downgrades. explainx.ai maps the campaign, the roast, and what Anthropic promised.
Jul 13, 2026
@Saboo_Shubham_ reposted OpenAI's official Codex plugin for Claude Code — 28k+ GitHub stars, months old, not a surprise drop. explainx.ai maps /codex:review, /codex:rescue, transfer flows, and why Claude→Codex works but not the reverse.
Jul 12, 2026
@ClaudeDevs and @thsottiaux both reset limits as GPT-5.6 Sol and Fable 5 collided. explainx.ai explains the competitive math — 60 vs 59 benchmark points, 30% Fable cost, and why neither can afford angry developers this week.
An op-ed titled "China's open-weights AI strategy is winning" by independent writer Ben Werdmuller hit 972 points and 775+ comments on Hacker News in July 2026, reigniting a debate that's been simmering since DeepSeek's first releases: is America's closed, proprietary AI strategy — Anthropic, OpenAI, keep-the-weights-locked-down — actually losing to China's push to make frontier-adjacent models free and downloadable? The essay's framing is blunt: locked-down business practices for a technology with no real moat but significant ecosystem benefits is, in Werdmuller's words, "an obviously losing strategy."
The piece landed the same week Kimi K3 (Moonshot AI, 2.8 trillion parameters) and a preview of Qwen 3.8-Max (Alibaba, 2.4 trillion parameters, open weights coming soon) both shipped, and days after Xi Jinping gave a speech explicitly endorsing open source as national AI policy. The timing turned what could have been a niche argument into one of the most-discussed threads on Hacker News in months.
| Claim | Werdmuller's argument | Strongest counter |
|---|---|---|
| Models have no moat | Switching between Claude and ChatGPT is frictionless, especially via API | True for raw capability; false for harness lock-in (Claude Code, Codex) |
| China is winning distribution | Open weights turn a GPU export disadvantage into ecosystem reach | Distribution ≠ profitability; someone still has to fund training |
| 80% of startups use Chinese models | Cited from an Economist quote via a16z's Martin Casado | Casado walked it back to ~16-24% within days |
| Chinese models are cheaper | Lower sticker price on API pricing pages | Ben Thompson: that reflects US supply scarcity, not lower true COGS |
| Distillation gives China a structural edge | Every US frontier release becomes free training signal for Chinese labs | Distillation is arguably how LLMs work generally — hard to selectively ban |
Werdmuller's central claim is that AI models, as standalone products, "have very little moat beyond what amounts to brand loyalty and superficial switching costs." Someone using ChatGPT today can use Claude tomorrow with minimal workflow disruption, particularly for engineering teams accessing models via API rather than a locked-in consumer interface. The real moat, in his framing, sits in the enterprise services layer wrapped around models — contracts, systems integration, and quality-of-life features — not in the underlying weights.
From that premise, he argues China's open-weights strategy is structurally smart for reasons beyond simple generosity:
He points to Alibaba and Moonshot's rapid-fire July 2026 releases and a16z partner Martin Casado's Economist-quoted claim that there's "an 80% chance" any given startup his firm sees is using a Chinese open-source model as evidence the strategy is already working, not just theoretically sound.
That Casado quote became the most contested data point in the entire discussion, and it's worth tracing carefully because it shows how a headline statistic can travel much further than its underlying claim supports. The original Economist line: "I'd say 80% chance [they are] using a Chinese open-source model." Read quickly, that sounds like 80% of startups run on Chinese AI.
Casado himself corrected this on X within days, and the correction matters: "Well, not quite. I'd say 20-30% use open source. Of those I'd say 80% use Chinese based models. So closer to 16-24%." That's a roughly 4x difference between the widely repeated headline number and Casado's own clarified figure. Multiple Hacker News commenters flagged the walk-back as the single most important correction in the thread — one noting it "should really be the #1 comment on this post," since it undermines a load-bearing piece of evidence in the original essay's argument for Chinese-model dominance.
The distinction underneath the correction is also worth noting on its own terms: Casado's revised number describes startups using open-source models at all — a category that includes plenty of internal tooling, evaluation, and dev-time experimentation rather than production-critical model choice. Several commenters made the same point about their own usage: they evaluate Chinese open models constantly, but that doesn't mean those models are what ships to customers. This mirrors the pattern we've tracked in Asia-based models now representing roughly 60% of OpenRouter tokens — real usage, but concentrated in specific segments (cost-sensitive bulk workloads, non-coding classification tasks) rather than a wholesale replacement of frontier models for core engineering work.
The most substantive economic counter-argument came not from the Hacker News thread directly but from Ben Thompson's Stratechery piece the same week, "Who's Afraid of Chinese Models?", which several commenters cross-referenced. Thompson's core correction: open weights are not free to serve, only free to download. Running Kimi K3 or any large model still costs real money in COGS (cost of goods sold) — GPU time, memory, serving infrastructure — and that cost is directly correlated to revenue in a way R&D spend is not.
His argument that Chinese models "seem cheaper" mostly because Anthropic and OpenAI are supply-constrained and charging a price premium reflecting scarce compute, not because Chinese inference is structurally cheaper on a true marginal-cost basis, reframes the entire debate. If demand for frontier intelligence currently exceeds available compute, closed labs can charge above their marginal cost; once that scarcity eases, Thompson expects prices to compress regardless of which country trained the model. That's a meaningfully different read than "China has cracked cheaper AI" — it's closer to "China has cracked cheaper AI right now, under current supply constraints," a distinction that matters for anyone modeling this out past 2026.
A large share of the Hacker News thread circled back to distillation — training a model using another model's outputs as signal — as the mechanism critics say explains Chinese open models catching up so fast without matching US compute or training budgets. The debate split cleanly into two camps that neither side fully dislodged the other from:
The "it's mostly distillation" camp points to Chinese labs' terms-of-service violations querying OpenAI and Anthropic APIs at scale, and argues this compresses the expensive last mile of reaching near-frontier capability without the R&D cost of getting there independently.
The "distillation is how this always worked" camp — and this was the more upvoted position across multiple subthreads — points out that LLMs themselves are a distillation of the open internet, scraped and compressed by the frontier labs in the first place. One heavily upvoted reply put it as: "Stop pilfering what I rightfully stole!" The argument: if training on scraped copyrighted human text is treated as fair use for American labs, it's difficult to coherently object to a Chinese lab training on the outputs of a US model that itself was trained on scraped data.
OpenAI's own head of strategic futures, Dean Ball, complicated the "it's just distillation" dismissal directly, posting about Kimi K3: "It's a very good model! I don't think its performance can be explained away by distillation or anything like that." That's a notable concession from inside a frontier lab, and it undercuts the simplest version of the American-exceptionalism counter-argument — that Chinese labs can only copy, not innovate. We covered the guardrail-access side of this same tension in why US cybersecurity defenders are currently blocked from using American frontier models, which forces some US teams toward Chinese open models specifically because domestic guardrails are more restrictive than the alternative.
One detail sharpened the entire debate mid-thread: a speech attributed to Xi Jinping, reported around July 18, 2026, that explicitly endorsed open weights as strategy: "We should seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing... AI is moving from the digital world into the physical world." That statement — surfaced by multiple commenters citing Stratechery's coverage — reframes Alibaba's decision to resume open-weighting Qwen 3.8-Max (after having stopped releasing weights for leading-edge models earlier in 2026) as coordinated national policy responding to top-level guidance, not an independent commercial calculation by one lab.
That distinction matters for how durable the strategy is. A single company's open-weights bet can reverse when the commercial math stops working — which is precisely what happened with Meta's Llama, cited repeatedly in the thread as a counter-example (more on that below). A state-level policy commitment, backed by a Xi Jinping speech and reinforced across multiple major labs simultaneously (Moonshot, Alibaba, DeepSeek, Zhipu/GLM, Tencent), is a different kind of commitment — closer to how we've covered China's own restrictions on overseas access to its top AI models, which shows the same state-level coordination cutting the other direction.
The most consistently repeated pushback across the thread: Meta already ran this exact experiment with Llama, and it didn't work out especially well for Meta. Llama was, for a stretch, the dominant open-weights model family in the West — and it produced enormous developer mindshare without translating into commensurate commercial success for Meta, whose in-house Llama 4 landed flat and whose original research team has largely departed the company.
Commenters split on what this proves. Some argued it shows open weights are simply a weak business strategy regardless of who runs it. Others — and this reading fits the China context better — argued Llama's problem wasn't openness itself but that Meta never had a clear complementary business the open model was meant to commoditize on Meta's behalf, unlike Nvidia (which benefits from cheap, widely-run models driving GPU demand) or a state actor pursuing manufacturing and robotics adoption rather than direct AI-service revenue. We've tracked the more recent entrant to this specific lane in Thinking Machines Lab's Inkling open-weights MoE release and David Siegel's argument that the open-source AI fight is bigger than software itself — both making versions of the "openness needs a complementary business model to pay for itself" case that the Werdmuller piece mostly skips.
Strip away the geopolitics and the practical guidance converges regardless of which side of the HN thread you find more persuasive: model choice in mid-2026 is genuinely closer to a commodity decision for a growing share of tasks than it was even a year earlier. If you're weighing running open-source models locally against a frontier API, or evaluating Kimi K3 specifically, the honest framing from this debate is: cost comparisons based on published API pricing alone are unreliable in either direction right now, because both US and Chinese pricing currently reflect supply constraints and strategic subsidy more than true marginal cost. Evaluate on task-specific benchmarks and your own inference bill, not on which flag is attached to the lab.
Sources: Ben Werdmuller — "American AI is locked down and proprietary. It's losing." · Hacker News discussion, 972 points · Stratechery — "Who's Afraid of Chinese Models?" · The Verge — Robert Hart on China's AI dominance · Martin Casado's X correction, July 2026.
Statistics, quotes, and comment counts reflect the Hacker News thread and linked sources as of July 21, 2026. Model pricing and availability change frequently — verify current figures against provider documentation before making cost comparisons.
