Thompson argues the H100/Blackwell export controls backfired: denied raw FLOPs, Chinese labs poured engineering into MoE routing, KV-cache reuse, and speculative decoding, and that serving-stack optimization is now the durable moat. He cites DeepSeek-class models running at 5-10x lower cost per million tokens than GPT or Claude equivalents on the same benchmarks.
Thompson's sharper claim is that this is not a catch-up story but a structural one: Anthropic and OpenAI cannot publish weights or race to the efficiency floor without cannibalizing the API revenue that funds them. Meta has visibly slowed on frontier training and Google keeps Gemini closed, leaving Chinese labs — who treat weights as marketing for cloud or as a side effect — as the only players with aligned incentives to push open-weight serving costs down.
By surfacing the piece to 441 points on HN, the submitter foregrounds Thompson's roll-call of DeepSeek V3/R1, Qwen 3, Kimi K2, and GLM 4.6 as evidence that this is a broad-based capability shift. The framing treats the plurality of Chinese labs shipping weights at frontier tiers as the newsworthy fact, not any single release.
Ben Thompson's latest Stratechery piece, "Who's afraid of Chinese models?", crystallizes a shift most infra engineers have felt for months but few product leaders have priced in: the frontier of open-weight models is now a Chinese story, and the price-per-token gap versus US-hosted closed models is no longer a rounding error.
The piece walks through the current lineup — DeepSeek's V3 and R1 line, Alibaba's Qwen 3 family, Moonshot's Kimi K2, Zhipu's GLM 4.6 — and notes that all four labs are shipping weights, not just APIs, at capability tiers that a year ago would have been Anthropic-and-OpenAI-only territory. The numbers Thompson cites are the ones that matter to anyone signing an inference bill: DeepSeek-class models serve at roughly 5–10x less cost per million tokens than GPT-class or Claude-class equivalents on the same benchmarks, and the delta is widening as the Chinese labs pour engineering into MoE routing, KV-cache reuse, and speculative decoding.
Thompson's read on the politics is characteristically dry. The US export controls on H100s and now Blackwell-class GPUs were supposed to slow China's frontier work. Instead, they forced the Chinese labs to optimize the entire serving stack — because if you can't win on raw FLOPs, you win on tokens-per-dollar. That optimization work is now the moat.
The interesting claim in the piece isn't "China is catching up." That framing is a year stale. The interesting claim is that the US frontier labs are structurally disincentivized from competing on serving efficiency, because their business model requires the closed API to be the product. Anthropic doesn't publish Claude's weights because Claude-the-API is the revenue. OpenAI is in the same box. Meta ships Llama weights but has visibly slowed on frontier training. Google ships Gemma but keeps Gemini closed.
Meanwhile the Chinese labs treat weights as marketing for their cloud businesses (or, in DeepSeek's case, as a side-effect of a hedge fund's research spend). They have no reason to hide the serving tricks — if anything, publishing them accelerates ecosystem adoption of their models over Western alternatives. The result is a widening asymmetry: one side is optimizing to keep the API expensive, the other is optimizing to make self-hosting cheap.
Thompson also picks at a soft spot most Western coverage skips: the "safety" argument for closed weights has quietly collapsed as a differentiator, because the open Chinese models are demonstrably no more dangerous in practice than the closed US ones, and the CCP-alignment concerns cut both ways — you can fine-tune them out in an afternoon on a single node. The remaining case for paying 10x for a closed API is convenience, tool-use reliability, and the specific tuning of Claude/GPT for agentic workloads — and even that gap is measured in months, not years.
The community reaction on HN (441 points, top of the front page) largely agreed with the diagnosis but split on the prescription. One camp — mostly infra people already running vLLM or SGLang in production — treated the piece as validation of what they've been doing. The other camp — mostly product engineers on Claude Code or Cursor — pushed back that raw benchmark parity doesn't translate to agentic tool-use parity, and that the last-mile reliability is exactly what you're paying Anthropic for. Both are right.
If your inference bill is a line item anyone in finance can find, you should already have a Qwen or DeepSeek path wired up behind a router. Not as your primary — the tool-use reliability gap on complex agent loops is real, and Claude Sonnet 4.7 still wins on multi-turn code editing by a meaningful margin. But as a fallback for the 60–80% of calls that are summarization, classification, embedding-adjacent reranking, or first-draft generation, the price gap now pays for the ops overhead of running your own endpoint (or renting one from Together, Fireworks, or DeepInfra) in weeks, not quarters.
The practical move for most teams in 2026 is a two-tier router: cheap Chinese open-weight for the bulk of traffic, frontier closed API for the calls that actually need it, with a cost dashboard that makes the split visible. The router doesn't have to be clever — a rule based on prompt length and whether tool-use is invoked gets you 80% of the savings. LiteLLM, OpenRouter, and Portkey all handle this out of the box now. The engineering cost is a week; the savings on a $50k/month inference bill are enough to fund the person who did the work several times over.
The second-order implication is on hiring. If you're building infra, MoE serving expertise (vLLM internals, expert parallelism, KV-cache tricks) is now a more valuable skill than knowing which prompt template Claude prefers. The market hasn't fully priced this in — job posts still over-index on "LangChain experience" — but it will.
Thompson's piece is a good excuse to update your priors, not a call to rip out your Anthropic contract. The frontier closed labs are still ahead on the specific workloads (long-horizon agents, computer-use, deep code editing) that justify their premium — for now. But the direction of the arrow is unambiguous: the price floor for "good enough" intelligence is falling by an order of magnitude every 12–18 months, and it's the Chinese open-weight labs setting that floor. If your product's unit economics depend on the closed API premium staying where it is, you have maybe 18 months to find a different moat.
> It’s striking the extent to which Claude Code and Codex are proving to be quite sticky; whichever harness you start working with is likely to be the one you stick with, and that figures to be even more the case with non-technical users.My experience has been quite the opposite. I was using Clau
I am afraid — if Chinese models go mainstream it has a clear way of pushing its narrative way beyond its otherwise borders. More like a Trojan horse it is for the Chinese.Here is a quick example of how Chinese deepseeks agent works kn its underlying model) when asked a tough questionhttps:/
The 2 things people need to remember:1) China can (and does) use the models to influence the west. They train in false information about Taiwan and Hong Kong. Or pretend like history is in favor of China.2) Ignoring the models containing false information, they are incredible. But you should be scar
I operate an analytics site (pretty big one B2B where client's backend feeds data into our system), and we see tons of traffic originating from northwestern China (Xinjiang) from Shenzhen Tencent Computer Systems Company Limited.There are also half a dozen other companies from China continuousl
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
The people who are most afraid of Chinese models are the VCs who poured into Anthropic and OpenAI at astronomically high valuations. Anthropic is valued at $1.2T and OpenAI is targeting $850B. These astronomical valuations were built on the premise that these labs would generate massive profits from