Why Chinese AI Labs Caught Up: Open Weights, Hardware Limits, and the New Model Race

01 The Surprise: Frontier Performance From Constrained Hardware

The measured facts came first. In January 2025, a Hangzhou company called DeepSeek released DeepSeek-R1, an open-weights reasoning model whose benchmark scores landed within touching distance of the best closed US models, and made the weights downloadable under an MIT license. The company behind it, per its Wikipedia summary, is "a Chinese artificial intelligence (AI) company that develops open weights large language models," based in Hangzhou, Zhejiang, and "owned and funded by High-Flyer, a Chinese hedge fund." Within weeks, Kimi from Moonshot and GLM models from Zhipu were posting comparable numbers, and by 2026 Chinese labs were recurring fixtures at or near the top of public leaderboards. The Asian Boss video examined at the top of this article captures the moment the wider public noticed: frontier-class capability, in domains from reasoning to video generation with models like SeeDance 2.0, arriving from labs that could not legally buy the best American chips.

What makes the moment analytically interesting is not that the labs caught up, but how. Our interpretation, flagged as such: this was not one breakthrough. It was the compounding of several efficiency decisions that hardware constraints forced earlier and harder than anywhere else. The rest of this piece separates what is measured from what is inference, because in this story the two get blurred constantly.

02 Efficiency as a Weapon: Distillation, MoE, and Smarter Data

The engineering story is documented in the labs' own technical reports. DeepSeek-V3 is a mixture-of-experts model with roughly 671 billion total parameters but only about 37 billion activated per token, which means the model gets the capacity of a very large network while paying the compute cost of a much smaller one. The same reports describe multi-head latent attention to shrink the memory the model has to keep alive during inference, fp8 mixed-precision training to double down on throughput, and heavy investment in data curation so that fewer, better tokens do more work. R1 added reinforcement-learning-driven reasoning on top of that base. Kimi K2 and GLM-4.5 followed the same pattern: trillion-scale or near-trillion-scale total parameters in mixture-of-experts layouts with tens of billions of parameters active per token, released with open weights.

Why does this matter strategically? Because every one of those techniques is a way to buy capability with something other than raw compute. Distillation moves knowledge from large teachers into cheaper students. Sparsity means the expensive parts of the network are only paid for when needed. Better data means fewer training passes for the same result. These are measured, published engineering choices. The interpretation that they were accelerated by export controls is ours, and it is widely shared by researchers in both countries: necessity compressed a timeline that abundant hardware would have stretched out.

03 Open Weights as Strategy, Not Charity

Releasing weights under permissive licenses is sometimes framed as generosity. It is better understood as distribution. Open weights let a lab reach every developer, company, and hobbyist on earth without a sales force, a data-center footprint, or an enterprise contract, and they make restriction attempts futile in a way that an API never could: once the weights are on laptops and local clusters, the model cannot be switched off. For a lab trying to enter a market where US providers were entrenched, that is the fastest possible route to adoption. It also commoditizes the thing incumbents charged for, which is a classic challenger move even outside AI.

The funding structure matters here, and it is a measured fact: DeepSeek is owned and funded by High-Flyer, a quantitative hedge fund, so its ability to spend on research does not depend on winning the enterprise SaaS race on day one. Our interpretation: open weights lower the cost of being trusted in some markets and raise the cost of monetization in others, and the labs calculating that tradeoff decided adoption was worth more than license revenue. Closed US labs, with their product moats and enterprise relationships, calculated differently. Neither calculation is obviously wrong, but they are different bets.

04 The Hardware Ceiling and What It Forced

The measured policy background: the US Department of Commerce has progressively restricted exports of advanced AI accelerators to China, first excluding top-tier data-center GPUs and later closing the workarounds that had allowed slightly downgraded variants. Chinese labs trained their frontier models on export-compliant hardware, and their own technical reports describe the lengths they went to in response: custom communication kernels tuned at a low level to squeeze bandwidth out of tightly constrained interconnects, aggressive precision reduction, and scheduling that wasted as few chip-hours as possible. When you cannot add GPUs, you get better at using the ones you have.

But the ceiling cuts both ways, and honest analysis has to say so. Training-side efficiency has diminishing returns at the very largest scales, and reasoning models that think for minutes per query are hungry at inference time, in deployment, where the chip deficit binds hardest. If US policy keeps tightening, Chinese labs face a real constraint on serving their own best models at global scale. The interpretation to watch: whether efficiency innovation keeps outrunning the tightening, or whether deployment economics eventually becomes the binding constraint that training efficiency cannot solve.

05 The Price War and Commoditization Pressure

The pricing facts are public and stark. DeepSeek's May 2024 V2 pricing touched off a price war among Chinese API providers, and by late 2024 and early 2025 the published list prices for premium models had diverged by an order of magnitude between American and Chinese vendors. The chart below is a snapshot of published vendor pricing pages from that window: US premium models at $2.50 to $3.00 per million input tokens, DeepSeek's models at 27 to 55 cents. Reported training costs point the same direction, with DeepSeek reporting its V3 final pre-training run cost at roughly $5.6 million, a figure that covers only that final run and not the full research program, a caveat often lost in retelling.

The commoditization pressure follows directly. When a downloadable model performs within the range customers actually feel, and serving it costs a tenth of the incumbent price, the incumbent's premium has to be earned somewhere else: in integrated products, in distribution, in trust, in multi-year enterprise relationships. That is the measured mechanism. Whether US labs' moats are deep enough to hold that line is the open question, and it is an interpretive one. Our reading: pricing power at the model layer is eroding everywhere, and the defensible value is migrating to the layers above and below it.

06 What to Watch Next

Five measurable things will settle the debate faster than any amount of commentary. First, the quality trajectory of open-weights models on independent benchmarks, not vendor-chosen ones. Second, whether inference-time compute demands of reasoning models collide with Chinese labs' deployment capacity, which will show up as throughput and latency on their APIs. Third, the pricing floor: API prices can only fall to the marginal cost of serving, and watching how close the majors get to that floor tells you how real the commoditization is. Fourth, the talent pipeline, because the engineering culture that produced the efficiency wave is a measurable quantity in publications and open-source contributions. Fifth, the next capability domains, like the video generation exemplified by SeeDance 2.0 in the source video, where it is anyone's guess whether the same open-weights pattern repeats.

The most likely world two years out, in our judgment, is not one where anyone "wins." It is one where frontier capability is abundant, cheap, and partly uncontrolled, sold at thin margins by some vendors and given away as leverage by others. That world is being built right now, one open-weights release at a time.