TL;DR: China’s AI story in 2026 is no longer about catching up. Chinese models went from a rounding error on Western model aggregators in late 2024 to roughly 61% of token volume on OpenRouter by mid-2026, while the US share fell from about 70% to 30% in a single year. Four frontier-class open models shipped inside a seven-week window this spring, and the price gap against closed frontier models is now the main reason Western engineering teams switch. The constraint has moved from capability to people: AI roles are over a quarter of China’s new-economy job postings, and for high-performance computing engineers there are roughly seven openings per available candidate. If you are building anything on Chinese models, the practical question in 2026 is not whether they are good enough. It is who you can hire to run them.
For two years the standard framing of Chinese AI was a gap measurement. How many months behind the frontier, how much compute was denied by export controls, how far the benchmarks trailed. That framing stopped being useful somewhere around the middle of 2026.
What replaced it is more interesting and more practical. Chinese labs picked a different distribution strategy, shipped it faster than anyone expected, and are now the default option for a large slice of the world’s cost-sensitive AI workloads. This piece walks through the five trends that actually matter in 2026, what the numbers behind each one are, and what each means if your job is to build a team rather than to write commentary.
Five things to take away
- Chinese models went from a rounding error to roughly 61% of OpenRouter token volume in under two years, while the US share fell from about 70% to 30%.
- Four frontier-class open models shipped in a seven-week window in spring 2026, and adoption now moves in days rather than quarters.
- Price is the wedge: 60% to 90% cheaper than closed frontier models for comparable work.
- China leads the world in AI research volume at 23.2% of global publications, but trails badly on private investment, which is where the US still dominates.
- The binding constraint is people. AI postings grew about twelvefold year on year, and high-performance computing runs near seven openings per candidate.
Trend 1: open weights turned out to be the distribution strategy
The single most consequential decision Chinese labs made was to release weights. Not papers, not APIs behind a waitlist, but downloadable models with permissive licences. It looked like a concession in 2024. In 2026 it looks like the winning distribution channel.
The measurable result is share. On OpenRouter, the aggregator most Western developers route through, Chinese models climbed from almost nothing in late 2024 to roughly 61% of token volume by mid-2026. The mirror image is just as stark: US models fell from about 70% of token share to roughly 30% in the year to June 2026, on Bloomberg’s reading of OpenRouter and Exponential View data, with DeepSeek alone accounting for 16.3% of all token volume.
The download numbers point the same way. Alibaba’s Qwen family passed 700 million cumulative downloads on Hugging Face by January 2026, having overtaken Meta’s Llama in late 2025, and crossed one billion by July. It now anchors a large share of new fine-tuned derivatives. When an early-stage company pitches with an open-source stack today, the odds are strong that the weights underneath are Chinese.
We covered the model families themselves in detail in our guide to Chinese open-source LLMs and the labs behind them. The short version is that four organisations now carry most of the weight: DeepSeek, Alibaba’s Qwen team, Moonshot AI with the Kimi series, and Z.ai with GLM, alongside MiniMax which listed in Hong Kong on 9 January 2026.
Why open weights beat a cheaper API
A cheaper API wins on price for as long as the price holds. Open weights win on control: you can host in your own region, fine-tune on proprietary data, pin a version that will never be deprecated under you, and audit what the model does. That combination is why regulated buyers who would never send data to a Chinese API endpoint are still running Chinese weights on their own infrastructure.
Trend 2: release cadence became the real moat
Individual model quality gets the headlines. Cadence is what compounds. Between 24 April and 16 June 2026, four frontier-class open models shipped inside a seven-week window: DeepSeek V4, MiniMax M3, Kimi K2.7-Code and GLM-5.2. Each of them landed within reach of closed frontier systems on agentic engineering benchmarks.
Adoption now moves as fast as the releases. On Vercel’s AI Gateway, GLM-5.2’s daily token volume grew about fiftyfold between its 16 June availability and the end of the month, its customer count grew faster still, and it took 76% of its own model family’s June tokens in barely two weeks. Vercel notes that no in-family migration it had charted previously moved that quickly.
For engineering leaders this changes the shape of the decision. A model choice is no longer a two-year architectural commitment. It is closer to a quarterly review, which puts a premium on abstraction layers, evaluation harnesses and the people who can maintain both. Teams that hard-wired one vendor’s SDK into their product in 2024 are the ones paying for it now.
Trend 3: price is the wedge, and it is a wide one
Capability parity gets a model onto the shortlist. Price is what moves the workload. Chinese open models generally run 60% to 90% cheaper than the leading closed models for comparable work, and the four spring releases reached similar agentic benchmark scores at less than a third of the cost of the top closed tier.
The size of that wedge shows up cleanly in first-party gateway data. Vercel reported that open-weight models took 29% of its gateway tokens in June 2026 on just under 4% of spend, a segment that had almost tripled since April. Roughly a third of the work, for about a twenty-fifth of the bill.
That gap matters most where token volume is large and quality tolerance is stable: code review bots, document extraction, support triage, batch classification, agent loops that retry. It matters least where a single answer carries commercial or legal weight. Most serious teams in 2026 run a mix, and route by task rather than by vendor loyalty. Our cost and performance comparison of US and Chinese models breaks the routing decision down benchmark by benchmark, and our review of Chinese AI coding assistants covers the tooling layer that sits on top.
Trend 4: domestic adoption is industrial rather than consumer
Western coverage of Chinese AI tends to focus on chatbots. The deployment that actually shows up in the numbers is industrial. China has led global surveys of enterprise adoption for some time: SAS and Coleman Parkes found 83% of Chinese organisations using generative AI, ahead of the UK at 70% and the US at 65%, though the same study put China behind the US on full implementation. The applications concentrate in manufacturing quality control, logistics routing, public services and healthcare imaging rather than in consumer assistants.
The research and money picture underneath is lopsided in an important way. On Stanford HAI’s 2026 AI Index, China now leads the world in research volume with 23.2% of global AI publications and takes 20.6% of global citations against 12.6% for the US. Private investment runs the other way and not by a little: US companies put $285.9 billion into AI in 2025, roughly 23 times China’s $12.4 billion. China is producing the papers and shipping the weights on a fraction of the capital. We track the funding side in our Chinese AI investment statistics roundup, and the cross-country picture in our ranking of countries with the highest AI adoption rates.
There is a second-order effect worth naming. Export controls on advanced accelerators pushed Chinese labs toward efficiency work: sparse architectures, memory compression, aggressive quantisation, better inference scheduling. Those techniques travel. A model trained under a compute ceiling tends to be cheap to serve, which is precisely why the price gap in the previous section is as wide as it is.
Trend 5: the bottleneck moved from chips to people
This is the trend that gets the least attention and has the most operational consequence. On a Maimai study cited by People’s Daily and reported by the South China Morning Post, AI roles reached 26.23% of China’s new-economy job listings in the spring 2026 hiring season, up from 2.29% a year earlier, with postings growing about twelvefold year on year. Supply has not moved at anything like that rate.
The shortage is sharpest in the specialisms that make an inference stack economical. The supply-to-demand ratio for high-performance computing engineers sits at 0.15, roughly seven open positions chasing each available candidate, with cloud computing not far behind at 0.27. Those are the hardest single hires in the market, and they are exactly the people you need if your plan is to self-host open weights.
Pay has responded the way you would expect. AI engineers command the highest average monthly salary of any technical role in China at around ¥20,804, ahead of chip engineers and general software engineers. China Daily reports that algorithm engineers specialising in large language models and generative AI average about ¥650,000 a year, the highest-paid technical track in the country. Fresh graduates in AI now clear ¥17,000 a month on average, and the strongest candidates are signing offers above ¥30,000 a month before they graduate.
Read those two sets of numbers carefully, because they measure different things. The ¥20,804 average covers all AI engineer postings nationally, including the long tail of smaller employers. What an international employer pays a specialist is higher: our China AI engineer rate card puts mid-level engineers at ¥40,000 to ¥65,000 a month and senior engineers at ¥65,000 to ¥90,000, which works out to roughly $50,000 to $140,000 a year across the full range. Budget against the band you are actually competing in, not the national average.
What this means if you are hiring
Four practical conclusions follow from the trends above.
Hire for the inference layer, not the training layer. Very few companies need people who can pretrain a foundation model. Almost every company running Chinese open weights needs people who can serve them well: quantisation, batching, KV cache management, GPU scheduling, evaluation harnesses. That skill set is scarcer than the job title suggests and it is what turns a cheap model into a cheap product. Our LLM engineers and AI model training specialists pages cover how those roles are usually scoped.
Expect to compete on more than salary. In a market where strong graduates hold multiple offers before they finish their degree, cash alone does not close. Interesting problems, access to compute, publication freedom and a clear technical ladder all move candidates who will not move for another 10%.
Decide early how you will employ people. Hiring in mainland China means an entity, or a compliant alternative to one. Social insurance and housing fund contributions add materially to gross salary, and the rules vary by city rather than nationally. Our guide to EOR providers in China covers the options if you are not ready to incorporate, and our China EOR service handles the employment side directly.
Widen the map before you widen the budget. The skills that make Chinese open models economical to run are not exclusive to China. Vietnam, the Philippines, Singapore and Malaysia all have engineers doing serious inference and applied AI work, often at a fraction of the compensation the Chinese majors are paying. Our map of AI companies and startups in Asia is a useful starting point, and the global AI talent shortage statistics put the regional numbers in context.
Frequently asked questions
Are Chinese AI models actually competitive with the closed frontier in 2026?
On agentic engineering and coding benchmarks the gap is small enough that price dominates the decision for most workloads. On the hardest reasoning tasks and on long-horizon reliability, the closed frontier models still hold an edge. The sensible position is that they are competitive for a large majority of production traffic and not yet interchangeable for all of it.
Is it safe to use Chinese open-weight models?
Running open weights on your own infrastructure is a materially different risk profile from sending data to an overseas API. Weights can be inspected, hosted in your own region and version-pinned. The residual questions are licence terms, provenance of training data and your own regulator’s view, all of which should be answered before deployment rather than after.
What does an AI engineer cost in China compared with the rest of Asia?
China sits at the top of the regional range. Mid-level AI engineers hired by international employers run roughly ¥40,000 to ¥65,000 a month, against about $3,400 to $5,300 a month in the Philippines and lower again in Vietnam. Compare the bands on our rate cards for China, the Philippines and Vietnam.
Can I hire engineers in China without setting up a legal entity?
Yes, through an employer of record, which employs the person on your behalf and handles payroll, social insurance and the housing fund. It is the standard route for teams hiring their first few employees in the market. Contractor arrangements carry real misclassification risk in China and are not a substitute for full-time roles.
Hiring the people behind the trend
Every trend on this page ends at the same place: a team that can pick the right model, serve it economically and keep it honest. Second Talent recruits and employs that team for you across China and the wider Asia region, with compliant employment handled end to end so you can hire in weeks instead of standing up an entity first.
Tell us what you are building and we will come back with profiles and full employment costs, or see our pricing first.
Sources
- South China Morning Post, Demand for AI talent in China outpaces job postings in other new-economy sectors (Maimai study via People’s Daily, March 2026)
- Stanford HAI, The 2026 AI Index Report
- Vercel, AI Gateway Production Index, July 2026
- Xinhua, Alibaba’s Qwen leads global open-source AI community with 700 million downloads
- China Daily, AI boom drives hiring on tech edge
- OfficeChai, Share of US models on OpenRouter has collapsed from 70% to 30% (OpenRouter and Exponential View data via Bloomberg)
- SAS and Coleman Parkes, China leads world in GenAI usage while US leads in full implementation
