A Different Kind of Chip War
The semiconductor competition between East Asia and the United States has generated years of headlines about fab yields, export controls, and advanced packaging. Less discussed — but accelerating in 2025 and into 2026 — is the parallel contest over large language models, where Chinese, Korean, and Japanese labs have begun releasing open-weight models that are competitive with leading Western systems on several benchmarks.
Open-weight release is the key variable. When Alibaba published Qwen 2.5 weights in September 2024, developers across Asia, Europe, and Africa gained access to a Chinese-origin model that scored comparably to Meta's Llama 3.1-70B on multilingual tasks, without routing requests through any US-based API. DeepSeek's R1 release in January 2025 drew wider attention when it matched GPT-4o on reasoning benchmarks while reportedly requiring significantly lower training compute — a claim that has been scrutinised but not definitively refuted.
Why Open Weights Matter for the Region
For East Asian enterprises and governments, open-weight models offer something proprietary APIs cannot: the ability to run inference inside their own data centres, under their own legal jurisdiction, without dependency on a foreign provider's availability or pricing decisions. South Korean financial regulators, for instance, have strict data-residency requirements that make cloud-based LLM APIs from US providers complicated to use for customer-facing banking applications. Running a fine-tuned Korean-language model on domestic infrastructure sidesteps that complication entirely.
Japan's government-backed GENIAC programme has directed funding toward Japanese-language model development specifically to avoid what officials have described as "undue dependence" on models trained primarily on English data. Sakura Internet and NEC have both announced Japanese LLM initiatives targeting enterprise use in finance, healthcare, and manufacturing — sectors where Japanese-specific domain knowledge and regulatory vocabulary matter.
In South Korea, Kakao's KoGPT and Naver's HyperCLOVA series have been running in production for several years, but 2025 brought a new wave of smaller, more deployable models aimed at on-device or private-cloud use cases rather than large-scale cloud API consumption.
DeepSeek's Cost Claim and Its Aftermath
DeepSeek's R1 attracted unusual scrutiny because its reported training cost — around $5.6 million in GPU hours, compared to the hundreds of millions typically cited for frontier models — suggested either a genuine efficiency breakthrough or a cost-accounting methodology that excluded earlier experimental runs and the chips used to develop the training infrastructure itself. Analysts at Epoch AI noted the difficulty of making direct comparisons given the opacity of training logs.
What the release definitively established, regardless of the training cost debate, is that a Chinese lab operating under US chip export controls could produce a model with frontier-competitive reasoning performance. Whether that was achieved through algorithmic innovation, efficient use of older hardware, or both, the result altered the assumption that export controls on H100 and A100 chips would create a durable performance gap.
Alibaba's Qwen Series and Enterprise Adoption
Alibaba's Qwen series has achieved the broadest deployment footprint of any East Asian open model, in part because Alibaba released multiple sizes — from 0.5B parameters to 72B — making the models usable on hardware ranging from consumer laptops to enterprise GPU clusters. Qwen 2.5-Coder, the coding-specialist variant, has been adopted by developers across Southeast Asia for tasks where latency and API cost matter more than marginal accuracy differences versus GPT-4o.
Alibaba Cloud's model API offers Qwen at prices that, in mid-2025, were lower than comparable OpenAI endpoints by a factor of roughly three to five depending on context length — a pricing advantage that has attracted adoption in cost-sensitive markets including Indonesia, Vietnam, and India, where developer budgets for AI API calls are more constrained than in the US or Europe.
Regulatory Context Across the Region
China has required AI models offered to Chinese consumers to complete a government approval process — the Measures for the Management of Generative AI Services, in force since August 2023 — which has channelled Chinese deployments toward domestic models by default. Outside China, East Asian open-weight models operate in a more permissive regulatory environment than their proprietary Western counterparts, since open weights released under permissive licences are harder to regulate at the border than API services.
The European Union's AI Act, now in effect, applies to models deployed in the EU market regardless of origin. Japan's AI governance framework, updated in 2025, focuses on voluntary commitments by developers rather than pre-deployment licensing. South Korea's AI Basic Act, passed in late 2024 and entering force gradually through 2026, requires impact assessments for high-risk AI applications but does not restrict open-weight model distribution.
The competitive dynamics between East Asian open-weight models and Western proprietary APIs will depend substantially on whether this regulatory divergence persists, or whether new multilateral frameworks begin to create comparable obligations across jurisdictions.