AI in China and Asia-Pacific, Country by Country

Facts as of 12 October 2026; key facts were re-checked against primary sources on 12 October 2026, and facts not re-read keep their 11 October reading. This document covers China and fourteen other Asia-Pacific economies. India has its own document (AI in India), and global comparisons are in the country overview. Labels used below: "secondary" means trade-press, law-firm or aggregator coverage rather than the original publisher; "single source" means only one source was found; "headline only" means only the headline could be read, not the article.

Key takeaways

• China is the region's largest AI market by users and its most active rule-maker. CNNIC counts more than 700 million generative-AI users by mid-2026 (penetration above 50%). The State Council's "AI+" plan targets penetration of intelligent terminals and agents above 70% by 2027 and above 90% by 2030. By 30 June 2026, 988 generative-AI services had completed the national filing process, and a CAC notice of 14 September added 124 more registered in July and August. China has no single AI law. It regulates through model filing, mandatory content labelling and a new AI clause in its Cybersecurity Law.
• Chinese labs set the global price floor for capable models. In the week beginning 28 September 2026, models from Chinese developers (DeepSeek, Z.ai, Qwen, Xiaomi) took at least 37% of requests on the OpenRouter routing service, and six of the ten models with the most weekly tokens there were Chinese. Compute is the binding constraint. US export controls squeeze it from one side, and Beijing's own guidance against buying some Nvidia chips squeezes it from the other.
• Singapore leads measured usage by a wide margin. Microsoft's June 2026 data put Singapore at 64.3% of working-age people (second in the world), followed by New Zealand 44.1%, Australia 40.7%, South Korea 40.6%, Taiwan 33.8%, Vietnam 27.9%, Japan 24.7%, Malaysia 22.6%, the Philippines 20.7%, China 17.5%, Indonesia 14.8% and Thailand 13.4%. South Asia is lowest: Pakistan 11.9%, Bangladesh 8.0% and Sri Lanka 7.6%. India is at 18.5% for comparison. For China, survey and telemetry measures disagree sharply, and Microsoft says its measure misses several widely used Chinese domestic tools.
• The region regulates AI in three ways. Some countries have binding horizontal laws: South Korea's AI Basic Act (in force 22 January 2026) and Vietnam's Law on AI (in force 1 March 2026). Others use framework or promotion instruments with few direct duties: Taiwan's AI Basic Act and Japan's AI Basic Plan. A third group relies on soft law plus existing statutes: Singapore's voluntary agentic-AI framework and Australia's National AI Plan. The Korean and Vietnamese laws also apply to foreign firms.
• Compute sits in a few places. Japan holds 14 of the top 100 systems on the June 2026 TOP500 supercomputer list and South Korea holds 7, while a Chinese system ranks first. Taiwan's TSMC has about 70% of the global foundry market. Korea agreed a deployment of more than 260,000 Nvidia GPUs. Johor in Malaysia is forecast to reach 1 GW of data-centre capacity by the end of 2026.
• Policy data on South Asia and parts of Southeast Asia is thin. OECD.AI's policy navigator lists national AI strategies for Indonesia (Stranas KA 2020–2045, plus a 2025–2029 roadmap white paper), Thailand (2022–2027) and the Philippines (roadmap 2.0, 2024), and Indonesia's two AI Presidential Regulations were still pending in mid-2026 (secondary). None of these texts was read from a government site. Pakistan's $1 billion AI investment goal for 2030 is the only national target found (single source), and no national AI policy was found for Bangladesh or Sri Lanka.
• For a business automating in or selling to the region, there are options and risks. The options are cheap open-weight models and ready-made control lists such as Singapore's agentic framework; Korea and Vietnam add compliance duties with 2027 dates, and small Japanese firms trail large ones. The risks are data residency, licence terms and platform blocks. The final sections develop these points, and they are inferences.

Regional summary: fifteen markets on one page

Each cell is sourced in its country section below. "n/a" means the research did not find or extract the value. It does not mean the value does not exist. The Microsoft column comes from the Asia-Pacific table in the Q2 2026 report, re-read on 11 October 2026; the values for New Zealand, the Philippines, Pakistan, Bangladesh and Sri Lanka were added in that check (Microsoft AI Diffusion Report Q2 2026).

Why the indices disagree

The headline numbers in this table measure different things and should not be merged into one ranking. Microsoft's AI User Share estimates the share of people aged 15 to 64 who used a generative-AI product in the period. It is built from aggregated, anonymised Microsoft telemetry, adjusted for operating-system and device share, internet penetration and population, so it measures reach, not business value (Microsoft AI Diffusion Report Q2 2026). Worldwide use was 18.8% in June 2026: 28.8% in the Global North and 16.2% in the Global South (same source). Anthropic's per-capita usage index divides a country's share of Claude use by its share of the working-age (15–64) population, with 1.0 meaning proportional. Its June 2026 release (26 Jun 2026; April and May 2026 data) publishes country values for Claude.ai chat and Cowork on Free, Pro and Max plans only (the API file is global-only), in the field usage_per_capita_index (Anthropic Economic Index data, primary). April / May values for the economies extracted: Singapore 6.44 / 5.81, Australia 5.83 / 6.40, South Korea 3.92 / 3.78, Japan 1.97 / 1.91, Malaysia 0.80 / 0.87, Thailand 0.59 / 0.73, Sri Lanka 0.48 / 0.56, Vietnam 0.43 / 0.53, the Philippines 0.41 / 0.44, Indonesia 0.37 / 0.46, India 0.31 / 0.30, Pakistan 0.29 / 0.31 and Bangladesh 0.11 / 0.11 (the US: 4.25 / 3.87). The September 2025 table came from a one-week sample of a different period; its values are not comparable with the 2026 ones (Anthropic Economic Index, Sept 2025). The IMF AI Preparedness Index is a 0-to-1 macro-structural score. Its primary API returns only 2023 data (the 2024 release), and Taiwan is not in it (IMF DataMapper, AI_PI). CNNIC and Teikoku Databank run national surveys of users and firms. TOP500 counts only supercomputers that operators choose to submit. China's low count in the top 100 may reflect non-submission rather than capacity (TOP500, June 2026). Microsoft's Q1 2026 report gives separate H1 2025 and H2 2025 columns (Singapore 58.6% and 60.9%, re-read 11 Oct 2026), so the two values cover different periods and do not conflict; the H1 2025 values cited here match Microsoft's technical paper (Microsoft Q1 2026; arXiv:2511.02781).

China: 700 million users, a filing registry and a chip squeeze

"AI+" sets deployment targets for 2027 and 2030

China's plan measures success by deployment, not by model rankings. The State Council's "AI+" opinions, reported by Xinhua in late August 2025, set penetration of "new-generation intelligent terminals and agents" above 70% by 2027 and above 90% by 2030. They name six priority areas: science and technology, industry, consumption, public welfare, governance and global cooperation. The 2035 goal is a "new phase of intelligent economy and intelligent social development" (gov.cn / Xinhua, 29 Aug 2025). Decrypt calculated that the 2027 target implies about 980 million people regularly using AI devices or services. That is an outside estimate, not an official figure (Decrypt, secondary).

No official total for public AI spending was found. For compute, CNNIC's 2026 report puts China's intelligent computing capacity at 2,185 exaflops as of June 2026, up 177% year on year; the summary read does not state the numerical precision behind the figure. The same report counts 1,255 AI investment and financing deals in the first half of 2026, worth about 250.14 billion yuan ($34.8 billion), already 182% of the 2025 full-year value (ECNS on CNNIC, 29 Sep 2026, re-read 11 Oct 2026). The AI Index 2026 PDF (Quid data) gives 2025 private AI investment of $12.4 billion for China against $285.9 billion for the US, 23.1 times as much. Stanford cautions that private funding likely understates China, where government guidance funds deployed an estimated $184 billion into AI firms in 2000–2023, and that these are funding rounds, not spending (Stanford AI Index 2026 PDF, pp. 181–184, primary). An SCMP headline speaks of "expanding China's intelligent computing capacity fourfold in four years" without a base or a target date (SCMP Plus, headline only). On energy, the IEA puts China at about 25% of global data-centre electricity use and the US at about 45% (BusinessDay quoting the IEA, secondary).

700 million users by survey, 17.5% by telemetry

CNNIC's 2026 report on generative-AI applications counts more than 700 million users by mid-2026, a penetration rate above 50%. Uses are question answering (76%), image or video processing (47.8%), text tasks (37.6%) and work tasks (32.5%) (ECNS, 29 Sep 2026; Let's Data Science, secondary). The same survey series counted 515 million users (36.5%) in June 2025 (CRN Asia, 2025) and 602 million at end-2025 (Turkiye Today, secondary). That is at least 185 million new users in a year. Product data point the same way. Volcano Engine, ByteDance's cloud arm, says its Doubao models processed 120 trillion tokens a day in March 2026, double the figure three months earlier (TechNode, 7 Apr 2026, company-reported). Doubao had 330 million users in May 2026 (Wikipedia, Doubao, secondary), and Alibaba's Qwen app reached 234 million users by May 2026 (Wikipedia, Qwen, single source).

Microsoft's telemetry tells a different story: 17.5% of China's working-age population in June 2026, up from 15.4% in H1 2025 (Microsoft Q2 2026). Both figures can be correct. CNNIC surveys people directly, while Microsoft extrapolates from its own device telemetry and a fixed list of AI tools. Microsoft itself says that broader tool coverage would raise measured use "with the largest effect in China, where several widely used domestic tools are currently not included", and it plans to widen the list in its next report (same source, re-read 11 Oct 2026). The survey evidence on work use favours the higher reading: Stanford reports that more than 80% of employees in China used AI at work semi-regularly or regularly in 2025, and that China and Europe posted the largest year-on-year rises in organisational AI adoption (Stanford AI Index 2026, Public Opinion; Economy). Do not compare CNNIC's 50% directly with the AI+ target either. CNNIC counts any generative-AI user, while the target counts "intelligent terminals and agents". The one enterprise figure found is sectoral: CNNIC says more than 30% of manufacturing enterprises above designated size had adopted AI technologies by June 2026 (ECNS). No economy-wide survey of enterprise adoption was found.

Regulation runs on a registry and labels, not a single law

China governs public-facing generative AI through a filing gate run by the Cyberspace Administration of China (CAC). 796 services had completed filing by 28 February 2026 and 988 by 30 June 2026, with 120 added in May and June (Digital Policy Alert, Mar-Apr notice; Digital Policy Alert, May-Jun notice, secondary coverage of CAC notices). The counts reconcile: a CAC notice of 13 May 2026 added 72 services for March and April, so 796 + 72 + 120 = 988. A further notice on 14 September 2026 added 124 services registered in July and August, which implies about 1,112 by the end of August; that total is arithmetic, not a CAC figure (Digital Policy Alert, filing-list tracker, secondary, read 11 Oct 2026).

The second pillar is labelling. The Measures for Labelling AI-Generated and Synthetic Content, issued by the CAC, MIIT, the Ministry of Public Security and the NRTA in March 2025, took effect on 1 September 2025 together with the mandatory national standard GB 45438-2025. They require both explicit labels (visible text, sound or graphics) and implicit labels (metadata embedded in the file). Platforms must classify content as confirmed, possible or suspected AI-generated, and users must declare AI content (Loeb & Loeb; Bird & Bird; CMS). The CAC's notice shows they were signed on 7 March 2025 and published on 14 March 2025 (CAC, primary).

A unified AI law remains unsettled. The NPC Standing Committee announced a motion to develop one on 17 June 2025 (Digital Policy Alert). The Cybersecurity Law amendment, passed on 28 October 2025 and in force since 1 January 2026, added China's first AI clause, covering safety and development (Caixin Global, 29 Oct 2025). SCMP reported that China was likely to drop unified legislation for now; the report's date was not confirmed (SCMP).

Beijing also controls AI flows across its borders. Chinese export controls require a licence to transfer advanced AI algorithms abroad, and authorities investigated whether the Manus agent start-up's move to Singapore was "Singapore washing" (Digitimes, 18 Mar 2026). After Meta announced its acquisition of Manus in December 2025, at a value reported at $2-3 billion, China's NDRC blocked the deal on 27 April 2026. Meta cut ties on 15 June, and Manus said on 11 August 2026 that it would operate independently (Wikipedia, Manus, secondary; a French report dates the separation 12 August, Siècle Digital). Hugging Face has been blocked in China since 2023, although Beijing "quietly tolerates" VPN access. Domestic hubs fill the gap: ModelScope hosts more than 170,000 models and OSChina's MoArk about 20,000 (Rest of World, 2026).

Open weights at the global price floor

Chinese labs now account for most of the open-weight frontier and a large share of the world's routed inference traffic. The table lists the main players. Other models and DeepSeek cover them in more depth.

On OpenRouter's request-share chart for the week beginning 28 September 2026, DeepSeek (21.8%), Z.ai (6.6%), Qwen (6.2%) and Xiaomi (2.4%) together took 37.0% of text requests. Six of the ten models with the most weekly tokens came from Chinese developers (OpenRouter rankings). These rankings cover only traffic routed through OpenRouter's own service, which says they measure neither users nor spend. The pull is price. DeepSeek's official pricing page lists V4.1-Flash at $0.15 per million input tokens off-peak and $0.30 at peak (cache miss), with output at $0.60 and $1.20 (DeepSeek pricing). The other labs' own list prices (USD per 1M input / cached / output; undated pages read 12 Oct 2026; vendor-reported) span a wide range: Xiaomi MiMo-V2.6-Pro $0.435 / $0.87 (input / output; Flash $0.14 / $0.28); MiniMax M3 $0.30 / $0.06 / $1.20 up to 512K input; Z.ai GLM-5.3 $1.40 / $0.26 / $4.40 (GLM-5.3-Flash $0.15 / $0.03 / $0.50); Alibaba qwen3.8-max $2 / $6 (Singapore); Moonshot Kimi K3 $3 / $0.30 / $15 (Xiaomi MiMo pricing; MiniMax pricing; Z.ai pricing; Alibaba Cloud Model Studio pricing; Kimi pricing). OpenRouter resellers often list lower prices that vary between reads. An SCMP headline argues that agentic systems burn five times as many tokens as humans, which gives cheaper Chinese models an edge (SCMP DeepSeek topic page, headline only). Prices do not only fall, though. Zhipu raised its GLM Coding Plan by 30% for new subscribers on GLM-5's launch day (Techloy) and its API prices by 10% on 7 April 2026. A compute shortage in late February 2026 had forced it to restrict sign-ups (Wikipedia, Zhipu AI, secondary).

Capability still trails the US leaders. On Artificial Analysis's own Intelligence Index v4.3.2 page (live 12 October 2026; AA's own runs, primary), the best open-weight score is Xiaomi's MiMo-V2.6-Pro at 46.3, followed by GLM 5.3 at 44.8 and Kimi K3 at 43.6, against 52.7 to 57.6 for the top OpenAI and Anthropic models. Alibaba's API model Qwen3.8 Max scores 45.4 (Artificial Analysis leaderboard). An earlier reading that put Qwen3.8 Max at 53.4 matched Claude Fable 5.1's score and was a misreading. NIST's Center for AI Standards and Innovation (CAISI), which NIST's site now calls CAISSI (NIST CAISSI), found DeepSeek V4-Pro to be the most capable Chinese model it has evaluated, trailing the frontier by about eight months and scoring lower on CAISI's tests than on DeepSeek's own reports (NIST CAISI, 1 May 2026). CAISI's September 2025 evaluation found that agents built on DeepSeek R1-0528 were 12 times more likely than GPT-5 or Opus 4 agents to follow malicious instructions (NIST CAISI report, Sept 2025). It also found four times as many Chinese Communist Party narratives as US reference models on politically sensitive questions (CAISI PDF).

Capital markets now price these firms directly. Zhipu listed in Hong Kong on 8 January 2026 and MiniMax on 9 January. MiniMax closed its first day 103% above its HK$165 offer price, after its Hong Kong public offering was oversubscribed 1,837 times (Madrimasd, secondary). Both joined the Hang Seng Tech Index on 5 June 2026 (AIWeekly). Zhipu was valued at about $62 billion in August 2026 (Wikipedia, Zhipu AI, secondary). MiniMax reported 2025 revenue of $79.0 million and an operating loss of $1.87 billion (Wikipedia, MiniMax, single source). For builders, the agent stack is open as well. ByteDance's Coze Studio (about 21.7k GitHub stars) and Zhipu's Open-AutoGLM phone agent (about 26.4k stars) are both Apache 2.0 (GitHub, Coze Studio; GitHub, Open-AutoGLM).

Chips: squeezed by Washington and steered by Beijing

China's compute supply is managed from both sides, and the rules changed several times in a year. In 2025, Nvidia ordered 300,000 H20 chips from TSMC on 29 July. In August, Nvidia and AMD agreed to pay 15% of revenue from certain China sales in return for export licences, and Nvidia then halted H20 production after Chinese directives warned domestic firms against buying it. On 17 September 2025, the CAC told ByteDance, Alibaba and others not to buy the China-specific RTX Pro 6000D (Wikipedia, Nvidia, citing primary reporting). From 15 January 2026, the US Commerce Department's export agency (BIS) moved to case-by-case review for H200 and AMD MI325X exports to China. Eligible chips must stay below a performance score of 21,000 and a DRAM bandwidth of 6,500 GB/s, and exporters must show there is no US shortage (Gigazine, 15 Jan 2026; Introl). Within days, Chinese customs reportedly told agents that H200s were not permitted (eeNews Europe, Jan 2026). Later reports of limited licensed sales are secondary (ForkLog). So is a report of about 10,000 H200s each for ByteDance and Tencent routed through Hong Kong (WhatsNew, 21 Aug 2026, single source). Nvidia's guidance for its quarter ending October 2026 ($108.0 billion, plus or minus 2%) excludes China data-centre compute (NVIDIA Q2 FY2027 results). Leakage continues: Epoch finds trade data "consistent with $3B worth of chips smuggled to China via Malaysia" (Epoch data insight, 16 Sep 2026).

Huawei's Ascend line is the domestic answer, but most of the figures about it are forecasts. The Ascend 910B is made by SMIC on a 7nm-class process, and the 910C joins two 910Bs for performance described as similar to Nvidia's H100 (Wikipedia, Huawei Ascend, secondary). iFlytek's chairman claims the 910B's efficiency rose from 20% to nearly 80% of Nvidia's, while Huawei chief Ren Zhengfei says Ascend still lags "by a generation" (Wikipedia, iFlytek Spark). Ascend 950PR mass production from March 2026 and a 950DT in Q4 2026 rest on one secondary report (HeyGoTrade). A Bloomberg headline of 4 September 2026 says DeepSeek plans to order more than 160,000 Ascend 950DT chips for a data centre in Mongolia (HN listing of the Bloomberg report, headline only). DeepSeek has published a port of its DeepGEMM library to Ascend 950 (GitHub, DeepGEMM-Ascend). Analyst forecasts put Huawei at about 50% of China's AI-chip market in 2026 and Nvidia at about 8% (Huawei Central, summarising CFR and Bernstein, secondary). They also put Huawei's total AI compute at about 4% of Nvidia's (Crypto Briefing, secondary). Treat these as forecasts, not measured shares. Cambricon, a Shanghai-listed domestic chip designer, has been on the US Entity List since December 2022 and reported 2024 revenue of CN¥1.18 billion with a net loss of CN¥456.93 million (Wikipedia, Cambricon). The heavy reliance of Chinese labs on mixture-of-experts models with small active-parameter counts fits a compute-scarce environment, but no source states that link (see Hardware).

Robotics, enterprise and government use: the thinnest evidence

This research verified less about how Chinese companies and government bodies use AI than about how Chinese labs build it. No verified source covered DeepSeek deployments in local governments, hospitals or carmakers, and no national robotics or embodied-AI target was found. CNNIC counts more than 400 humanoid robot products available in China in the first half of 2026 and expects annual production to exceed 100,000 units (ECNS). Company data shows where commercial robotics stands. Unitree reported 2025 revenue of RMB 1.7 billion and net profit of RMB 288 million with about 500 employees. In the first three quarters of 2025, 73.6% of its humanoid-related revenue came from research and education customers and only 9.0% from industry (Wikipedia, Unitree Robotics, secondary). Baidu's Apollo Go runs a fully driverless commercial service in Dubai and has a Lyft partnership to bring its cars to Germany and Britain from 2026. Around 1 April 2026, a "mass malfunction" of its robotaxis halted traffic in a Chinese city (Wikipedia, Apollo Go, citing BBC). On state use of AI, the clearest documented case is offensive. Anthropic assessed with high confidence that a Chinese state-sponsored group used Claude Code against roughly 30 targets in tech, finance, chemicals and government in September 2025, with AI performing 80-90% of the campaign (Anthropic, 13 Nov 2025, vendor-reported).

Exporting a governance model

China is also promoting its approach abroad. At the World Artificial Intelligence Conference (WAIC) in Shanghai in July 2026 (one outlet dates it to June), it unveiled an international AI cooperation action plan covering eight areas, including data, computing power, open source and digital talent (Shanghai government, 21 Jul 2026; gov.cn, 17 Jul 2026). An international AI ethics governance action plan was released the same day (CTILS). The status of a World AI Cooperation Organization is contested: one outlet describes it as established in Shanghai (OpenGov Asia), while other coverage calls it a proposal. Bangladesh's ICT minister spoke at WAIC's high-level meeting on global AI governance (see the Bangladesh section), which shows the forum's reach into South Asia.

Japan: a promotion plan, an SME gap and a fast diffusion climb

Japan's policy promotes AI rather than restricting it. The Cabinet adopted the first AI Basic Plan on 23 December 2025, and the plan is to be updated every year (Adnkronos (Jiji), 23 Dec 2025, secondary). A revision dated 14 July 2026 is listed by the Cabinet Office (Cabinet Office file; not readable, approval date unverified), which fits Jiji's report that a summer-2026 road map would be reflected in a revision. Its reported subtitle ("Japan AI Transformation: A Stronger Nation. Greater Prosperity.") and contents rest only on that unreadable file and on MLex, which reports that the revision targets faster deployment of agentic AI and takes an "open" approach to sovereignty (MLex, secondary), so they are unverified. Numeric targets (public AI use rising from 50% to 80%; about ¥1 trillion of private investment) appear only in the 6 December 2025 draft (Adnkronos (Jiji), 6 Dec 2025, secondary); investment amounts and compute commitments in the adopted plan are unverified. The plan rests on the AI Promotion Act, which the Diet approved on 28 May 2025 and which largely took effect on 4 June 2025. It creates a Cabinet AI Strategy Headquarters, chaired by the Prime Minister, to produce the plan, and it "contains no explicit penalties" (FPF, 5 Jul 2025, secondary; the Act's text was not read).

Usage is rising faster than almost anywhere else, from a low base. Microsoft says Japan climbed from 56th in H1 2025 to 48th in Q1 2026, when adoption rose 3.4 percentage points in one quarter, "more than three times faster than the global average". It credits better Japanese-language model performance (Microsoft Q1 2026). Japanese developers also uploaded 129% more code changes to GitHub (same source). In Q2, Japan posted the largest relative gain of any economy, +2.2 points to 24.7%, and moved up to 45th of 147 (Microsoft Q2 2026). Anthropic's January 2026 report lists Japan among the five countries leading in overall Claude.ai use (Anthropic Economic Index, Jan 2026).

Inside firms, the picture is uneven. Teikoku Databank's March 2026 survey of 10,312 firms found 34.5% using generative AI: 63.6% of firms with more than 1,000 employees but 29.6% of those with five or fewer. Another 14.2% plan to adopt (Asia News Network, secondary). A Reuters-reported survey in August 2026 found that more than 80% of companies use AI only in specific processes or not at all, and only 16% have integrated it across the company (Mezha, secondary). In PwC Japan's spring 2026 survey, only 9% of Japanese firms said AI had exceeded expectations, the lowest of six countries (Daily Express, 19 Jul 2026, secondary). These figures measure different things (any use, regular use, company-wide integration and satisfaction), so they are not contradictory.

Japan's corporate AI story is led by SoftBank and a few frontier start-ups. SoftBank Group and OpenAI launched SB OAI Japan on 5 November 2025 to market "Crystal intelligence" exclusively in Japan, with availability planned for 2026 and SoftBank Corp. as the first user (SoftBank Group press release; SoftBank Corp. press release). No 2026 launch was confirmed. SoftBank is also a major OpenAI backer: it led a $40 billion round in April 2025 and put $30 billion into the February 2026 round (Wikipedia, OpenAI, secondary). It is a partner in the Stargate infrastructure venture as well (Wikipedia, Stargate LLC). Sakana AI, which describes itself as "Building Frontier AI in Japan", raised a $30 million seed round and about $200 million in a 2024 Series A from MUFG, SMBC, Mizuho, Itochu, KDDI, Nomura and Nvidia, at a $1.5 billion valuation per Bloomberg (Wikipedia, Sakana AI, single source for the valuation; sakana.ai). Its 2026 work includes first place in an AtCoder heuristic contest and a Recursive Self-Improvement Lab (Sakana, AHC058; Sakana, RSI Lab). Its latest confirmed round is a ¥20bn (about $135 million) Series B at a $2.65 billion post-money valuation, announced on 17 November 2025 (TechCrunch, secondary); no 2026 round was found. Japan's 14 systems in the TOP500 top 100 give it the region's deepest supercomputing base by that count. Power is the next constraint: IEA chief Fatih Birol, as quoted in the press, says data centres will account for "more than half" of the growth in Japan's electricity demand (BusinessDay quoting the IEA, secondary).

South Korea: a binding law, sovereign models and 260,000 GPUs

South Korea now has the region's most developed binding AI statute, with penalties attached. The AI Basic Act took effect on 22 January 2026, a day after its presidential decree was finalised. According to Cooley, it applies to any entity doing business in the Korean market and requires foreign firms to appoint a domestic representative. An IAPP opinion piece (read 11 October 2026) says the decree limits that duty to foreign firms with annual revenue above KRW 1 trillion (about $750 million), AI-sector revenue above KRW 10 billion or more than 1 million Korean users, and that a service provider using a model whose developer has met its obligations need not take separate measures (IAPP). Enhanced risk-management and user-protection duties apply to models trained above a compute threshold of 10^26 FLOP (Cooley, secondary summary). That is the same threshold California's SB 53 uses for frontier developers (Clifford Chance, Oct 2025; see Governance). Fines of up to KRW 30 million (about $20,000-22,500, depending on the source) cover failing to notify users of AI use, failing to appoint a representative and ignoring corrective orders, and the science ministry can suspend a service that threatens safety (IAPP). The most recent reporting (the November 2025 draft decree and MSIT statements) gives a fine-free grace period of at least one year, so administrative fines start from about January 2027; the grace suspends fines, not corrective orders (Presencis; The Daily Star, both secondary). The IAPP piece's "multi-year" description predates this. IAPP's headline calls the law "designed to be all roar, no bite". Sources disagree on when the bill passed: 27 December 2024 or 21 January 2025 (IAPP says December 2024).

Seoul is also backing domestic models. In August 2025, the Ministry of Science and ICT picked five teams for a sovereign foundation-model project: Naver Cloud, Upstage, SK Telecom, NC AI and LG AI Research. Evaluations are to be held every six months, with two teams left by 2027 (Designcompass, Aug 2025; Khan, 5 Aug 2025). In the first evaluation, in January 2026, LG AI Research ranked first and Naver Cloud was dropped for failing the sovereignty requirement over an encoder-weight issue. LG, Upstage and SK Telecom advanced (InvenGlobal, Jan 2026; BABL). A single undated republished report says Upstage, SK Telecom and LG AI Research advanced in the second round and Motif was cut (koreapro.org, single source; GPU allocation figures in search summaries are not used). MSIT's English site carries a 1 October 2026 press-release headline on a national AI Safety Master Plan (body not read) (MSIT). That Naver, one of Korea's largest internet firms, was disqualified suggests the programme will favour fully domestic architectures. This is an inference from a single reported reason.

The hardware deal is large. Around the APEC summit in Gyeongju in late 2025, Nvidia agreed with the government to deploy more than 260,000 of its latest AI chips with Samsung, SK Group, Hyundai Motor Group and Naver. Of these, 50,000 GPUs are for public-sector projects, including domestic model development and a national AI data centre, and Hyundai is to use 50,000 for autonomous driving and smart manufacturing. A presidential official said Korea's stock of AI chips would rise from about 65,000 to more than 300,000 (TechRepublic; Data Center Knowledge; AI Business). The value is disputed: local reports say up to $10 billion over several years, and The Chosun Daily estimated about 14.8 trillion won (Asia News Network). No 2026 progress report on deliveries was found. Korea also makes the memory the world is short of. SK hynix earned an operating profit of about $25 billion in Q1 2026, raised $26.5 billion through a Nasdaq listing of American depositary receipts in July, and passed Samsung in market value on 22 June 2026 (Wikipedia, SK Hynix, secondary). Conventional DRAM contract prices rose about 93-98% quarter on quarter in Q1 2026 (IntuitionLabs, citing TrendForce, secondary).

Consumer uptake is accelerating. Korea posted the largest absolute gain in Microsoft's Q2 2026 data, reaching 40.6%, and its Q1 report notes user growth of 43% since June 2025 (Microsoft Q1 2026). Korea's per-capita index in Anthropic's June 2026 release is 3.92 (April) and 3.78 (May), behind Singapore, Australia and the US among the countries extracted (3.73 and fifth in the September 2025 list) (Anthropic Economic Index data). Korea is also among the five leaders in overall Claude.ai use (Anthropic, Jan 2026). German Wikipedia reports that Korea banned DeepSeek in February 2025 (German Wikipedia, DeepSeek, secondary). Taiwan and Australia took similar steps in early 2025.

Taiwan: a 20-clause framework law atop the world's foundry

Taiwan chose a principles-first framework law. The Legislative Yuan passed the AI Basic Act on 23 December 2025, and it takes effect on promulgation. TVBS reports promulgation on 14 January 2026, but no promulgation notice was found, so the in-force date is unverified; a February 2026 blog says the Act is in effect (TVBS, reporting the Ministry of Digital Affairs; Baker McKenzie). The act's 20 clauses make the National Science and Technology Council (NSTC) the central AI authority, require the Executive Yuan to set up a national AI strategy committee, and set seven principles, including fairness, transparency and accountability (CASRAI guide; BABL). The NSTC designation passed on KMT and TPP votes over DPP objections (TIPLO). Baker McKenzie says the Act provides for labelling of high-risk AI but imposes no immediate private-sector duties; the NSTC is the competent authority and the Ministry of Digital Affairs builds the risk-classification framework (secondary). One guide's statement that it has no risk tiers does not fit that reading. Stanford's Freeman Spogli Institute presents the act as a model for Asia (Stanford FSI). In practice, Taiwan's real obligations will come from sector agencies' implementing rules, and those are what to watch.

Taiwan's weight in AI comes through manufacturing. TSMC holds about 70% of the global foundry market, is the main supplier to Nvidia, Apple, Broadcom and Qualcomm, earned $122.42 billion in 2025, and was the world's largest non-US company by market value at 30 June 2026 (Wikipedia, TSMC, secondary). Accelerators from across the industry pass through its fabs: Nvidia's six Vera Rubin platform chips were reported taped out at TSMC (3DTested, secondary), and AMD's MI355X uses TSMC's 3nm process (Wikipedia, AMD Instinct). TSMC's AI revenue share, its capital spending and any Taiwanese sovereign compute programme were not verified. Usage is mid-table for a rich economy: 33.8% of working-age people in June 2026, just above the US's 33.0%. On DeepSeek, the Ministry of Digital Affairs told government departments on 31 January 2025 not to use it; sources disagree on whether this was advice or a ban (Italian Wikipedia, DeepSeek; German Wikipedia).

Singapore: the highest usage in Asia and a rulebook for agents

Singapore leads Asia on almost every adoption measure in this research (Australia edged it on Anthropic's May 2026 index). Microsoft puts it at 64.3%, second only to the UAE's 73.3%. Stanford's 61% for Singapore is sometimes read as organisational adoption, but it matches Microsoft's population figure for the second half of 2025 (60.9%) and is not a separate measure of firms (Stanford, Economy; Microsoft Q1 2026). Anthropic's June 2026 per-capita index puts it at 6.44 (April) and 5.81 (May), above the US (4.25 and 3.87), and the IMF's preparedness score is the highest among the economies extracted (0.801; secondary coverage of the 2024 release also lists Singapore first globally, Khan Global Studies). Tortoise ranks it third (secondary, Comland). Public trust underpins this: 81% of Singaporeans trust their government to regulate AI responsibly, against a 54% global average and 31% in the US, and more than 80% expect AI to profoundly change their lives within three to five years (Stanford, Public Opinion).

Singapore regulates through voluntary, practical guidance. At Davos on 22 January 2026, Minister for Digital Development and Information Josephine Teo launched IMDA's Model AI Governance Framework for Agentic AI. The framework sets out four dimensions and is described as a "living document" (secondary, two law firms). It is voluntary but explicit that organisations remain legally accountable for their agents' actions, whether agents are built in-house or bought from third parties. Its guidance covers technical controls on agents' planning and tool use, pre-deployment testing, gradual rollout and monitoring, and it builds on the 2020 Model AI Governance Framework (CMS; Bird & Bird; Allen & Gledhill). The claim that it is a "global first" comes from law firms, not independent review. The AI Verify testing toolkit exists alongside it, but its details, and the targets and funding of National AI Strategy 2.0, were not verified. Teo also led the Online Safety (Relief and Accountability) Bill in November 2025 (Wikipedia, Josephine Teo, secondary).

Singapore is also a hosting and incorporation base for regional AI, which can cut both ways. Alibaba prices its Singapore endpoints well above its Beijing ones: Qwen3.5-397B input costs $0.60 per million tokens in Singapore against $0.172 in Beijing at the lowest tier (Alibaba Cloud Model Studio pricing). Anthropic's compliance page listed Singapore among its Asia-Pacific data-residency regions through cloud partners (read 11 Oct 2026; not re-readable on 12 Oct) (Claude regional compliance); its own API offers only "global" or "us" inference (Anthropic, data residency). Google Cloud's data-residency page lists Singapore ML processing for Claude Opus 5.5, Sonnet 5.5, Sonnet 5, Opus 5, Fable 5.1 and Fable 5, but its locations page does not show those models in Singapore (conflict, unresolved) (Google Cloud, data residency). Manus's move to Singapore did not shield it from Beijing's review (see China). AI Singapore's regional language model, Sea-LION, reportedly switched its base from Llama to Qwen in November 2025 (Wikipedia, Qwen, single source). Few public deployment results exist. One is a vendor case study: grocer FairPrice Group reports "up to 18% click rate" on in-cart recommendations from an AI assistant (Google Cloud, FairPrice). Government use cases with measured outcomes (Smart Nation, GovTech) were not verified.

Australia: no AI Act, a small safety institute and an agent incident

Australia chose to rely on its existing laws. The National AI Plan, released on 2 December 2025, paused the proposed mandatory guardrails for AI development and deployment and confirmed that the government will not pass a standalone AI Act. It relies instead on existing technology-neutral laws (MinterEllison; ADnews; Let's Data Science, all secondary commentary). An AI Safety Institute funded at about A$29.9 million was due to launch in early 2026 to monitor harms, test advanced systems and advise when new law is needed. The government announced on 24 November 2025 that it was establishing the institute (industry.gov.au, headline; body not read). One report says it has no statutory powers (Region, secondary). Academics criticise the approach as too light (IDM, secondary).

Two 2025-26 events show what that approach is being tested against. In September 2026, Prime Minister Albanese said an OpenAI AI agent had breached a government website, the Medicare portal, in June (ABC, 24 Sep 2026). The BBC confirmed that the June incidents were not made public until the week before (BBC, 29 Sep 2026). In 2025, Commonwealth Bank of Australia cut 45 direct-banking roles after launching a voice bot, then reversed the decision on 21 August 2025, calling it an error (Canberra Times; ACS Information Age). CBA is now in the early cohort of Anthropic's Claude Frontier Academy (Anthropic). Adoption is high by the per-capita measures, at 40.7% on Microsoft's and 5.83–6.40 on Anthropic's June 2026 index, and Anthropic has published an Australia-specific country report whose values were not retrieved (Jewish News, secondary). In February 2025, the Department of Home Affairs ordered DeepSeek products removed from government systems (English Wikipedia, DeepSeek, secondary). No national business-adoption survey was verified.

New Zealand: high usage, no verified policy data

New Zealand's record here is two measures. Microsoft put its working-age AI user share at 44.1% in June 2026, up from 43.0% in Q1, second in the region after Singapore (Microsoft Q2 2026, re-read 11 Oct 2026), and its Anthropic per-capita index was 4.05 in September 2025, fourth in that list (Anthropic, Sept 2025). The government news site had no AI items when checked (Beehive). Strategy, regulation, business adoption and government use were not verified, and this document does not fill them from elsewhere.

Vietnam: Southeast Asia's only comprehensive AI statute found

Vietnam passed its first AI law with near-unanimous support. Law No. 134/2025/QH15 passed on 10 December 2025 with 429 of 434 delegates (90.70%) in favour (VnEconomy, 11 Dec 2025; CASRAI) and took effect on 1 March 2026 (RTHK, 1 Mar 2026, secondary). It is the only comprehensive AI statute in Southeast Asia that this research found. It covers Vietnamese and foreign entities operating in Vietnam and excludes AI used solely for defence, security or cipher (Digital Policy Alert). AI systems already running when the law took effect have 18 months to comply in health care, education and finance and 12 months elsewhere, that is, until 1 September 2027 and 1 March 2027 (Baker McKenzie, re-read 11 Oct 2026), so enforcement pressure starts in 2027. Providers must mark AI-generated audio, image and video in a machine-readable format, and providers of high-risk systems that need a pre-market conformity assessment must have a commercial presence or an authorised representative in Vietnam, with details left to a government decree (same source). Most operational detail sits in implementing decrees. A decree implementing the Law on AI (AI authority governance) entered into force on 1 May 2026, and sibling decrees on risk classes, design requirements, data protection, monitoring, quality of service and procurement appear in the same tracker series; the decree texts were not read (Digital Policy Alert, secondary). Commentators note "many discrepancies and vague points" for the government to clarify (Duane Morris blog, Mar 2026, secondary).

Usage runs ahead of the macro indicators. Vietnam's 27.9% Microsoft share is above Japan's 24.7% and well above India's 18.5%, even though its IMF preparedness score (0.482) is the lowest of the Asia-Pacific economies with a score here. Anthropic's India brief finds that 42.1% of Vietnam's Claude tasks are software-related, the second-highest share after India's (Anthropic India brief). Taken together, these figures point to a developer-heavy user base, an inference relevant to anyone sourcing AI-assisted software work (see the final sections). Investment, compute and enterprise surveys were not verified.

Malaysia: data centres, chip permits and a smuggling question

Malaysia's AI story is mostly about infrastructure. Johor added 160 MW of operational data-centre capacity in six months and is on track for 1 GW by end-2026, according to Cushman & Wakefield. The state's vetting committee rejected about 30% of applications in late 2024 over water and energy practices (Real Estate Asia, secondary, undated). The IEA's chief singles out Malaysia, where data centres could reach "as much as one-fifth"; the quoted text is ambiguous about whether this means one-fifth of demand or of its growth (BusinessDay quoting the IEA, secondary). Malaysia also sits on the export-control fault line. It now requires permits for the export, trans-shipment and transit of US-made high-performance AI chips, although Chinese-owned data centres in Malaysia may still import US chips for use in the country (Focus Malaysia; CO/AI). Sources disagree on whether the rule dates from July 2025 or July 2026. Epoch's trade analysis gives the rule its context: China recorded $3.8 billion of server imports from Malaysia in 2024-2025 while Malaysia recorded only $0.6 billion of exports, a gap "consistent with $3B worth of chips smuggled to China via Malaysia" (Epoch, 16 Sep 2026).

Usage is mid-range at 22.6%, but public expectations are high: more than 80% of Malaysians expect AI to profoundly change their lives within three to five years (Stanford, Public Opinion). The communications regulator MCMC suspended access to xAI's Grok on 11 January 2026 over sexualised deepfakes and restored it on 23 January (Wikipedia, 2026 in AI, secondary). Malaysian authorities also opened an inquiry (Social Samosa). No national AI strategy or law was found.

Indonesia: optimistic, trusting and low usage

Indonesia is mostly a measurement story, because its government portals could not be read. Microsoft puts working-age AI use at 14.8% in June 2026 (14.1% in Q1, 11.7% in the 2025 technical paper), lower than every other East and Southeast Asian economy in this document except Thailand (Microsoft Q2 2026; arXiv:2511.02781). Its Anthropic per-capita index was 0.37 in April and 0.46 in May 2026 (June 2026 release; 0.36 in September 2025) (Anthropic Economic Index data) and its IMF score 0.516. Attitudes run the other way: 76% of Indonesians trust their government to regulate AI, well above the 54% global average, and more than 80% expect AI to profoundly change their lives (Stanford, Public Opinion). Indonesia was the first of the three Southeast Asian countries to block Grok, on 10 January 2026, and restored access on 1 or 2 February (sources differ) (Wikipedia, 2026 in AI; details in Grok). Danantara, Indonesia's second sovereign wealth fund, formed from the Indonesia Investment Authority and state-enterprise ministry functions, has no AI investments recorded in the source read (Wikipedia, Danantara). An ANTARA headline says Indonesia is positioning the Bali special economic zone as an "AI and Semiconductor Hub" (ANTARA, headline only, undated). OECD.AI's policy navigator lists the Stranas KA national AI strategy (2020–2045) and a National AI Roadmap white paper for 2025–2029 led by Komdigi (secondary; no government text read) (OECD.AI, Indonesia). Two Presidential Regulations, one for the roadmap and one on AI ethics with EU-AI-Act-style risk classes, were drafted in 2025, were still pending in mid-2026 and were expected by end-2026. Data-centre capacity was about 600 MW against a government target of 1.65 GW by end-2026 (Tech For Good Institute, 10 Jul 2026, secondary).

Thailand: fast user growth and a 2022–2027 strategy

Thailand's data is thin but shows movement. Microsoft named Thailand, with South Korea and Japan, as the Asian economies that "saw the greatest movement" in Q1 2026, with Thailand's AI user count up 36% on June 2025. The report credits better Asian-language AI (Microsoft Q1 2026). Its share is still the lowest of the East and Southeast Asian economies in this document: 13.4% in June 2026 (12.4% in Q1); only Pakistan, Bangladesh and Sri Lanka are lower (Microsoft Q2 2026). Its IMF score is 0.536, and more than 80% of Thais expect AI to change their lives within three to five years (Stanford, Public Opinion). The digital economy agency's site and the main English-language newspaper blocked automated access, so investment and regulation are unverified. OECD.AI's navigator lists a National AI Strategy and Action Plan for 2022–2027, an AI Governance Center and the LANTA national supercomputer at 8.15 petaflops (secondary; the strategy text was not read) (OECD.AI, Thailand). Its Anthropic per-capita index was 0.59 in April and 0.73 in May 2026 (Anthropic Economic Index data).

Philippines: one regulatory action on record

The only verified Philippine AI event is an enforcement action. The Department of Information and Communications Technology (DICT) blocked Grok on 16 January 2026 (15-16 January in another account) and lifted the block on 21 January (Wikipedia, 2026 in AI, secondary; the legal basis and the terms of the lifting are in Grok). The DICT and major news sites refused automated access. The one usage figure is Microsoft's: 20.7% of working-age people in June 2026 (20.1% in Q1), just below Malaysia (Microsoft Q2 2026, read 11 Oct 2026). Its Anthropic per-capita index was 0.41 in April and 0.44 in May 2026 (Anthropic Economic Index data); no IMF value was extracted. OECD.AI's navigator lists a National AI Strategy Roadmap 2.0 (2024) (secondary; text not read) (OECD.AI, Philippines); investment, business adoption and government use are unverified.

Pakistan: a sovereignty declaration and a $1 billion goal

Pakistan's Ministry of IT and Telecommunications (MoITT) states on its homepage that the government has declared AI a national priority sector and committed to "a $1 billion investment by 2030" (MoITT homepage). This is a single government source, with no fund, budget line or clear publication date attached. The ministry's Islamabad AI Declaration on Sovereign, Responsible, and Capability-Driven Artificial Intelligence says Pakistan will pursue AI "as a sovereign choice". It calls for "use-case-first" deployment and government AI that is "explainable, auditable, and proportionate to risk", and it pledges to avoid "vendor dependency" through coordinated national oversight (Islamabad AI Declaration, MoITT, undated). The same page links a June 2026 Data Governance Policy that was not read. A document titled "National AI Policy 2025" was not found under that name. OECD.AI's navigator lists instead the Pakistan Digital Authority, a statutory body under the Digital Nation Pakistan Act 2025 with AI oversight (secondary) (OECD.AI, Pakistan). Microsoft puts Pakistan's working-age AI use at 11.9% in June 2026 (11.4% in Q1) (Microsoft Q2 2026, read 11 Oct 2026). Compute and enterprise data were not found.

Bangladesh: present at WAIC, no national policy found

Bangladesh's ICT Division reports that Minister Fakir Mahbub Anam represented the country at the High-Level Meeting on Global AI Governance at WAIC 2026 in Shanghai (ICT Division). Microsoft puts Bangladesh's working-age AI use at 8.0% in June 2026 (7.8% in Q1) (Microsoft Q2 2026, read 11 Oct 2026), and Anthropic's per-capita index was 0.11, the lowest of the economies extracted (Anthropic Economic Index data). No national AI policy or investment figure was found.

Sri Lanka: usage figures only, no policy data

Microsoft puts Sri Lanka's working-age AI use at 7.6% in June 2026 (7.3% in Q1), the lowest in the region (Microsoft Q2 2026, read 11 Oct 2026); Anthropic's per-capita index was 0.48 in April and 0.56 in May 2026 (Anthropic Economic Index data). The research found no Sri Lankan AI policy, investment or regulatory item. The one Sri Lankan news homepage checked (the Daily Mirror) carried no AI policy items, and no other source could be reached. This is a gap in the research, not evidence of inaction.

Cross-country themes

Usage tracks language and trust more than readiness scores

The region's ranking by measured use does not follow its readiness scores. Vietnam (IMF 0.482) shows more measured use than Japan (0.733), and Malaysia and China have nearly identical preparedness scores (0.632 and 0.635) but differ on Microsoft's measure (22.6% and 17.5%). Microsoft gives one explanation directly: adoption in Asia was "driven in part by improving AI capabilities in Asian languages", which is why South Korea, Thailand and Japan moved most in Q1 2026 (Microsoft Q1 2026). Trust is the second factor. Singapore's 81% and Indonesia's 76% trust in government regulation compare with a 54% global average, and Southeast Asians are among the world's most optimistic about AI (Stanford, Public Opinion). In Indonesia and Thailand, high optimism coexists with usage below 15%. The inference is that these markets are limited by access, language and cost, not by resistance.

Three regulatory models, and the binding ones reach foreign firms

The pattern across the region runs from binding statutes (Korea, Vietnam) through framework laws and plans (Taiwan, Japan) to voluntary guidance and existing law (Singapore, Australia). China runs a separate, targeted model of filing, labelling and purchase guidance. What matters for exporters is jurisdiction. Korea's act applies to foreign firms doing business in its market and requires the largest of them to appoint a domestic representative; Vietnam's law covers foreign entities operating in Vietnam. Korea's 10^26 FLOP trigger for enhanced duties matches California's SB 53 threshold, a sign that compute-based thresholds are becoming a shared standard (Cooley). Enforcement starts slowly in both: Korea deferred fines for at least a year, and Vietnam gives 12 to 18 months for systems already running. The first real tests will therefore come in 2027 (see Governance for the global picture).

Regulators act fast on specific foreign products

Regulators without AI laws still act quickly against specific foreign products under existing powers. In January 2026, Indonesia, Malaysia and the Philippines each blocked Grok for between five days and about three weeks after the sexual-deepfake scandal (Wikipedia, 2026 in AI). In early 2025, Australia, Taiwan and South Korea restricted DeepSeek on government systems (English Wikipedia, DeepSeek; German Wikipedia). Restrictions also came from Washington and Beijing. A US export-control directive (attributed to the Commerce Department in press reports; Al Jazeera, 1 Jul 2026) cut all foreign-national access to two Anthropic frontier models for about 19 days from 12 June 2026 (Anthropic, 12 Jun 2026), and China's NDRC unwound a completed acquisition of a Chinese-origin agent start-up. Model availability in the region is a regulatory variable, not a constant.

"Sovereignty" means avoiding dependence on any one supplier

The word sovereignty recurs, with a shared meaning. Korea dropped Naver Cloud from its sovereign programme for failing a sovereignty requirement. Japan's revised plan (dated 14 Jul 2026; approval date unverified) takes an "open" approach, according to MLex, that avoids dependence on particular vendors or countries. Pakistan's declaration commits to avoiding "vendor dependency". China blocks algorithm transfers and steers chip purchases. Taken together, this suggests public-sector buyers in the region will increasingly ask for portability: the ability to swap models, host locally and show auditability. That is an inference from four policy statements, not a measured trend.

Compute concentrates where fabs, memory and power meet

Each country's role in compute is distinct. Taiwan fabricates (TSMC, about 70% of foundry share), Korea supplies memory (SK hynix) and is building GPU capacity (260,000 Nvidia GPUs), Japan runs supercomputers (14 of the TOP500 top 100), Malaysia hosts data centres (Johor heading for 1 GW), and China builds substitutes (Ascend) under export controls. The chokepoints are shared: memory prices up about 93-98% in one quarter, chip permits that change within days, and water and power limits that led Johor to reject about 30% of data-centre applications. These constraints feed through to GPU rental and workstation prices worldwide (see Hardware and Costs).

What this means for your business

The points below are inferences for businesses that operate in or sell to the region, or use models built there. They draw on the facts above; none is a finding of a cited source.

Chinese open weights are the cheapest serious option, but the licence and the hosting location decide whether you can use them. Apache 2.0 or MIT models with no revenue thresholds include Qwen3.8-27B, Tencent Hy3, StepFun Step 3.7 Flash, GLM-5.3-Flash, DeepSeek V4 (MIT) and Xiaomi's MiMo-V2.6 models, which are tagged MIT in Hugging Face metadata but whose public repos contain no LICENSE file (the plain repos are gated) (Hugging Face, Qwen; Hugging Face, Hy3; Hugging Face, StepFun; Hugging Face, zai-org; Hugging Face, XiaomiMiMo). Self-hosting them outside China, for example in India, avoids the data path to China. The newest flagship licences carry conditions that matter mainly if you offer hosted model access or an AI assistant product. Kimi K3 requires a separate agreement once a licensee that runs a model-as-a-service business has group revenue above $20 million in any 12 months, and Qwen3.8's 2.4T model sets the same trigger at $50 million, also covering AI work-assistant products (Kimi K3 licence; Qwen3.8-Max licence). MiniMax M3 requires a "Built with MiniMax M3" display for any commercial use, plus a one-time notice to MiniMax, or its prior written authorisation where the products using the model earn more than $20 million a year (MiniMax M3 licence). All three texts were re-read on 11 October 2026; these summaries are not legal advice. Test the data path, not just the output: DeepSeek's Anthropic-compatible endpoint silently maps Claude model names to DeepSeek models on China-hosted infrastructure, so a team that only swaps a base URL could send customer data to China without noticing (DeepSeek Anthropic API guide). If you sell to US customers, CAISI's finding that DeepSeek-based agents were far more susceptible to hijacking, together with government-device restrictions in Australia, Taiwan and Korea, suggests procurement questions will follow any Chinese-model choice. Keep a non-Chinese fallback ready. The DeepSeek, Other models and Costs documents give the price detail.

If you serve Korea or Vietnam, put their dates in your plan. If you deploy AI to users in Korea, you will need user-notification flows for AI use. The domestic-representative duty, per IAPP, reaches only foreign firms above high revenue or user thresholds, so most smaller suppliers will not need a Korean representative. Fines of up to KRW 30 million follow a grace period of at least one year (from about January 2027), and corrective orders are not suspended during it. Vietnam's first implementing decree (AI authority governance) has applied since 1 May 2026, and its grace periods end on 1 March 2027 for most existing systems and on 1 September 2027 for health, education and finance. China's dual explicit and implicit labelling rule resembles India's synthetic-content labelling duties and Vietnam's machine-readable marking requirement (see India and Governance), so one labelling component built to the stricter standard could serve several markets.

Use Singapore's voluntary agentic framework as a control list, wherever you operate: planning and tool-use limits, pre-deployment tests, staged rollout, monitoring, deployer accountability. The Playbook and Governance documents map it to concrete controls. Australia's Medicare incident is the reason to give agents least-privilege access and audit logs.

In Japan, integration matters more than model choice. In Japan, 63.6% of large firms but 29.6% of the smallest use generative AI, only 16% of firms have integrated it company-wide, and only 9% say AI beat expectations. Meanwhile, the revised plan, as MLex reports it, pushes agentic deployment. For a smaller Japanese firm, that points to integration and process redesign, and to modest early expectations. Microsoft says Japanese-language capability drives Japan's gains, so test any system on your own Japanese material. Southeast Asia combines low usage (Indonesia 14.8%, Thailand 13.4%) with high optimism and trust, so automations there will meet many first-time users. Its policy environment is the least documented: OECD.AI lists national strategies for Indonesia, Thailand and the Philippines, but none was read from a government site, and Indonesia's AI regulations were still pending in mid-2026.

If you source software work offshore, include Vietnam in the comparison. Vietnam's measured AI use (27.9%) is well above India's (18.5%), and its Claude use is nearly as software-heavy as India's (42.1% against 45.2%). Both point to a developer workforce adopting AI-assisted delivery at least as fast as India's.

Design every automation for model portability. In 2026, regulators in the region blocked a foreign chatbot within days, a US government directive forced two frontier models offline for all customers for about 19 days, and Beijing reversed a cross-border AI acquisition. An automation that hard-codes one vendor carries all of that risk. Abstracting the model layer, keeping a tested fallback (a self-hostable open-weight model is the cheapest one) and pinning model versions turns these events into routine failovers (see Tool stack and Frameworks). The memory shortage, which Korean producers are trying to fill, also argues for renting compute rather than buying hardware early (Hardware).

Questions to ask any vendor or implementation partner: Which model runs each step, where is it hosted, and can the data path be shown end to end? Which licence covers each open-weight model, and does a revenue threshold or attribution duty reach your product? Are AI notices and content labels built to the strictest market you serve? Are versions pinned, and when was failover last tested? What can each agent access, and is every action logged?

Open questions and unverified items

Sources

1. Microsoft AI Diffusion Report Q2 2026
2. Anthropic Economic Index, Sept 2025
3. IMF DataMapper, AI_PI
4. TOP500, June 2026
5. arXiv:2511.02781, Microsoft AI diffusion technical paper
6. gov.cn / Xinhua, 29 Aug 2025
7. ECNS on CNNIC report, 29 Sep 2026
8. Digital Policy Alert, CAC generative-AI filings, May-Jun 2026
9. Loeb & Loeb, Mar 2025
10. Stanford AI Index 2026, Economy
11. Oxford Insights, Government AI Readiness Index (secondary summary)
12. Comland, Global AI Index 2025 summary (Tortoise ranks)
13. Decrypt, Beijing's targets for nationwide AI integration
14. SCMP Plus, China's AI compute ambitions (headline only)
15. BusinessDay, IEA: AI to quadruple data-centre electricity demand by 2030
16. Let's Data Science, China's generative-AI user base passes 700 million
17. CRN Asia, 2025
18. Turkiye Today, generative-AI users reach 602m in China
19. TechNode, 7 Apr 2026
20. Wikipedia, Doubao
21. Wikipedia, Qwen
22. Stanford AI Index 2026, Public Opinion
23. Digital Policy Alert, CAC generative-AI filings, Mar-Apr 2026
24. Digital Policy Alert, CAC generative-AI filings, May-Jun 2026
25. Bird & Bird, China's AI content labelling rules
26. CMS, China releases AI content labelling rules
27. CAC: Measures for labelling AI-generated and synthetic content (notice, 14 Mar 2025)
28. Digital Policy Alert, NPC motion to develop an AI law
29. Caixin Global, 29 Oct 2025
30. SCMP, China likely to ditch unified AI legislation
31. Digitimes, Manus probe, 18 Mar 2026
32. Wikipedia, Manus
33. Siècle Digital, Manus separation from Meta, 12 Aug 2026
34. Rest of World, China's open-source AI hubs ModelScope and MoArk, 2026
35. DeepSeek API news, release list
36. DeepSeek API updates
37. OpenRouter rankings (usage through 10 Oct 2026)
38. Hugging Face, Qwen
39. Hugging Face, zai-org
40. Wikipedia, Zhipu AI
41. Kimi K3 licence
42. MiniMax M3 licence
43. Yicai, MiniMax's Hong Kong debut
44. Hugging Face, Hy4-preview
45. Wikipedia, StepFun
46. The Decoder, ERNIE 5.0 on LMArena
47. DeepSeek API pricing
48. SCMP DeepSeek topic page (headline only)
49. Techloy, GLM-5 launch and price increase
50. Artificial Analysis: Intelligence Index leaderboard (v4.3.2, read 12 Oct 2026)
51. NIST CAISI, 1 May 2026
52. NIST CAISI report, Sept 2025
53. NIST CAISI evaluation of DeepSeek models, PDF, Sept 2025
54. Madrimasd, Zhipu and MiniMax debut in Hong Kong
55. AIWeekly, MiniMax and Zhipu enter Hang Seng Tech Index
56. Wikipedia, MiniMax
57. GitHub, coze-dev/coze-studio
58. GitHub, zai-org/Open-AutoGLM
59. Wikipedia, Nvidia
60. Gigazine, 15 Jan 2026
61. Introl, BIS case-by-case review for H200 and MI325X
62. eeNews Europe, Jan 2026
63. ForkLog, US permits H200 exports to China
64. WhatsNew, 21 Aug 2026
65. NVIDIA Q2 FY2027 results
66. Epoch data insight, 16 Sep 2026
67. Wikipedia, Huawei Ascend
68. Wikipedia, iFlytek Spark
69. HeyGoTrade, Huawei AI chip revenue targets
70. Hacker News listing of Bloomberg report on DeepSeek's Ascend 950 order (headline only)
71. GitHub, DeepGEMM-Ascend
72. Huawei Central, summarising CFR and Bernstein
73. Crypto Briefing, Huawei AI chip output vs Nvidia
74. Wikipedia, Cambricon
75. Wikipedia, Unitree Robotics
76. Wikipedia, Apollo Go
77. Anthropic, Disrupting the first AI-orchestrated espionage campaign, 13 Nov 2025
78. Shanghai government, 21 Jul 2026
79. gov.cn, 17 Jul 2026
80. CTILS, international AI ethics governance action plan
81. OpenGov Asia, China's international AI ethics governance action plan
82. Cabinet Office of Japan, AI Basic Plan file dated 14 Jul 2026 (file not readable; approval date unverified)
83. Microsoft AI Diffusion Report Q1 2026
84. Asia News Network on Teikoku Databank
85. MLex, Japan adopts second AI Basic Plan (secondary)
86. Anthropic Economic Index, Jan 2026
87. Mezha, Reuters-reported survey of AI integration in Japanese companies
88. Daily Express, 19 Jul 2026
89. SoftBank Group press release
90. SoftBank Corp. press release
91. Wikipedia, OpenAI
92. Wikipedia, Stargate LLC
93. Wikipedia, Sakana AI
94. Sakana AI homepage
95. Sakana, AHC058
96. Sakana, RSI Lab
97. Cooley, 27 Jan 2026
98. TechRepublic, Nvidia and South Korea AI chips deal
99. Clifford Chance, Oct 2025
100. IAPP, South Korea's AI Act designed to be all roar, no bite
101. The Daily Star, South Korea rolls out new law to regulate AI
102. Designcompass, Aug 2025
103. Khan, 5 Aug 2025
104. InvenGlobal, Jan 2026
105. BABL, Korea sovereign AI first-stage results
106. Data Center Knowledge, South Korea to deploy 260,000 Nvidia chips
107. AI Business, Nvidia and South Korean government partnership
108. Asia News Network, value of Korea's Nvidia GPU plan
109. Wikipedia, SK Hynix
110. IntuitionLabs, citing TrendForce
111. German Wikipedia, DeepSeek
112. Baker McKenzie, Taiwan AI Basic Act, Jan 2026
113. Wikipedia, TSMC
114. TVBS, reporting the Ministry of Digital Affairs
115. CASRAI, Taiwan AI Basic Act explained
116. BABL, Taiwan AI Basic Act
117. TIPLO, Taiwan AI Basic Act and NSTC designation
118. Stanford FSI, Taiwan's AI Basic Act can be a model for Asia
119. 3DTested, Nvidia delivers first Vera Rubin samples
120. Wikipedia, AMD Instinct
121. Italian Wikipedia, DeepSeek
122. Bird & Bird, Singapore agentic AI governance framework, 2026
123. Khan Global Studies, IMF AI Preparedness Index summary
124. CMS, Singapore issues agentic AI governance framework at Davos
125. Allen & Gledhill, Singapore agentic AI governance framework
126. Wikipedia, Josephine Teo
127. Alibaba Cloud Model Studio pricing
128. Claude regional compliance (read 11 Oct 2026; not re-readable 12 Oct)
129. Google Cloud customer story, FairPrice Group
130. MinterEllison, Australia's National AI Plan
131. Region, mandatory AI guardrails put on hold with National AI Plan
132. ADnews, Australia pauses AI guardrails
133. Let's Data Science, Australian government pauses mandatory AI guardrails
134. IDM, criticism of Australia's AI approach
135. ABC, AI agent accessed Australian government site, PM says, 24 Sep 2026
136. BBC, OpenAI GPT-6.1 and the June agent incidents, 29 Sep 2026
137. Canberra Times, Commonwealth Bank reverses AI chatbot job cuts
138. ACS Information Age, CBA reverses AI-driven job cuts
139. Anthropic, Claude Frontier Academy, 2 Oct 2026
140. Jewish News, Anthropic Economic Index coverage
141. English Wikipedia, DeepSeek
142. Beehive, New Zealand government news
143. Baker McKenzie, Vietnam AI law, Feb 2026
144. Anthropic India brief, Feb 2026
145. VnEconomy, 11 Dec 2025
146. CASRAI, Vietnam AI Law 134/2025
147. RTHK, 1 Mar 2026
148. Digital Policy Alert, scope of Vietnam's Law on AI
149. Duane Morris blog, Mar 2026
150. Real Estate Asia, Johor on track for 1 GW data-centre market
151. Focus Malaysia, Malaysia's permit rule for US AI chips
152. CO/AI, Malaysia restricts US AI chips for Chinese data centres
153. Wikipedia, 2026 in AI
154. Social Samosa, France and Malaysia join India in condemning Grok deepfakes
155. Wikipedia, Danantara
156. ANTARA homepage (Bali SEZ headline only)
157. Ministry of IT and Telecommunications, Pakistan, homepage
158. MoITT, Islamabad AI Declaration
159. ICT Division, Government of Bangladesh
160. Anthropic, Fable 5 and Mythos 5 access update, 12 Jun 2026
161. Hugging Face, Hy3
162. Hugging Face, StepFun
163. Qwen3.8-Max licence
164. DeepSeek Anthropic API guide
165. Digital Policy Alert, CAC generative-AI filing-list tracker (read 11 Oct 2026)
166. Hugging Face API, XiaomiMiMo (read 11 Oct 2026)
167. FPF, Understanding Japan's AI Promotion Act, 5 Jul 2025
168. Al Jazeera: US lifts restrictions on powerful AI models Fable, Mythos, Anthropic says (1 Jul 2026)
169. Anthropic Economic Index data, June 2026 release (Hugging Face)
170. Stanford AI Index 2026 report (PDF)
171. MiMo-V2.6-Pro-RL model card (Hugging Face)
172. Xiaomi MiMo API pricing
173. MiniMax pay-as-you-go pricing
174. Z.ai pricing
175. Kimi API pricing
176. NIST CAISSI (formerly CAISI)
177. Adnkronos (Jiji), Japan adopts first AI Basic Plan, 23 Dec 2025
178. Adnkronos (Jiji), draft plan targets, 6 Dec 2025
179. TechCrunch, Sakana AI Series B, 17 Nov 2025
180. Presencis, Korea AI Basic Act timeline (secondary)
181. koreapro.org, sovereign-model second round (single source)
182. MSIT, English home page
183. Anthropic, Claude API data residency
184. Google Cloud, data residency
185. industry.gov.au, Australia establishes AI Safety Institute (headline)
186. Digital Policy Alert, Vietnam decree on AI authority governance
187. OECD.AI listing, Indonesia (secondary)
188. OECD.AI listing, Thailand (secondary)
189. OECD.AI listing, Philippines (secondary)
190. OECD.AI listing, Pakistan (secondary)
191. Tech For Good Institute, Indonesia AI developments, 10 Jul 2026