Analysis date: July 2027 | FX: 1 USD = 7.8 HKD, 1.35 SGD, 150 JPY, 1,350 KRW, 83 INR, 0.78 GBP, 1.52 AUD, 0.92 EUR
This analysis compares 15 government AI compute subsidy programs across Asia, the USA, the UK, the EU, and Canada. Programs are evaluated on five dimensions: scale (total funding), model (how subsidies are delivered), accessibility (who can apply), outcomes (measurable results), and innovation quality (breakthrough impact per dollar spent).
Key finding: Smaller programs with the right design can outperform larger programs. The USA NAIRR (OC cost: $35M) produces world-class medical, scientific, and materials research outputs that dwarf what programs spending 10-100x more deliver. Korea AIVP (~$19M/yr) achieves quantifiable 30% sales growth and 45% productivity gains for SMEs. Singapore ECI ($111M) institutionalizes AI across 1,000 enterprises. By contrast, HK AISS ($385M) — one of the largest programs relative to population — still shows limited public outcomes 18 months in.
| Program | Country | Total Funding | Duration | Subsidy Terms | Notable Outcomes |
|---|---|---|---|---|---|
| AI Subsidy Scheme (AISS) | Hong Kong | HK$3B (~US$385M) | 3 yrs (2024-27) | Up to 70% AISC list price (up to 90% in exceptional cases); min 1 H800 for 3 months | Rolling applications since Oct 2024; successful applicant list published Feb 2026; limited public outcomes visible |
| Enterprise Compute Initiative (ECI) | Singapore | S$150M (~US$111M) | 1 yr per cohort, staggered | Cloud credits: Google S$200K training, AWS S$350K base + S$600K complementary, Microsoft S$250K Azure + S$700K co-dev, Oracle S$250K per company / S$1.9M enterprise. Consultancy: 70% of actual costs, cap S$105K per company (total consulting cap S$150K) | ~1,000 firms paired; 70+ AI Centres of Excellence launched by May 2026; Visa, DP Architects, Ocean Network Ex (AI Trailblazers); 70% of SG companies adopted AI; projected 3% GDP growth contribution |
| RIE AI Research Fund | Singapore | S$1B (~US$741M) | 5 yrs (2025-30) | RIE grants via NRF; S$500M for HPC specifically; 400% tax deduction on qualifying AI spending (capped S$50K) | MERaLiON v2 LLM (90,000+ global downloads); Sovereign healthcare AI "Enigma" for clinical trials |
| METI AI Compute Subsidies | Japan | Y72.5B (~US$483M) | 2024-ongoing | Up to 70% of project costs under Economic Security Promotion Act | 5 AI cloud infrastructure projects approved; designated as "specified critical products" for national security; part of Y10T (~$66B) broader AI/semiconductor investment |
| ABCI 3.0 | Japan | Y36B+ (~US$238M) | Jan 2025-ongoing | Free access via application; priority for national institutes, universities, startups | PLaMo 2.0 (31B, trained from scratch), Swallow LLM series, LLM-jp-4-32B-A3B (beat GPT-4o on Japanese MT-Bench 7.82 vs 7.29), Rakuten AI 3.0 (700B MoE, Japanese MT-Bench 8.88), Namazu, cotomi; #4 global HPL-MxP; #15 TOP500; 60+ foundation model projects active; 3,000+ users since original ABCI launch (2018) |
| AI Voucher Programme (AIVP) | South Korea | K27.6B/yr (~US$19M) | Annual since 2020 | Up to 80% SME / 70% mid-sized; cap K200M/project (~US$138K); K230M for AI Semiconductor track; in-kind contributions accepted (effective cash burden 2-5%) | 1,000+ projects since 2020; 2,093+ registered vendors; K430.3B linked sales; K48.4B linked exports; K69.7B retained investment; 864 FTE jobs created; independent evaluation shows +6pp employment, +30% sales growth, +37.6-45% productivity gains across cohorts |
| AI Computing Center | South Korea | K4T (~US$2.9B) | To 2030 | Govt + private co-invest via SPC; Samsung SDS leading bid; K19.8B (2025) for HPC leasing support | Target: 15,000 GPUs; 2,832 GPUs already allocated to 52 national AI projects (2026); still under construction |
| NAIRR | USA | $35M (OC) + pilot aggregation | Jan 2024-ongoing | Free access via public-private partnership; 14 federal agencies + 28 private partners; allocations: 50K-1.4M GPU hours | 820+ projects across 50 states + DC + PR; 82 classroom awards; 23 infrastructure projects. Protein function prediction, Cervical cancer radiation planning (60min-><4min), Surya space weather model, Digital twin of US water cycle, High-entropy alloy discovery, Molecular simulation acceleration, Humanoid farming robotics, Copyright study in Science magazine |
| IndiaAI Compute Pillar | India | R10,372 Cr (~US$1.24B) | Multi-year | Up to 40% reduced cost; 10 empaneled cloud providers via competitive tender | 38,000+ GPUs deployed (4x initial 10K target); 3,100+ AI startups; India jumped from #7 to #3 globally in AI capability; $679.8M startup funding Q1 2026; Neysa raised $600M; BharatGen AI (gov-funded multimodal LLM, Jun 2025) |
| AI Research Resource (AIRR) | UK | Up to P2B (~US$2.54B) | 2023-2030 | 3 access routes: Rapid Access (up to 20K GPU hours, 3 months, micro/SMEs); Gateway (up to 10K GPU hours, 3 months, first-time users); Innovator (50K-150K GPU hours, 6 months, larger projects). Plus P500M Sovereign AI Fund. Free access, no cost sharing. | Isambard-AI: #11 TOP500, #4 Green500; Dawn: 1,024 Intel DC GPU Max 1550; Cardiomyopathy protein modelling; UK LLMs in Welsh, Gaelic, Cornish; Skin cancer detection bias identification; AI agent security modelling; UK-wide pollution modelling at 1sqm resolution; Agricultural AI; Prince of Wales visited Bristol facility |
| National AI Plan | Australia | A$460M+ (~US$303M) | 2024-ongoing | AI Capability Centres: A$44M for 4 centres (up to A$11M each); GovAI Chat sovereign AI platform in trials; ATSE recommends up to A$5B for sovereign AI compute | 4 centres being established; 15 data centre projects worth A$51.9B endorsed through IDA; GovAI Chat in trials for APS; $19B annual value target for public sector by 2030; still early-stage |
| GenAI4EU / EuroHPC | EU | e93.5B Horizon Europe (~US$101B); e7B EuroHPC 2021-27 | 2021-27 | Project-based grants via Horizon Europe calls; AI Factories: free compute for SMEs/startups (Playground: 2 working days; Fast Lane: 50K GPU-hours, 4-day approval); AI Gigafactories via InvestAI regulation | 14 supercomputers, 19 AI Factories, 13 AI Factory Antennas, 10 quantum systems, 58 R&I projects, 38 participating states; 76 expressions of interest for AI Gigafactories across 16 member states. JUPITER (24K NVIDIA GH200, ~90 ExaFLOPS); MareNostrum 5 (4,480 NVIDIA H100, ~1 ExaFLOPS); Largest single compute allocation in EU history |
| Canadian Sovereign AI | Canada | $2B total (~US$1.45B) | 2024-27 | Three-pronged: AI Compute Challenge ($700M for private data centres); Public Supercomputing Infrastructure ($1B via SCIP Program); AI Compute Access Fund ($300M for compute subsidies); Near-term $200M augmenting existing infrastructure | Cohere $240M data centre investment finalized (Mar 2025); SCIP CFP launched Jan-Feb 2026; Access Fund applications closed Jul 2025; $2.8M Northern AI projects; 1,000+ stakeholder consultations; sovereign data stays in Canada |
| TAIWAN AI RAP | Taiwan | TWD 31.1B (~US$1B) for 10 AI initiatives (2026) | Pilot May 2025; ongoing | Free GPU compute (~NT$650K / US$20K credits per team); low-code AI dev platform; NVIDIA Taipei-1 (512 H100) free access; Compute Alliance (Hon Hai, Wistron, AMD, NVIDIA, etc.) | TAIDE sovereign LLM (Llama-based); Taipei-1 supercomputer operational; RAP platform launched; Compute Alliance formed; 2040 target: TWD 1.5T AI output, 500K jobs |
| Local Gov Compute Vouchers (算力券) | China (municipal/provincial) | ~US$200M+ aggregated across cities | 2023-ongoing | Vouchers subsidizing 80% of AI rental fees; Shanghai CN¥600M (~US$84M) + CN¥100M LLM fund; Zhejiang up to ¥8M RMB (~US$1.1M) per company; Chengdu CN¥100M; plus model vouchers (模型券) for API access | Decentralized approach across 10+ cities; compute utilization up from 20-30% idle rate; SME access to data centers; most aggressive per-firm subsidy globally; paired with national EDWC infrastructure strategy |
Baseline for comparison.
~30 approved projects (data.gov.hk, as of July 2026):
Projects span medical/synthetic biology, large language models, embodied AI, generative AI, materials science, environmental monitoring, and enterprise AI. Full list in open data CSV.
Breakdown by category: 11 local institutions (47%), 2 R&D centres (11%), 1 government bureau (6%), 1 enterprise (6%), 1 start-up (6%) — heavily academic, minimal commercial uptake
Breakdown by research area: 5 medical/synthetic biology, 4 large language models, 2 LLM applications, 1 embodied AI, 1 generative AI, 1 AI agents, 1 new materials, 1 environmental, 1 connectivity
No public detail on individual project outcomes, breakthroughs, or economic impact metrics
Japan AI Strategy Overview:
ABCI 3.0 (AI Bridging Cloud Infrastructure):
Community: 82 Classroom awards; 23 infrastructure projects; 2026 Annual Meeting; Slack workspace; monthly newsletters
Not a single national program but a coordinated decentralized strategy across 10+ cities and provinces, encouraged by the National Development and Reform Commission (NDRC) and tied to the national "Eastern Data, Western Computing" (东数西算) infrastructure plan.
| City/Province | Budget | Subsidy | Notes |
|---|---|---|---|
| Shanghai | CN¥600M (~US$84M) | Up to 80% of AI rental fees | Plus CN¥100M earmarked for LLM training; most aggressive municipal program |
| Zhejiang | Not disclosed | Up to ¥8M RMB (~US$1.1M) per company | Home province of DeepSeek; highest per-firm cap globally |
| Beijing | Not disclosed | Compute vouchers + model vouchers | AI+ action plan; shared computing platform for LLM training; model vouchers (模型券) for API access + pre-trained model licensing |
| Chengdu | CN¥100M (~US$14M) | Subsidized data center access | First city to pilot compute vouchers (2023); targeted at research institutions |
| Shandong | CN¥30M + CN¥1B planned | Subsidized compute rental | Province-level; CN¥1B additional for AI infrastructure buildout |
| Shenzhen | Not disclosed | Compute vouchers | 4,000 PFLOP/s intelligent computing center target; AI action plan |
| Henan, Ningbo, others | Not disclosed | Compute vouchers | Multiple additional cities launching voucher schemes |
A parallel innovation: instead of subsidizing raw compute (to train models), model vouchers subsidize buying AI services — API access to existing models or licensing of pre-trained models. Beijing pioneered this approach. This is a distinct policy tool not seen in other jurisdictions.
| Program | Key Achievements |
|---|---|
| USA NAIRR | Protein function prediction (published), Cervical cancer radiation planning (60min->4min, published), Surya space weather model (published), Copyright study in Science, Digital twin of US water cycle, Humanoid farming robotics |
| UK AIRR | Cardiomyopathy protein modelling; UK-wide 1sqm pollution tracking; Skin cancer bias detection; AI agent security; Minority language preservation (Welsh, Gaelic, Cornish) |
| Program | Key Achievements |
|---|---|
| Japan ABCI 3.0 | 9 Japanese LLMs spawned: PLaMo 2.0 (31B), Swallow, LLM-jp-4-32B-A3B (beat GPT-4o), Rakuten AI 3.0 (700B MoE, MT-Bench 8.88 best-in-class), Namazu, cotomi; #4 HPL-MxP; #15 TOP500; 60+ foundation model projects |
| IndiaAI | BharatGen AI (gov-funded multimodal LLM); India from #7 to #3 globally in AI; 3,100+ AI startups; $679.8M startup funding Q1 2026 |
| Canada | Cohere (world-leading LLM company) — $240M data centre investment |
| Singapore RIE | MERaLiON v2 (90K+ downloads, 8 languages, emotion recognition); Enigma healthcare AI |
| Taiwan AI RAP | TAIDE sovereign LLM (Traditional Chinese-optimized); NVIDIA Taipei-1 supercomputer; Compute Alliance pooling public-private resources; 2040 target: TWD 1.5T AI output |
| Program | Key Achievements |
|---|---|
| Singapore ECI | 1,000 companies paired; 70+ AI Centres of Excellence; Visa, DP Architects, Ocean Network Ex; 70% of SG companies adopted AI; 3% GDP contribution |
| Korea AIVP | K430B linked sales, K48.4B exports, 864 FTE jobs; +30% sales growth, +6pp employment, +37.6-45% productivity gains (independently validated via DID+TWFE); Budget declining as market created |
| Program | Status |
|---|---|
| EU EuroHPC | 14 supercomputers, 19 AI Factories — massive scale but individual project outcomes less visible |
| Korea AI Center | 2,832 GPUs to 52 projects, early-stage |
| Japan METI | 5 infrastructure projects approved, still building |
| Australia | 4 capability centres being established; GovAI Chat in trials |
| HK AISS | Rolling applications since Oct 2024; applicant list published Feb 2026; limited public outcomes; 18 months in |
| Canada (infrastructure) | SCIP CFP launched Jan 2026; Access Fund in review |
| China Local Gov Vouchers | 10+ city programs; Shanghai CN¥600M; Zhejiang US$1.1M per firm; voucher + model voucher dual system |
| Model | Programs | Effectiveness |
|---|---|---|
| Free access (no cost sharing) | Japan ABCI, UK AIRR, USA NAIRR, Taiwan AI RAP | Highest — removes all barriers; enables widest participation |
| Voucher/cap per project | Korea AIVP (K200M), Singapore ECI (S$105K consultancy), China compute vouchers (up to US$1.1M per firm) | High — creates two-sided marketplace; measurable ROI |
| % of compute cost | HK AISS (70-90%), IndiaAI (40%), Japan METI (70%) | Moderate — still requires user budget; creates selection bias |
| Cloud credits | Singapore ECI (S$200K-S$1.9M per CSP) | Moderate — vendor-locked to specific CSP |
| Research grants | Singapore RIE, EU Horizon | High for research — slow for enterprise adoption; high per-project scale |
| Rank | Program | Total |
|---|---|---|
| 1 | EU Horizon Europe | ~$101B |
| 2 | South Korea AI Center | ~$2.9B |
| 3 | UK AIRR | ~$2.54B |
| 4 | Canada Sovereign AI | ~$1.45B |
| 5 | IndiaAI Compute | ~$1.24B |
| 6 | Taiwan AI RAP (total AI budget) | ~$1B (2026) |
| 7 | Singapore RIE AI | ~$741M |
| 8 | Japan METI | ~$483M |
| 9 | HK AISS | ~$385M |
| 10 | Australia AI Plan | ~$303M |
| 11 | Japan ABCI 3.0 | ~$238M |
| 12 | China Local Gov Vouchers (aggregate) | ~$200M+ (partial) |
| 13 | Singapore ECI | ~$111M |
| 14 | USA NAIRR OC | ~$35M (pilot aggregation much larger) |
| Program | Evaluation Framework | Public Data |
|---|---|---|
| Korea AIVP | Independent GRI evaluation (DID+TWFE); NIPA internal metrics | Full metrics published (sales, exports, jobs, productivity) |
| Singapore ECI | DISG assessment; Morgan Stanley analysis | 1,000 firms, 70+ CoE, named success stories |
| USA NAIRR | XDMoD metrics portal; project highlights; 2026 Annual Meeting | 820+ projects, detailed case studies, open-source code |
| UK AIRR | DPR peer review; DSIT monitoring | Use cases published; TOP500 rankings |
| Japan ABCI | OECD.AI documentation; benchmark rankings | 9 LLMs, benchmark scores, project counts |
| IndiaAI | Government announcements | GPU counts, startup numbers, rankings |
| Canada | Stakeholder consultations | Cohere investment, CFP details |
| EU EuroHPC | EuroHPC JU reporting | Factory counts, benchmark data |
| HK AISS | Committee vetting; progress reports to DPO | Applicant list PDF only; no economic impact metrics |
| Australia | Capability centre evaluation | Early-stage; limited outcomes |
| Korea AI Center | MSIT reporting | GPU allocation numbers only |
| Japan METI | METI/Economic Security Act | Project approvals only |
| Taiwan AI RAP | NSTC/NCHC platform metrics | Pilot stage; limited public outcomes |
| China Local Gov Vouchers | City-level reporting (varies) | Opaque — few programs publish outcomes or independent evaluations |
| Model | Programs | Effectiveness |
|---|---|---|
| Free access (no cost sharing) | Japan ABCI, UK AIRR, USA NAIRR, Taiwan AI RAP | Highest — removes all barriers; enables widest participation |
| Voucher/cap per project | Korea AIVP (K200M), Singapore ECI (S$105K consultancy), China compute vouchers (up to US$1.1M per firm) | High — creates two-sided marketplace; measurable ROI |
| % of compute cost | HK AISS (70-90%), IndiaAI (40%), Japan METI (70%) | Moderate — still requires user budget; creates selection bias |
| Cloud credits | Singapore ECI (S$200K-S$1.9M per CSP) | Moderate — vendor-locked to specific CSP |
| Research grants | Singapore RIE, EU Horizon | High for research — slow for enterprise adoption; high per-project scale |
| Rank | Program | Total |
|---|---|---|
| 1 | EU Horizon Europe | ~$101B |
| 2 | South Korea AI Center | ~$2.9B |
| 3 | UK AIRR | ~$2.54B |
| 4 | Canada Sovereign AI | ~$1.45B |
| 5 | IndiaAI Compute | ~$1.24B |
| 6 | Taiwan AI RAP (total AI budget) | ~$1B (2026) |
| 7 | Singapore RIE AI | ~$741M |
| 8 | Japan METI | ~$483M |
| 9 | HK AISS | ~$385M |
| 10 | Australia AI Plan | ~$303M |
| 11 | Japan ABCI 3.0 | ~$238M |
| 12 | China Local Gov Vouchers (aggregate) | ~$200M+ (partial) |
| 13 | Singapore ECI | ~$111M |
| 14 | USA NAIRR OC | ~$35M (pilot aggregation much larger) |
| Program | Evaluation Framework | Public Data |
|---|---|---|
| Korea AIVP | GRI evaluation (DID+TWFE); NIPA internal metrics | Metrics published via OECD case study (sales, exports, jobs, productivity — NIPA source) |
| Singapore ECI | DISG assessment | 1,000 firms, 70+ CoE, named success stories |
| USA NAIRR | XDMoD metrics portal; project highlights; 2026 Annual Meeting | 820+ projects, published papers in Science, Nature, Bioinformatics Advances |
| UK AIRR | DPR peer review; DSIT monitoring | TOP500 rankings; use cases published |
| Japan ABCI | OECD.AI documentation; benchmark rankings | 9 LLMs, MT-Bench scores, TOP500 rankings |
| IndiaAI | Government announcements | GPU counts, startup numbers (PIB-verified) |
| Canada | Stakeholder consultations | Cohere investment details, CFP documents |
| EU EuroHPC | EuroHPC JU reporting | Factory counts, benchmark data (official site) |
| HK AISS | Committee vetting; progress reports to DPO | ~30 approved projects in CSV; no economic impact metrics |
| Taiwan AI RAP | NSTC/NCHC platform metrics | Pilot stage; limited public outcomes |
| China Local Gov Vouchers | City-level reporting (varies) | Opaque — few programs publish outcomes or independent evaluations |
| Australia | Capability centre evaluation | Early-stage; limited outcomes |
| Korea AI Center | MSIT reporting | GPU allocation numbers only |
| Japan METI | METI/Economic Security Act | Project approvals only |
HK AISS: aisc.cyberport.hk/aiss/; digitalpolicy.gov.hk; data.gov.hk; 2025-26 Budget Speech
Singapore ECI: disg.gov.sg/enterprise-compute-initiative/; OECD.AI; edb.gov.sg
Singapore RIE: OECD.AI; IMDA (May 2025)
Japan METI: METI (Apr 2024); OECD.AI; GENIAC Cycle 4 (Jun 2026)
Japan ABCI 3.0: abci.ai/en/; OECD.AI (Jul 2026); Rakuten AI 3.0
Korea AIVP: OECD case study (Apr 2026); KITIM
Korea AI Center: Samsung SDS; DataCenterDynamics
USA NAIRR: nairrpilot.org/; NSF 25-546; Zenodo proceedings (2026); Science (Apr 2026)
IndiaAI: indiaai.gov.in; OECD.AI; BharatGen launch (Jun 2025)
UK AIRR: UKRI (Jun 2026); Bristol use cases; sovereignai.gov.uk
Australia: industry.gov.au; Tech Council
EU EuroHPC: eurohpc-juropa.eu; EU Digital Strategy; Alice Labs (2026)
Taiwan AI RAP: NCHC; CBI/NSTC (May 2025); BowerGroupAsia; Taipei-1 supercomputer
China compute vouchers: FPRI (Sep 2025); RAND; ENGtechnica (Sep 2025); Tom's Hardware
Canada: ISED; Cohere $240M (Mar 2025); RBC
Korea AIVP independent evaluation: Gyeonggi Research Institute (GRI), DID+TWFE methodology, firm-level panel 2015-2024 — cited in OECD case study (Apr 2026)
Korea AIVP internal evaluation: NIPA management system data (2020-2023) — cited in OECD case study
Japan LLM comparison: lilting.ch (Apr 2026)
Singapore GDP impact: Morgan Stanley report — cited via edb.gov.sg AI round-up (Jul-Sep 2025)
NAIRR Zenodo annual report: zenodo.org/records/20434486 (2026)
EU market analysis: Alice Labs; Segler Consulting
Converted from markdown for Discord viewing | Source: ~/Desktop/ai_compute_programs_comparison.md