AI Compute Subsidy Programmes - Global Comparison

Points of interest

Transparent and frequent measurement enables programmes to more effectively achieve Hong Kong's goal of becoming an innovation hub by identifying issues early. Three distinct levels of transparency can be observed across this comparison: recipient transparency (who received support), allocation transparency (what resources each project received), and outcome transparency (what those resources produced). The USA's NAIRR programme is the only programme in this comparison to publish all three systematically: it maintains a project directory with per-project allocations and collects and reports project outcomes.
Peer review complements administrative assessment by leveraging specialist expertise available beyond the programme team, and alleviates its information burden. The UK's AIRR requires each applicant to review up to five other proposals, while NAIRR uses volunteer experts to review applications. Both approaches broaden the domain expertise available to administrators while reducing the programme team's evaluation burden.
Incumbents dominate compute access through greater financial resources and scale, limiting the emergence of promising start-ups and SMEs. Leading programmes use an access ladder which addresses the two major components of access - cost burden and compute allocation - through routes designed for different users and scales. NAIRR offers free Sandbox, Classroom, Start-Up, and Research tracks; AIRR offers free Gateway, Rapid, and Innovator routes; EuroHPC combines free Playground, Fast Lane, and Large Scale access with a pay-per-use commercial route. Korea's national allocation system extends the pattern with a ring-fenced GPU pool for SMEs/start-ups.
Most co-pay based programmes subsidise private cloud usage, while most no-cost programmes run on state-funded compute centres.

Programme	Region	Type	Target	Budget	Mechanism	Terms	Duration	Outcomes
No-cost allocation
NAIRR Pilot	USA	Public-private	USA-based researchers and educators from academic institutions, non-profits, federal agencies, and start-ups/small businesses with federal grants	HK$234M (US$30M) Private companies have now donated ~US$100M in compute since initial launch	Free GPU access. Applications vetted by volunteer experts (peer review). Compute is pooled from universities and research institutes.	Decision time Approximately 2 weeks for Start-Up applications; approximately 1 month for Research applications. Allocation size / cap Start-Up track: up to 2,000 GPU-hours over 3 months. Other tracks vary by opportunity. Payment timing No applicant payment; access is free. Programme details Applications submitted via NAIRR web portal; reviewed by volunteer peers. Tracks: Research: ~1-month approval process 12-month project limit Monthly reviews Institutional email required Start-Up: 2-week approval process 3-month GPU allocation Up to 2,000 GPU-hours Graduate students require a letter of support from their supervisor Classroom: GPU resources for undergraduate/graduate AI courses requiring access to compute for coursework Must be USA-based Sandbox: Exploratory access for anyone No proposal required Source	Jan 2024-ongoing	Scale and uptake 600+ projects across 50 states (NSF). 14 federal agencies + 28 private partners. Published outcomes and transparency Free allocations via peer review, including Start-Up and Classroom tracks; per-project resource allocations are published. UCSD cervical-cancer brachytherapy planning AI (Lance Moore; NAIRR250110 on SDSC Voyager) - planning from hours to under 4 minutes; Brachytherapy journal HONeYBEE multi-modal oncology embeddings (npj Digital Medicine; NAIRR pilot funding) Wider portfolio also spans protein prediction, space weather, and farming robotics (Science / Nature Scientific Data / Bioinformatics Advances)
AI Research Resource (AIRR)	UK	Public	UK-based researchers, research institutes, start-ups, and SMEs; 3 routes to give compute access to different sized projects.	HK$20.9B (£2B)	Free GPU access. Applications vetted by other applicants. All applicants must be reviewers.	Decision time Up to 2 months. Allocation size / cap 10,000-150,000 GPU-hours depending on route. Payment timing No applicant payment; access is free. Programme details Projects submit training jobs to a queue. Running jobs get deducted from the project's allocated GPU-hours. Tracks: Rapid: up to 20K GPU-hours / 3 months (UK micro/SMEs only; no closing date) Gateway: up to 10K GPU-hours / 3 months (first-time users; no closing date) Innovator: 50K-150K GPU-hours / 6 months (closed Jan 2026) All routes require application; no cost to applicants; light-touch review with approvals within 2 months Source	2023-2030	Scale and uptake AIRR-specific project and recipient totals are not published. Its main compute centre reported ~1,000 research projects and 4,000+ users in its first year, but these are not strictly AIRR projects. Published outcomes and transparency Information on AIRR recipients and project outcomes is scarce. University of Oxford - 10,000 GPU-hours, personalised cancer vaccine design from 1B+ datapoints across 33 tumour types (AI for Science Strategy) University of Edinburgh & Imperial College London - generative foundational model of medical images, via Gateway route (AI for Science Strategy) Shuang Ao - 50,000 GPU-hours on Isambard-AI AIRR cluster (RAi UK Annual Report 2025) researching AI uncertainty quantification, automatic failure detection, and safety alignment. Bristol anniversary case work: heart disease, dementia diagnosis, cancer pathology, UK-language LLMs (Welsh/Gaelic/Cornish), air pollution forecasting, dairy farming, AI safety*
Sovereign AI Fund	UK	Public	UK AI start-ups (pre-seed to growth); 5 priority sectors: compute & infrastructure, foundation models, health & life sciences, scientific discovery, trust & safety	HK$5.24B (£500M)	Equity investment + free compute access + visas + grants	Decision time Not published. Allocation size / cap Up to 1 million GPU-hours per start-up. Payment timing Compute is free; separate equity investments are made at market terms. Programme details The Sovereign AI Fund has a reserved allocation of AIRR resources, which it sub-allocates to promising projects. This bypasses the slower, standard AIRR review process and deploys compute quickly. Up to 1M GPU-hours per start-up via AIRR (est. HK$147M / £14M+ deployed in first tranche) £1M-£10M equity cheques at market terms; pre-seed to growth stage Fast-track visa decisions within 1 working day; 10 free visas per company Right of First Refusal on future investment for compute recipients Source	Launched April 2026	Scale and uptake First compute tranche: 3M+ GPU-hours to 6 start-ups. ~30 firms in active discussions for compute access. Published outcomes and transparency First equity investment: Callosum (AI infrastructure orchestration). Named first cohort: First cohort (Apr 2026): equity - Callosum (Cambridge, heterogeneous compute orchestration) compute - Prima Mente (Alzheimer's/Parkinson's biological models), Cosine (sovereign AI for defence & regulated industries), Cursive (DeepMind alumni, behavioural prediction agents), Doubleword (inference governance), Odyssey (multi-modal world models), Twig Bio (CANOPY, engineering biology)
Taipei-1	Taiwan	Public-private	Industry, government, academia, and research institutions across Taiwan	No public figure (NVIDIA-donated compute)	Free GPU supercomputer access via proposal	Decision time Not published. Allocation size / cap Not published. Payment timing No applicant payment; access is free. Programme details Free access to NVIDIA's 512 H100 GPUs + 64 OVX systems, negotiated by the Ministry of Economic Affairs for Taiwanese R&D. Proposals reviewed by expert committee on a rolling basis First CFP April 2024; open to industry, academia, and research Source	Completed Dec 2023; first CFP Apr 2024-ongoing	Scale and uptake No public project count or recipient list. Published outcomes and transparency Very opaque programme outcomes. Found one acknowledgement to the programme in academic publications (Warp-STAR paper). No public project directory published.
EuroHPC AI Factories	EU	Public	EU-based AI start-ups and SMEs (free, Industrial Innovation tiers); public/private researchers (free, AI for Science track); larger industrial users (pay-per-use commercial track). 19 Factories + 13 Antennas across EU and associated countries.	HK$62.6B (€7B) EuroHPC infrastructure 2021-2027. ~HK$6.3B (€700M) GenAI4EU grant envelope (separate GenAI deployment grants). HK$179B (€20B) AI Gigafactories fund (not yet allocated).	Free GPU access + consultancy. Applications vetted by appointed ARC.	Decision time Approximately 2 working days for Playground and 4 working days for Fast Lane; Large Scale uses two cut-offs per month. Allocation size / cap Fast Lane: up to 50,000 GPU-hours; Large Scale: more than 50,000 GPU-hours. Payment timing Free for eligible access tracks; commercial access is pay-per-use. Programme details Applications via EuroHPC access portal; free at point of use. Reviewed by an appointed Access Resource Committee (ARC). Antennas in smaller countries (e.g. UK → HammerHAI) route applicants to a mothership Factory. Tracks: Playground: Limited resources for entry-level users Lightweight ARC review ~2 working days Rolling, no cut-off dates Fast Lane: Up to 50,000 GPU-hours Lightweight ARC evaluation ~4 working days Rolling, no cut-off dates Large Scale: 50,000+ GPU-hours Full peer review with external experts + ARC ranking 2 cut-offs per month AI for Science: full peer review, 28 ERC evaluation panels for reviewer matching Commercial: pay-per-use, no peer review, up to 20% of resources Source · Access Policy (PDF)	2024-2027 (AI Factories); EuroHPC since 2021	Scale and uptake 2,334 projects awarded across all EuroHPC supercomputers (not AI Factory-specific). Published outcomes and transparency No public AI Factory project directory or per-programme recipient list.
Enterprise Compute Initiative (ECI)	Singapore	Public	SG-registered companies: ≥10 staff, ≥2 tech team, prior AI PoC, CEO sponsorship	HK$909M (S$150M)	Free cloud credits + subsidised consultancy	Decision time Not published. Allocation size / cap Provider-specific cloud credits range from S$200,000 to S$1.9 million; consultancy subsidy capped at S$105,000. Payment timing Cloud credits are free; consultancy payment timing is not published and applicants pay 30%. Programme details Cohort-based via DISG; CEO sponsorship required. Cloud credits: Google S$200K; AWS S$350K+S$600K; Microsoft S$250K+S$700K; Oracle S$250K/S$1.9M Government pays 70% of consultancy (capped S$105K); applicant pays remaining 30% Requires ≥10 staff, ≥2 tech team, prior AI PoC, and a permanent AI Centre of Excellence Source	1 yr per cohort, staggered	Scale and uptake No public recipient or project total. Published outcomes and transparency Little information on programme outcomes thus far outside of three case studies (all well-established companies).
Applicant co-payment and voucher subsidies
AI Subsidy Scheme (AISS)	Hong Kong	Public	Universities, R&D centres, government bureaux, AI start-ups (<7yr, <250 staff), strategic enterprises	HK$3B	Subsidised compute rental (applicant co-pay)	Decision time Not published. Allocation size / cap Small Usage track: no more than 1 card for up to 12 months. Standard-track cap not published. Payment timing Applicant prepays its 30% share upfront. Programme details Projects apply via Cyberport, which then submits an application to the AISS committee Government pays 70% of list price (90% exceptional) Applicant prepays remaining 30% upfront Minimum tenancy 1 month; Small Usage track (≤1 card, ≤12 months) with streamlined approval Source	2024-2027	Scale and uptake 31 approved projects, nearly HK$1.4B committed (Jun 2026). Published outcomes and transparency Published research and company awards. Sample projects: Karta-X - HK edge AI / smart city platform; Self-Improving Edge Object Detection System with Synthetic Data Generation and Active Learning NYXLAB - AI-powered SecOps for Managed Security Services (No published outcomes of funded compute thus far) InfiX.ai - PolyU-spinout enterprise GenAI infrastructure (Co-GenAI / model fusion / low-resource training); Collaboration Generative Artificial Intelligence (Co-GenAI) Platform (90% subsidised by AISS) MediConCen - HK InsurTech claims automation; AI application in Document Authentication and Intelligent Data Extraction (Publicly uses Google Cloud / Gemini). No AISS attribution thus far. Kite Plus Limited - Foundation Models Research & Development (LLMs); no public detail on the company or operations found
IndiaAI Compute Pillar	India	Public	Academia, MSMEs, start-ups, research community, students, and government agencies via IndiaAI Compute Portal	HK$8.43B (₹10,372 Cr) IndiaAI Mission total outlay; Compute Pillar is one of seven pillars.	Subsidised compute rental (applicant co-pay)	Decision time Not published; approved credits must start within 30 days. Allocation size / cap Subsidised compute limit not published. Payment timing Post-paid usage; applicant pays at least 60%, and IndiaAI remits its subsidy to providers quarterly. Programme details Applications via IndiaAI Compute Portal. Post-paid through 10 empanelled providers (Intel, AMD, NVIDIA, AWS); IndiaAI remits subsidy quarterly. Up to 40% subsidy; applicant pays remaining 60% (100% above subsidised amount) Credits must start within 30 days of approval Start-ups have reported delays in formal agreements and IP ownership concerns (May 2026) Source	2024-ongoing	Scale and uptake 190 total projects: 46 start-ups/MSMEs, 30 early-stage start-ups, 78 government, 27 academia, 5 students. 12 companies selected from 500+ proposals for foundation model development (Jan 2026). Published outcomes and transparency 10 start-ups in Station F programme (Paris, 2 cohorts, 4-month immersive, IndiaAI + HEC Paris). Applicant pays ~60% of compute; 5 of 12 had not signed formal agreements as of May 2026. 30 India-specific applications approved by Jul 2025. BharatGen - sovereign multilingual/multi-modal AI (IIT Bombay consortium); ₹988.6Cr under IndiaAI Mission Sarvam 30B & 105B - reasoning models trained from scratch on IndiaAI compute - covered by Forbes Varya - 14B-param distilled video model (Avataar.ai); $0.005/sec; 10-27x cheaper than rivals (Jun 2026)
Shanghai AI Vouchers (模塑申城)	Shanghai	Public	AI enterprises, start-ups, SMEs deploying LLMs; model API consumers; data buyers	HK$1.07B (¥1B)	Compute + model + data vouchers (three voucher types)	Decision time Not published. Allocation size / cap Model and data vouchers: up to ¥5 million each per entity. Compute-voucher cap not published. Payment timing Not published. Programme details Three voucher pools (Jul 2025): ¥600M compute (算力券), ¥300M model (模型券), ¥100M data (语料券). Compute: up to 30% rental subsidy (to 100% with district co-funding) Model: up to 50% of API/deployment costs, max ¥5M/entity Data: up to 30% of purchases, max ¥5M/entity Source (Jul 2025) · 模塑申城 (Dec 2024)	Jul 2025-ongoing	Scale and uptake No published recipient ledger. WAIC 2025 announced ¥1B voucher package. Published outcomes and transparency No project-level outcomes published. Programme targets and start-up incentives include: AI小镇 (AI Town) bonus: 100% subsidised ¥1M compute + ¥1M model + ¥1M data vouchers for qualifying start-ups Original "模塑申城" framework (Dec 2024) targets 100 EFLOPS, 50+ industry corpus benchmarks, 3-5 model incubators
Hangzhou AI Compute Vouchers	Hangzhou	Public	AI start-ups, SMEs, and enterprises in Zhejiang/Hangzhou; home province of DeepSeek (predates programme)	HK$268M (¥250M/yr city-level)	Compute rental vouchers (city-level) + provincial AI vouchers (national treasury bond funding)	Decision time Not published. Allocation size / cap Hangzhou city: up to ¥8 million per company; district schemes have separate caps. Payment timing Provincial vouchers use post-paid reimbursement; timing for city and district schemes is not published. Programme details City and provincial voucher pools, plus district top-ups. Hangzhou city: ¥250M/yr (¥1B over 4 years), 30% subsidy, max ¥8M/company; higher caps for generative multi-modal foundation-model firms Zhejiang province: 10-30% subsidy via national treasury-bond funding; post-paid reimbursement Binjiang district: 50% subsidy via Hangzhou AI Computing Centre, max ¥1M/entity Xiaoshan district: separate ¥100M/yr compute voucher pool Source (Zhejiang) · First batch recipients	2023-ongoing (最早¥50M/yr pilot Jul 2023; expanded Jun 2025)	Scale and uptake Named recipients from first batch. Three companies received vouchers (Apr 2025): 杭州核新软件技术有限公司 + two others. Zhejiang's first 43 "智算云企" named (Dec 2025) include 老板电器 (Robam, ~¥3.5M/yr cloud AI spend). Hangzhou "六小龙" (DeepSeek, Unitree, etc.) were built before vouchers; vouchers now target the next generation. Xinhua: "杭州现象" coverage. Published outcomes and transparency Recipient names are available, but no systematic project outcomes are published.
Shenzhen 训力券 (Training Power Vouchers)	Shenzhen	Public	AI and embodied-intelligence start-ups and enterprises; model trainers and inference users	HK$214M (nearly ¥200M first batch) HK$4.82B (¥4.5B) city-district AI policy total across all channels (separate model vouchers, data vouchers, R&D grants, and AI fund).	Compute rental vouchers - applicant co-pay	Decision time Approximately 60 working days. Allocation size / cap Minimum ¥500,000 annual compute rental; up to ¥10 million per entity annually. Payment timing Applicant contracts with a provider first; subsidy is redeemed after verified use. Programme details Contract with a registered provider first, then apply online (~60 working days). Vouchers offset costs; applicant pays the balance. 50% subsidy (60% if start-up <1 year old) Minimum spend: ¥500K/yr compute rental before eligible Annual cap: ¥10M/entity; unused amounts carry forward No double-dipping: total public subsidy across sources cannot exceed 50% Providers: registered (入库) only; cannot be both provider and recipient Redemption (兑现): post-use verification against completed service, not pre-paid credit Source: 2025 Second Batch Application Guide · First batch announcement	Mar 2025-ongoing	Scale and uptake First batch Mar 2025: nearly ¥200M total, ~40 companies, up to ¥10M each. Published outcomes and transparency Named recipients with disclosed amounts - among the most transparent Chinese municipal programmes. No systematic project outcomes published. 元戎启行 (DeepRoute.ai) - autonomous driving 逐际动力 (LimX Dynamics) - embodied AI / humanoid robots 云天励飞 (IntelliFusion) - edge AI / smart city 鲲云科技 (Corerain) - custom AI chips 睿迈生物科技 - AI × biotech
AI Compute Access Fund	Canada	Public	Canadian for-profit SMEs: <500 FTE, Series A+ or revenue-generating, Canadian R&D team, AI product with commercialisation plan, compute agreement in place. Indigenous unincorporated businesses eligible.	HK$1.71B (up to C$300M) initial fund. HK$3.98B (up to C$700M) Access Fund expansion committed under AI for All strategy (Jun 2026).	Subsidised compute rental (applicant co-pay). ISED-led assessment with external AI expert consultation.	Decision time 60-90 business days. Allocation size / cap C$100,000-C$5 million per project over up to 3 years. Payment timing Quarterly reimbursement claims. Programme details ISED two-stage process: Statement of Interest → Full Application → assessment (60-90 business days) → Contribution Agreement. ISED decides after external AI expert consultation. Government pays 2/3 of Canadian cloud costs, 1/2 of non-Canadian (non-Canadian ineligible after Mar 2027) C$100K-C$5M per project over up to 3 years (through Mar 2028) Non-repayable, conditionally repayable, or repayable (set at assessment) Quarterly reimbursement claims; first call closed Jul 2025 (oversubscribed) Source (Program Guide)	2025-ongoing	Scale and uptake C$66M committed to 44 SME projects (May 2026) across life sciences, health care, energy, manufacturing, agriculture, finance, and transport. First cohort; further awards as assessments finalise. Published outcomes and transparency Named recipients and contribution records are published; measured project outcomes are not yet available. Examples: Variational AI (Vancouver) - AI drug discovery, small molecule synthesis for biotech/pharma licensing (CBC) SimpleDirect (Toronto) - AI task automation (invoicing, scheduling); C$3.015M contribution (Open Government) SenseNet - AI wildfire detection (BetaKit) Spare - AI public transit route optimisation (BetaKit) Separately, the broader Canadian Sovereign AI Compute Strategy includes ~C$1.9B for infrastructure build (AI Compute Challenge + SCIP + near-term augmentation) - data centre and supercomputer construction, not direct SME compute access.
National AI Project GPU Allocation	South Korea	Public	Universities, research institutions, government agencies, and enterprises via national AI project proposals; separate SME/start-up track	HK$7.75B (₩1.46T). Approved May 2025 supplementary budget. Procures GPUs for subsidised allocation; NAIP retains ownership, cloud providers operate the hardware.	Subsidised GPU access and fixed GPU allocation via cloud for enterprises. MSS SME/start-up track is free.	Decision time Approximately 2 months in the first round (applications in April 2026; support from June 2026). Allocation size / cap 16-128 dedicated GPUs per project for 1-12 months. Payment timing Prepaid monthly for paid tracks; MSS SME/start-up access is free. Programme details Applications via aiinfrahub.kr. Three-stage evaluation: expert committees score proposals on strategic importance and impact; adjustment committees set GPU scale via per-project interviews; Science and Technology Ministers' Meeting grants final approval (cross-ministry deliberation, not an independent review body). MSIT (Apr 2026) Providers: NHN Cloud (7,656 B200), Naver Cloud (3,056 H200), Kakao (2,424 B200) Fixed allocation (고정할당): dedicated GPUs, not shared. 16–128 GPUs per project, 1–12 months, prepaid monthly. Idle or misused GPUs repossessed with no refund. Pricing: industry ₩800K/GPU/month (B200), youth start-ups ₩400K, academia free. National AI Project co-pay unpublished; MSS SME/start-up track free. First round: applications Apr 2026, support from Jun 2026. Korea.net (May 2025) · MSS (Apr 2026) · MSIT (Apr 2026)	2025-ongoing	Scale and uptake 52 projects selected across 28 ministries (Apr 2026). First-round allocation of 2,832 GPUs to national projects and 264 GPUs to SMEs/start-ups. Published outcomes and transparency Project-level GPU allocations and intended uses are published; delivered outcomes are not yet available. National project highlights: SDV autonomous driving AI model - MOTIE, 304 GPUs (largest single allocation) Industry-Specific Foundation Model - MSIT, 256 GPUs Korean AI Weather/Climate Model - KMA, 128 GPUs AI Basic Healthcare - MOHW, 80 GPUs AI-Converged Content Production - MCST, 64 GPUs SME/start-up (via MSS): 264 GPUs - 64 for manufacturing AI (production/quality-control) + 200 for start-up incubator across 3 tracks (Strategic AI 85, Industry-Specific 85, Startup for All 30). Free delivery via NHN Cloud. National AI Computing Centre (₩2.9T Samsung SDS consortium data centre in Haenam) is separate infrastructure build under construction; excluded per scope rules.
Private equity and accelerator models
Google AI Futures Fund / KDDI	Japan	Private	AI-native Japanese start-ups; pre-seed to Series A with traction preferred; 5 focus areas: future of work, knowledge, software dev, creativity, entertainment	Up to HK$15.6M (US$2M) equity per start-up + up to HK$2.73M (US$350K) in credits	Equity for cash, compute, and consultancy	Decision time Not published; first selections scheduled for September 2026. Allocation size / cap Up to US$350,000 in credits and US$2 million in equity per start-up. Payment timing Credits are provided with an equity investment; no applicant payment timing applies. Programme details Rolling applications via aistartup@kddi.com; first selections Sept 2026. Up to US$350K in Google AI tools and Cloud credits Early access to Google's latest models + Google/KDDI technical and business support Equity stake required; must use Google AI models in a materially impactful way (not exclusive) Source	Launched July 2026, rolling basis	Scale and uptake No projects announced yet. Published outcomes and transparency Too early to assess; first selections are scheduled for September 2026.