Official sources were checked on 26 October 2026, using applicant-guide Rev 3, effective 1 July 2026. This is a preparation method, not a determination of eligibility, approval timing or an individual project's costs. An application button does not mean resources have been awarded to a particular team.
Read “Up to 70%” With the Price and Approval Conditions
The Digital Policy Office describes a three-year scheme supporting eligible users of Cyberport's Artificial Intelligence Supercomputing Centre (AISC). The normal subsidy is up to 70% of the service list price. Cyberport administers the relevant application and usage arrangements. Digital Policy Office overview
“Up to” is not a guaranteed rate for every application. Nor does it mean 70% of all the company's AI costs. Data preparation, staff, other cloud services, product integration and supporting infrastructure need separate budgeting against actual quotations and the approved scope.
Start with a defined task: a fine-tuning run, a bounded inference evaluation or a reproducible set of research experiments. Reserving the largest possible resource and then looking for work to fill it reverses the planning process.
Identify the Applicant Before Writing the Proposal
The public FAQ includes local institutions and R&D centres among the applicant categories. The AI-startup category includes incorporation under the Companies Ordinance within seven years, a specified Hong Kong connection, relevant AI work, and fewer than 250 employees globally. Hong Kong registration or an incubator label alone does not settle eligibility. Official AISS FAQ
A university team should establish whether the applicant is the university, an appropriate research unit or a separate company, and who has authority to make commitments. University support for research does not automatically transfer the university's applicant status to a spinout. Support under another programme does not establish that all AISS conditions are met either.
The entity question also affects budgeting, procurement, service agreements and results management. Resolve it before discovering that financial evidence belongs to a different entity or that the intended users and signing arrangements do not align.
Plan Resource Units and Months From the Current Guide
Rev 3, section 5.1, states a minimum application of one H800 GPU core, meaning one-eighth of an H800 GPU card, for one month. Actual resources remain subject to assessment. The older “one card for three months” summary should not be treated as the current guide's only starting point. Current applicant guide
At this check, the website FAQ and its linked v2/2025 FAQ still state one card for three months, which differs from Rev 3. Use the explicitly versioned current guide for initial budget discussions, but confirm actual application units, resources and prices with the administrator before committing payment. Do not merge the two versions into an invented set of conditions. Small usage and other applications also have different documentation and review routes. Official AISS entry
Engineers should turn “we want GPUs” into a reviewable task description: model and dataset scale, training or inference, expected runtime, storage and environment dependencies, and the basis for those estimates. A hardware name alone does not establish model throughput or completion time; this article makes no such performance claim.
Connect the Application to the Experiment Schedule
| Preparation area | Facts to explain | An inadequate substitute |
|---|---|---|
| Research or product question | What will be tested and how results will be judged | A broad ambition to “develop a large model” |
| Data readiness | Permitted data, preparation state and access rights | Obtaining data after securing compute |
| Computing requirement | Measured basis, units, months and dependencies | Treating the maximum resource as the minimum need |
| Cash and supporting costs | Costs outside the award and payment arrangements | Calculating the subsidy without budgeting cash |
| Delivery and usage records | Milestones, owners, runs and results | Treating activation as completed research |
This is an editorial planning aid, not the official scoring sheet. Evidence should reflect real capabilities and progress. Do not invent datasets, partners, papers, customers or computing-use records.
For example, a team may need to prepare an image dataset before model experiments can begin. Connecting the resource start to that readiness is more useful than promising immediate use. This is a scheduling example, not an approved project's arrangement or a claim that resources can be deferred indefinitely.
Approval Is Followed by Payment, Use and Reporting Duties
The guide connects activation to the agreement, prepayment and resource availability. Lower actual usage does not automatically produce a refund. It also sets out continuing usage, progress and final-report obligations. Assign responsibility for those requirements and dates rather than making production plans around an application number. Guide sections 9–12
An application number, approval notice, service agreement, activated resource and project result are different evidence stages. University or company procurement, data-security and payment processes should be aligned early. “Applied” should not be reported as “computing capacity obtained”.
Projects already receiving public support need a separate review of overlapping scope and duplicate subsidy, with the relevant funders. Do not assume that every public research grant conflicts, or that all support can be combined. Individual scope and expenditure decisions belong to the administrators and appropriate professional review.
Separate Research, Engineering and Commercial Responsibilities
For teams moving from Hong Kong research towards GBA engineering or international markets, compute is one resource among several. The useful preparation is a shared plan connecting experiment objectives, data conditions, owners and subsequent validation. Our guide to company–university collaborative research in Hong Kong addresses a related division of work.
An early conversation with Jimhang Capital can start with a non-confidential project outline and resource questions. Do not send raw customer data, credentials, private pitch decks or university material that cannot be disclosed. Contact justin@jimscapital.cn. This is editorial research, not a personal statement by Justin Zhan. It does not suggest that Jimhang Capital is a scheme agent or official appointee, or guarantees an award, model performance or financing. Actual applications and conditions require current official documents and case-specific review.