{"resourceId":"taiwan-moea-taipei1-sixth-call-terms-free-h100-compute-20260602","versions":[{"version":"external-2a413ace492763b7a2fb5fccdb2719de9af7de3061b23f4362d2e3419d94937e","resource":{"id":"taiwan-moea-taipei1-sixth-call-terms-free-h100-compute-20260602","title":"MOEA's sixth Taipei-1 call sets the terms for free NVIDIA H100 and L40 compute: whole-system blocks, non-commercial use, no sensitive data, non-negotiable NVIDIA contracts","organization":"Taiwan Ministry of Economic Affairs, Department of Industrial Technology (A+ programme office); NVIDIA","sector":"Government-administered access to a vendor-run AI supercomputer","geography":"Taiwan","publishedAt":"2026-06-02 (background; newly relevant as MOEA negotiates the next phase)","publicationDate":"2026-06-02","eventDate":"2026-06-02","sourceName":"MOEA Department of Industrial Technology, A+ Enterprise Innovation R&D programme","sourceLabel":"Taipei-1 compute application: sixth call notice","sourceUrl":"https://service.moea.gov.tw/EE514/tw/aiip/50-7605.html","evidenceClass":"standards-guidance","outcomeClass":"emerging","topics":["infrastructure","governance-procurement","data-security"],"finding":"The sixth call, which allocates the remaining time to February 26, 2027, offers Taiwanese companies, schools, research institutes and start-ups free Taipei-1 compute, which NVIDIA shares from July 2024 to February 2027. Capacity is 16 DGX H100 systems (8 GPUs each) and 8 OVX 2.0 systems (8 L40 GPUs each). A team must take all 16 DGX for six weeks (four of them computing, plus onboarding and offboarding) or all 8 OVX for twelve weeks; smaller users must apply jointly. Review weighs commercial potential and large-compute experience. Use must be non-commercial. Users must not process or enter proprietary or sensitive information. NVIDIA's NDA and user agreement are standard and cannot be changed. NVIDIA headquarters operates the system, and publicity naming Taipei-1 needs NVIDIA's approval.","sledRelevance":"This is a detailed public rulebook for sharing a vendor's AI supercomputer with universities and firms. States and university consortia designing shared compute can reuse its allocation model and should note its data and contract limits.","evidence":"Primary government call notice with eligibility, capacity, schedule and terms; it does not report usage or results.","architectureImplications":"Whole-system time blocks favour large training runs and force small users to pool projects.","governanceImplications":"Results disclosure follows government requirements, and generative AI use must be disclosed under Executive Yuan guidance.","securityPrivacyImplications":"The ban on proprietary or sensitive data rules out regulated public-sector data.","caveats":"Published June 2, filed as background because the Nov 1 report says the current phase is ending and a next phase is under negotiation. The notice covers this call only.","streamIds":["nvidia"],"roles":{"sales":"This notice is a concrete model for public clients designing shared AI compute with a vendor. It shows who can apply, how time is allocated and what the vendor keeps control of. Useful for states, university systems and research consortia comparing options. Be candid about the limits: no sensitive data, non-commercial use only, and NVIDIA's contracts cannot be negotiated. Those limits matter for any agency use case involving resident or regulated data.","engineering":"Allocation is all-or-nothing: the full 16-node DGX H100 system for six weeks, or the full OVX cluster for twelve, so projects must be ready to scale on arrival. Small teams must combine. Because sensitive data is barred, workloads should use public or synthetic data. NVIDIA runs the platform and can schedule maintenance with two weeks' notice. Engineers advising a similar programme should plan data preparation, checkpoints and job scheduling before the window opens.","delivery":"For a public client adopting this model, delivery work includes the call process, review panel, joint-application rules and post-use reporting. Build in the disclosure duties: results reporting to government, acknowledgment in papers and vendor approval for publicity. Assign an owner to track vendor maintenance notices. These are proposed adaptations; the notice describes rules, not outcomes, and Taiwan's own results are only partly visible in award lists."},"roleDetails":{"sales":{"customerProblem":"Public research users lack access to large GPU systems.","stakeholders":["University research computing","State economic development","Vendor"],"discoveryQuestions":["Can users bring sensitive data?","Who controls publicity and results?"],"valueHypothesis":"A clear allocation model widens access to scarce compute.","potentialEngagement":"Design review for a shared-compute programme.","unsupportedClaims":["Any usage or outcome figures","Suitability for regulated data"]},"engineering":{"fit":"Relevant to shared training-scale compute.","architectureIntegration":"Scheduled whole-system blocks on DGX H100 and OVX L40 clusters.","prerequisites":["Job-ready code","Public or synthetic datasets"],"constraints":["No proprietary or sensitive data","Six- or twelve-week blocks only","Non-commercial use"],"security":"Sensitive data is barred; vendor runs the system.","proofOfValue":["Proposed: dry-run jobs before the window","Proposed: checkpoint plan for maintenance windows"]},"delivery":{"work":"Run the call, review and reporting process.","dependencies":["Vendor contracts","Review panel"],"ownership":"Programme office.","skillsAdoption":"Train applicants to form joint proposals.","governanceCheckpoints":["Eligibility and data review","Post-use disclosure"],"acceptanceCriteria":["Proposed: every allocated block fully used"],"risks":["Unused blocks","Data-rule breaches"]}},"retrievedAt":"2026-11-02T14:37:09Z","enrichedAt":"2026-11-02T14:43:04Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Universities and start-ups get access to H100-class systems they could not buy.","procurementImplications":"Vendor contracts are fixed; the public side cannot negotiate user terms per project.","operatingModelImplications":"The vendor sets maintenance windows with two weeks' notice and controls publicity.","sourceVerification":{"openedUrl":"https://service.moea.gov.tw/EE514/tw/aiip/50-7605.html","referenceExcerpt":"使用單位不應在該環境中處理和輸入任何專有或敏感資訊。","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}