IndiaAI Compute Portal in 2026: A GPU Access Checklist for Indian Startups

Updated 8 November 2026. A practical guide for Indian founders. All examples are illustrative unless stated as official portal information.Why founders should look at the portalAI founders often discover that the model is not the biggest challenge. Compute bills, experiment design and procurement delays can decide whether a prototype becomes a usable product. India's IndiaAI Compute Portal offers a route to request computing infrastructure and related services, subject to eligibility, verification, approval and payment. This guide is for a founder who needs to evaluate that route without assuming free GPUs or automatic admission. The portal is a government-backed access mechanism, not a substitute for a technical plan or a reliable product hypothesis. Before applying, identify the exact workload, expected hours, data requirements, delivery milestone and maximum cash you can spend. A clear application starts with a clear experiment.What the official portal currently saysThe official IndiaAI Compute Portal lists tracks for startups and MSMEs, early-stage startups, researchers, students, IndiaAI fellows and government entities. The displayed startup and MSME criteria include DPIIT recognition, AI or ML experience, and annual revenue of ₹50 lakh or funding of ₹100 lakh. An early-stage startup track lists four years of AI or ML experience, including research, and a letter from the IndiaAI CEO or an officer of Joint Secretary rank or above. These are portal-listed criteria, not a promise that every company with one document qualifies. The portal also asks applicants to review the detailed eligibility tab for their category. Do not treat a funding pitch, accelerator acceptance or Udyam certificate alone as confirmation of eligibility.Application steps, from account to allocationThe portal describes sign-in through government identity services such as DigiLocker, e-Pramaan or Jan Parichay. An applicant completes the registration form and uploads category-specific documents. Designated officials verify the registration before the applicant can request compute. The compute request needs a project proposal explaining technical approach, expected impact and a draft bill of materials. The portal says requests below 5,000 GPU hours are auto-approved after the relevant registration and eligibility process, while larger requests go to the project management and execution committee. This does not remove the need for a valid account, a suitable provider configuration or available capacity. The portal says requests should be submitted from the first through the twenty-fifth of the month and gives a committee publication timetable for the following month. Recheck the live page before filing.Build a technical request before estimating moneyDescribe the task as a workload, not a general ambition. A founder building a multilingual customer support classifier might need a limited training experiment, evaluation runs and inference testing. A computer-vision team may need a different memory profile and data pipeline. List model architecture, dataset size, expected GPU memory, batch size, training duration, storage, network needs and evaluation checkpoints. Separate one-time training from ongoing inference because their economics differ. Record which assumptions come from benchmarks and which are guesses. Run a small representative test on a paid or existing environment if possible. Its observed throughput can turn an optimistic proposal into a credible bill of materials. Never request a powerful GPU simply because it is available; select capacity for a measurable outcome.Illustrative GPU-hour calculationSuppose a team plans 12 experiments, each using one GPU for 18 hours, followed by 80 hours of evaluation and testing. The first estimate is 12 × 18 + 80 = 296 GPU hours. Add a documented contingency of 20%, or about 59 hours, for a planning total of roughly 355 hours. If a hypothetical listed configuration costs ₹75 per GPU-hour, gross compute would be ₹26,625 before any subsidy, tax, storage or other charges. The ₹75 rate is an example, not a current portal quotation. If a subsidy were approved at an illustrative 30%, the compute component would fall by ₹7,987.50 to ₹18,637.50. The actual subsidy, eligible charges and invoice treatment depend on the live portal and approval. A founder should maintain separate columns for quoted price, approved support, non-compute charges and final payable amount.What to compare before choosing a providerCompare available GPU types on memory, performance for your framework, location, data transfer, storage cost, instance minimums, support and cancellation conditions. A cheaper hourly price can become more expensive if a workload takes twice as long. Ask whether a provider supports the libraries and drivers your team already uses. Check how checkpoints and data are preserved when an instance stops. Calculate total experiment cost as compute hours multiplied by the relevant rate, plus storage, data movement and other billed services. Check whether support, taxes and idle time are charged. The official portal notes that approved users still bear costs beyond the subsidy and pay service providers according to their billing cycles. A subsidy reduces an eligible bill; it does not make cost control optional.The approval clock mattersThe portal says service approvals are valid for 30 calendar days and usage must start within that period to avoid expiration. This matters if a founder applies before the dataset, code and evaluation process are ready. Schedule the application around engineering readiness. Assign one person to monitor approval messages, confirm provider access, test credentials and start a small job promptly. Keep a record of approvals and expiry dates. For a large request, plan for committee review rather than building your launch timeline around immediate allocation. If a deadline matters to a paying customer, create a fallback compute plan with a realistic budget. Government-backed access should improve economics, not create a single point of operational failure.Prepare a founder-ready document packMake a shared folder containing company registration and recognition documents applicable to your track, founder identity records, evidence of relevant AI experience, financial or funding evidence where requested, project summary, technical architecture, expected GPU hours, rate assumptions, data-management plan and a named technical contact. Use the portal's current checklist as the authority on mandatory documents. Write a one-page project note explaining the customer problem, why compute is required now, what you will measure and what the next decision will be. A strong outcome is specific: for example, reducing a model's evaluation error on a held-out dataset or demonstrating a usable inference latency. Avoid promises such as building a national AI platform in one month without supporting milestones.A 30-day execution planDays 1–5: Define the use case, collect representative data legally and choose a baseline model. Days 6–10: Benchmark a small sample, estimate memory and compute requirements, and review the portal's eligibility criteria. Days 11–15: Prepare documents, draft the bill of materials and verify pricing. Days 16–20: Submit the registration or compute request at the appropriate stage and record the status. Days 21–25: Prepare reproducible scripts, evaluation datasets and monitoring while awaiting the required approvals. Days 26–30: If access is granted, run a limited pilot and review actual spend. If access is pending, keep the workload ready and update the timeline. Do not treat this plan as a guaranteed approval schedule.Risk controls for founders using subsidised infrastructureUse role-based permissions, multi-factor authentication, restricted IP access where supported, secrets management and a clear record of who can download datasets. The portal describes controls including IP whitelisting, MFA, role-based access and audits, but the applicant still owns many operational decisions. Avoid uploading sensitive customer information without a documented lawful basis, vendor review and appropriate protection. Test with synthetic or anonymised examples where feasible. Limit the number of people who can launch expensive jobs. Set budget alerts outside the portal if needed. Record model versions, datasets and evaluation results so another team member can reproduce the experiment. These steps make a compute grant more useful because they turn temporary infrastructure into durable knowledge.When the portal may not be the best optionA founder using a hosted AI API for a few hundred customer conversations may not need dedicated GPU infrastructure. Compare the total cost of an API, a small managed inference service and rented GPUs before investing time in an application. A company that has not met the portal's eligibility criteria should not plan around receiving subsidised capacity. Teams with urgent customer deadlines may value instant commercial access more than a lower eventual rate. If you need only a demonstration, start with the smallest technically sound option. The goal is to prove demand and performance, not to maximise allocated compute. Revisit the portal when workload volume, model ownership or research needs justify the effort.Common application mistakesThe first mistake is confusing the portal's general announcement with a confirmed individual subsidy. The second is copying a large GPU-hour number from another startup without measuring your workload. The third is overlooking storage, transfer, idle time and taxes. The fourth is applying before code or data is ready, which can waste the approval validity window. The fifth is assuming an auto-approved small request means identity verification and other controls disappear. The sixth is treating a preliminary cost estimate as a provider quote. The seventh is forgetting a fallback plan for production workloads. Fix these issues with a documented experiment, current eligibility evidence and a named person accountable for spend.Frequently asked questionsIs IndiaAI compute free?No. The portal describes subsidised approved projects and user payments for costs beyond the subsidy. Verify the actual rate and support before committing.Can every DPIIT-recognised startup apply?Recognition is only one part of the displayed startup criteria. Review the full current eligibility requirements and your category.Are requests under 5,000 GPU hours guaranteed?The portal describes auto-approval for requests below that threshold after the relevant account and verification steps. It should not be read as guaranteed capacity or waived eligibility.When should we submit?The portal currently specifies the first through twenty-fifth of the month for compute requests. Check the live process and any updates.What if we are pre-revenue?Review the early-stage startup track and its distinct evidence requirements rather than assuming the regular startup track applies.A practical decision to make todayBefore opening an application, write down one sentence describing the result your team must prove with compute. Then calculate the smallest GPU-hour budget that can test it, list the documents you can actually provide and identify one alternative provider. If those three items are clear, use the official portal to check current eligibility and prices. If they are not clear, spend the next week on workload design rather than a rushed submission. This approach gives a founder a defensible experiment whether or not the request is approved.Primary sourcesIndiaAI Compute Portal, eligibility and FAQs; IndiaAI Mission official website.