- The short: what’s new in the ~3.8k GPU tranche
- Onboarding flow & thresholds (who gets auto-OK vs PMEC review)
- Pricing, subsidy & the new voucher model
- Practical hurdles: power, cooling, network, datasets
- “Apply” checklist you can copy-paste
The short
- Fresh capacity: IndiaAI’s third tender adds ~3,850 GPUs (first tranche to include Google Trillium TPUs) on top of a pool that crossed 34k GPUs by end-May.
- Pay-as-you-go: Access via the Compute Portal from ~₹67/GPU-hour with up to 40% subsidy for eligible users.
- What’s next: Government has indicated a coupon/voucher-style mechanism alongside a larger secured pool reported at ~38k units.
What’s new in this tranche
Round-3 adds ~3,850 GPUs and the first inclusion of ~1,050 Trillium TPUs, alongside earlier H100/H200/L-class additions. Treat mix as evolving until commercial close & delivery schedules firm up.
| Signal | What it means |
|---|---|
| Trillium TPUs join pool | Training/inference options beyond Nvidia/AMD; check framework support & quotas per project. |
| Pool scale (public) | ~34k+ GPUs by late May; more recent reporting points to ~38k secured capacity. |
| Portal pricing | Baseline ~₹67/GPU-hr; subsidy up to 40% for priority users. |
Onboarding flow (startups, academia, gov) — how approvals work
- Register on the IndiaAI Compute Portal with org/institute email + mobile → OTP verification.
- Submit IDs (auto-verification where possible):
- Students APAAR ID
- Startups DPIIT recognition; MSMEs Udyam ID
- Researchers/Faculty Google Scholar / ORCID / Scopus author IDs
- Project proposal (problem, novelty, national impact, team track). Institute verifiers (students/research) and MeitY Startup Hub CoEs (startups) review.
- Monthly window: Requests collated 1st–25th; PMEC meets end/start of month; results published by the 10th.
- Thresholds that trigger PMEC: requests > 5,000 GPU-hours or > 50 GPUs; below that, category verifiers can approve.
Auto-OK paths Researchers meeting simple h-index/citation bars, and IndiaAI Fellowship awardees, typically see faster allocation post-verification.
Pricing, subsidy & vouchers
- Rack rate: ~₹67/GPU-hour via the portal (roughly one-third typical global rates).
- Subsidy: PMEC may award up to 40% subsidy on approved project BOMs; paid to providers quarterly.
- Voucher/coupon model: Announced to simplify access and cap user costs; integration underway.
Practical hurdles you’ll still hit
Infra frictions
- Power/cooling DC slots depend on local MW & thermal capacity; expect staged delivery.
- Network Egress can dwarf compute costs; budget for object-store and inter-DC traffic.
Data & platform
- Datasets AIKosha (IndiaAI datasets) is live; check licensing & privacy flags.
- Heterogeneous silicon Trillium/Nvidia/AMD/Gaudi mixes → validate framework support early.
Apply checklist (copy–paste)
- Account: Org/institute email + mobile ready.
- ID docs: APAAR (students), DPIIT (startups), Udyam (MSMEs), Scholar/ORCID/Scopus (researchers).
- One-pager: Problem, novelty, target users, milestones, expected GPU-hours, models/tooling stack.
- Team bios: brief CVs + links (code/pubs/patents).
- Data plan: sources, licensing, storage/egress budget; any AIKosha sets.
- Budget: rack at ₹67/GPU-hr; request subsidy with justification; note if you exceed 5,000 hours / 50 GPUs.