India Built the GPU Capacity, But Are Enterprises Using It?

India announced a serious bet on AI infrastructure with the IndiaAI Mission, committing over ₹10,000 crore to build a national compute layer – GPUs that Indian startups, researchers, and enterprises could access at subsidized rates rather than paying global market prices in dollars. On paper, the numbers looked promising: more than 38,000 GPUs listed on the IndiaAI Compute Portal, six empaneled cloud providers, a target of 100,000 GPUs by end of 2026. The intent was clear: to close the gap between India’s AI ambitions and the compute infrastructure needed to back them. 

But somewhere along the way, the gap it hasn’t closed is between capacity that’s listed and capacity that enterprises can actually provision when they need it. Here’s the reason behind it. 

The Scale of the Bet 

The IndiaAI Mission was approved with an outlay of ₹10,371 crore (roughly $1.07 billion) – making it one of the larger sovereign AI infrastructure programs globally. The compute pillar sits at the center of it: a shared national GPU pool accessible through the IndiaAI Compute Portal at subsidized rates of around ₹65 per GPU hour, compared to global cloud rates that typically run several multiples higher. 

For eligible applicants, the pricing makes sense. Academic institutions running Indic-language research, government agencies with sovereign AI cloud in India, and early-stage startups with time to navigate the application process have real reasons to pursue subsidized access. The scheme was designed with those users in mind. 

The problem is that most enterprises aren’t those users.
Why the GPU Capacity Sits Idle 

The friction is a combination of structural constraints that compound on each other. 

Start with disbursement. Of the ₹10,371 crore allocated to the IndiaAI Mission, only around ₹400 crore has actually been released so far. Hardware procurement and deployment timelines follow from that, which partly explains why empaneled providers have faced installation delays despite appearing on the portal. 

Access adds another layer. Eligibility requirements mean not every organization qualifies, and larger compute requests are subject to additional review. Experts warn that a first-come-first-served approach could systematically favor institutions with strong proposal-writing capacity over those with more productive use cases – a structural bias toward bureaucratic capability rather than technical ambition. 

For startups that do clear eligibility, the operational constraints are significant. Fine-tuning workloads are capped at a seven-day lease, which means teams building and iterating on models face renewal uncertainty mid-run. Companies selling AI services typically commit to fixed customer pricing over months – sudden compute cost changes or access gaps in the middle of a workload aren’t something a product roadmap can absorb easily. 

The voucher system adds a third constraint. Subsidized access typically covers 40 to 60 percent of compute costs, with users paying the remainder. For a company already managing cash flow, the partial subsidy still requires significant capital allocation. And without multi-year pricing certainty, that commitment is hard to plan around. 

Underneath all of this is a reliability question. AIRAWAT, India’s national AI supercomputing infrastructure, has had documented downtime incidents. As one expert told Business Standard: “Building compute is the easy part; running it reliably with SLAs researchers can actually plan around is what the Mission has not yet solved completely.” For enterprises with customers and revenue tied to their AI systems, uptime isn’t a nice-to-have, it’s the baseline. 

What Enterprises Actually Need 

Here’s the usage reality: more than 60% of startups requesting access through the IndiaAI Compute Portal want GPUs for inference, not training. Another quarter wants fine-tuning capacity. The segment running full foundation model training from scratch is small and getting smaller as open-source models mature. 

Inference is not a grant-cycle workload. It doesn’t arrive on a schedule. It scales with user traffic, spikes unexpectedly, and needs to support real-time inference in India. The compute underneath it has to be on-demand, reliably available, and governed by SLAs that mean something. A seven-day lease cap and uncertain renewal cycles are architectural mismatches for that use case – not edge cases to work around, but fundamental incompatibilities.

The broader ecosystem gap compounds this. Global cloud providers combine compute with developer tooling, managed networking, deployment infrastructure, and technical support. Access to a GPU at ₹65 per hour doesn’t include any of that. For an enterprise moving from model to production, the time spent stitching together surrounding infrastructure is often the actual bottleneck – not GPU availability itself. 

The IndiaAI scheme was designed to democratize compute access for research and model development. That’s a legitimate goal, and for the users it was designed for, it works. But the majority of enterprise AI workloads in 2026 aren’t about building models – they’re about running them, reliably, at scale, with predictable costs and no operational surprises. 

Neysa: Built for Enterprise Workload

Neysa Velocis is built for exactly this gap. On-demand GPU infrastructure, enterprise SLAs, no eligibility gates, no lease caps. Compute that’s available when a workload needs it, not when a review cycle completes. For enterprises running inference at scale, iterating on fine-tuned models, or building AI products with customers depending on uptime, that’s the difference between infrastructure you can build on and infrastructure you have to work around. 

Curious about what enterprise GPU infrastructure looks like without the scheme friction?

Reach out to the Neysa’s product solutions team.