The defining AI story of the year so far isn’t a flashy consumer chatbot or a viral image model. It’s plumbing. One of the largest Indian funding rounds in the first half of the year went to Neysa, an AI-cloud and GPU-infrastructure company that reached unicorn status in under three years. According to Business Standard, citing Tracxn data published in June 2026, Neysa raised roughly $600 million, placing it among H1 2026’s three largest Indian startup rounds.
That a company selling compute — not an app sitting on top of it — attracted that kind of capital tells you where investors think the durable value sits. As India races to add raw processing power, a homegrown ‘GPU-as-a-service’ layer is quietly forming between the imported silicon and the applications everyone else is busy building. This is a feature about that layer, why it is forming now, why it may prove unusually defensible, and where it could come unstuck.
The layer taking shape
Neysa’s pitch is deceptively simple: take expensive, scarce GPUs and rent them out as a managed cloud service, complete with the orchestration, networking, and software tooling that turn bare metal into something a developer can actually use. It is the same logical position that the big global cloud providers occupy — sitting between the chip vendors and the people writing models and products — but built for the Indian market and Indian customers.
The economics of GPU-as-a-service are what make this interesting. The capital required to buy, house, power, and cool large GPU clusters is enormous, and most startups and even mid-sized enterprises cannot justify owning that hardware outright. A specialised provider amortises that cost across many tenants, sells capacity by the GPU-hour, and lives or dies on how fully those expensive machines stay booked. Get utilisation right and the margins are real; get it wrong and the depreciation eats you.
Reaching unicorn status in under three years on the strength of infrastructure — rather than a hit product — is a signal in itself. It suggests investors believe demand for compute will outpace supply for long enough that whoever controls the rentable middle layer captures durable economics. In a market awash with model wrappers and thin AI features, the company selling the shovels looks, to a certain kind of backer, like the safer bet.
Why now
Three forces are converging to make this the moment for an Indian compute layer.
The first is demand. Indian enterprises, startups, and government bodies are all moving from AI experimentation to deployment, and deployment is hungry for GPUs — for fine-tuning, for inference at scale, and increasingly for training domestic models. That demand is rising faster than the country’s installed compute base can comfortably absorb, which is precisely the condition under which a service layer thrives.
The second is public policy. The IndiaAI Mission has been onboarding subsidised public GPUs at a deliberately low price point. According to figures published by ExplainX referencing the IndiaAI Mission in June 2026, the programme had onboarded somewhere in the region of 34,000 to 38,000 GPUs at a subsidised rate of around ₹65 per GPU-hour, with a stated target of 100,000 public GPUs by December 2026. (These figures should be read against official PIB and MeitY releases.) Cheap, state-backed compute lowers the barrier to entry for the whole ecosystem and, crucially, normalises the idea of buying compute as a metered utility.
The third is private build-out. Large Indian conglomerates including Reliance and Tata, alongside the global hyperscalers expanding their local regions, are pouring capital into data-centre and GPU capacity. The same ExplainX summary suggests private deployments could push national GPU capacity past 200,000 by year-end. Whatever the precise number, the direction is unambiguous: India is adding compute at pace, and a denser supply base creates more room for specialised providers who can package, broker, and optimise that capacity for specific workloads.
Why infra is defensible
The most interesting argument for the compute layer is not that it is exciting — it isn’t — but that it is hard to dislodge once established.
Start with capital intensity. Building a credible GPU cloud requires real estate, power agreements, networking, cooling, and tens of millions of dollars of hardware that depreciates whether or not it is used. That cost is a moat in both directions: it keeps casual competitors out, and it locks customers in, because migrating live training and inference workloads between providers is genuinely painful. Switching costs in infrastructure are some of the stickiest in technology.
Then there is data residency. A growing number of Indian organisations — in financial services, healthcare, and government especially — have strong reasons, regulatory or political, to keep their data and their compute on Indian soil. A domestic AI-cloud provider can offer data-residency-friendly compute as a first-class feature, something foreign hyperscalers can match only by building out local regions, and something a domestic player can credibly own as part of its identity.
Finally, the competitive geometry is favourable. An infrastructure company does not have to win the brutal, capital-incinerating race to build the best foundation model. It can stay deliberately neutral, serving whoever is training and serving models — including the foundation labs themselves. In a landscape where it is far from clear which model will dominate, selling the underlying compute to all of them is a structurally calmer place to stand. You profit from the gold rush without having to find the gold.
The hard part
None of this is easy money, and it would be irresponsible to pretend otherwise.
The most obvious vulnerability is the chip supply chain. India’s GPU layer runs overwhelmingly on imported silicon, and in practice that means Nvidia. A domestic ‘compute sovereignty’ story built on hardware designed and manufactured abroad is sovereign only up to a point. Allocation, pricing, and export-control dynamics all sit outside Indian control, and any provider’s growth is ultimately gated by its ability to actually acquire chips. Building a business on top of someone else’s scarce product is a real dependency, not a footnote.
Second is the utilisation-and-pricing trap. The whole model rests on keeping expensive hardware busy. Idle GPUs are pure loss, and depreciation does not pause for slow quarters. Worse, the existence of heavily subsidised public GPUs at around ₹65 per GPU-hour sets a low anchor price for the entire market. A private provider has to justify a premium through performance, reliability, support, and software — or get squeezed between subsidised public capacity below and the hyperscalers’ scale above. Pricing power, in other words, is not guaranteed by demand alone.
Third is energy and the cost curve. GPU clusters are voracious consumers of power, and power in India is neither uniformly cheap nor uniformly reliable. Sustaining margins as the hardware ages and newer, more efficient chips arrive means relentless attention to power purchase agreements, cooling efficiency, and the timing of hardware refresh cycles. The cost curve only bends in your favour if you manage it aggressively; left alone, it bends the other way.
Put together, these are the reasons infrastructure is a grind rather than a gold rush. But that is also, paradoxically, the case for it. The work is hard, capital-intensive, and unglamorous — which is exactly why fewer players can do it, and why those who do it well may still be standing when this year’s crop of chatbots has been forgotten. If India is going to build a lasting position in AI, the most durable bet may not be the smartest model. It may be the layer of compute humming quietly underneath all of them.
