For most of the last decade, India’s role in the global computing economy was that of a tenant. We wrote the software, staffed the support desks, and ran workloads on someone else’s silicon in someone else’s data centre. The hardware that actually mattered — the chips, the packaging, the GPUs — was designed in California, fabricated in Taiwan, and assembled in China. India consumed compute; it did not own it.
That posture is quietly changing. Over the past eighteen months, a cluster of announcements — a multi-billion-dollar packaging project in Odisha, fabs breaking ground in Gujarat, and tens of thousands of subsidised GPUs coming online under the IndiaAI Mission — has begun to outline a different ambition. Beyond chasing hyperscaler data centres, India is attempting to own the chip-and-compute layer itself. For founders building in AI and deep tech, this is not abstract industrial policy. It changes the cost of training a model, the legality of where your data lives, and the kinds of companies that are now fundable in India.
The chip layer gets real
The headline event is in Odisha, where the state signed a memorandum of understanding with Intel and 3DGS for a semiconductor packaging project valued at roughly $3.3 billion, according to reporting compiled by Angel One and EE Times. Packaging — the step where bare silicon dies are encased, connected, and made usable — has historically been treated as the unglamorous back end of chipmaking. In the age of chiplets and advanced 3D stacking, it has become one of the most strategically contested parts of the value chain. India staking a claim here, rather than only in design, signals genuine intent.
Crucially, this is no longer a slide-deck economy. The manufacturing layer is moving from plan to production. Tata’s fab at Dholera, Gujarat, carries an investment of around Rs 91,526 crore and is targeting mature nodes in the 110nm-to-28nm range — the workhorse chips that go into cars, appliances, and industrial systems rather than bleeding-edge AI accelerators, per DQ India. Meanwhile, Micron’s assembly and test plant at Sanand is already shipping DRAM and NAND memory to customers reportedly including Dell, Asus, and Qualcomm — from Indian soil. A chip that is packaged in Gujarat and sold to a global PC maker is a different proposition from a press release about future capacity.
What ties these threads together is the framing the government now calls Semicon 2.0, where design is treated as the priority rather than an afterthought. The logic is sound: India already has a deep bench of chip-design engineers working inside the local R&D centres of global semiconductor firms. Owning the intellectual property — the architectures, not just the assembly — is where margins and leverage live. Whether policy can convert captive talent into homegrown design houses is the open question, but the direction of travel is unambiguous.
Subsidised compute as industrial policy
If fabs are the long game, the IndiaAI Mission is the near-term lever. The Mission has built a shared GPU pool and is renting access to it at deliberately suppressed rates — onboarding roughly 38,000 GPUs at subsidised prices reported at around Rs 65 per GPU-hour, according to IndiaAI and EE Times. For context, comparable high-end GPU capacity on commercial clouds can cost several multiples of that. The stated target is 100,000 GPUs by the end of 2026.
Treating compute as subsidised industrial infrastructure — the way an earlier era treated electricity or ports — is a meaningful policy choice. It tells startups, researchers, and academics that the cost of experimentation will be socialised rather than left entirely to venture funding and hyperscaler list prices. The capacity is being delivered through empanelled private providers building what amount to sovereign clouds, with Yotta’s Shakti Cloud among the most visible. The intent is that AI workloads for Indian users can run on infrastructure that is physically and legally located in India.
There is a healthy debate to be had about whether subsidised GPU-hours are the best use of public money, and whether the pricing is sustainable once subsidies taper. But as a signal — that the state sees compute as a strategic input worth underwriting — it lands. The Mission has effectively created a floor under which the cost of training in India is unlikely to rise, at least for the subsidy window.
What it unlocks for startups
For founders, the most immediate consequence is straightforward: cheaper training and inference. A two-person team fine-tuning an open-weight model, or a vertical SaaS company building retrieval-augmented features, faces a materially lower compute bill when subsidised GPU-hours are on the table. That compresses the capital required to reach a working prototype, which in turn changes what early-stage investors are willing to back.
The second unlock is data-residency-friendly AI. A growing list of Indian customers — banks, insurers, government departments, healthcare providers — cannot or will not send sensitive data to compute located offshore. Sovereign cloud capacity inside India turns a compliance headache into a feature. Startups that can credibly say their entire stack, training included, runs within Indian jurisdiction have a commercial edge in exactly the regulated sectors where AI budgets are largest.
The third, and most durable, opportunity is in deep tech and chip design itself. A maturing domestic semiconductor ecosystem creates demand for an entire supporting cast: electronic design automation tooling, materials and chemicals, test and inspection, packaging IP, and the kind of fabless design startups that have been rare in India precisely because the surrounding infrastructure did not exist. As fabs and packaging lines come online, the addressable market for these companies stops being purely hypothetical.
- Cheaper experimentation: subsidised GPU-hours lower the cost of training and inference for early-stage AI teams.
- Compliance as a moat: in-country compute makes data-residency a selling point in BFSI, healthcare, and public sector deals.
- New deep-tech surface area: fabs and packaging create pull-through demand for EDA, materials, test, and fabless design startups.
The hard part
None of this is guaranteed, and the gap between announcement and outcome in Indian industrial policy is wide and well-documented. Three constraints deserve sober attention.
The first is energy and talent — and time. Fabs are among the most power-hungry and water-intensive industrial facilities on earth, and they demand uninterrupted, ultra-clean supply that India’s grid does not uniformly provide. A 28nm fab does not produce competitive chips on day one; yields take years to mature, and the engineers who know how to coax those yields are scarce and globally mobile. This is a multi-year payoff measured in process generations, not funding rounds. Founders banking on a fully self-sufficient domestic supply chain by next year will be disappointed.
The second is chip access under global export controls. The most advanced AI accelerators are subject to tightening US export restrictions, and India’s ability to acquire frontier silicon — whether to stock the IndiaAI GPU pool or to feed advanced design ambitions — is partly hostage to geopolitics it does not control. Building mature-node fabs domestically does little to solve frontier-compute access in the short term. The 100,000-GPU target assumes a procurement environment that could shift with a change in Washington’s posture.
The third is the oldest problem of all: policy follow-through versus announcement. India has a long history of signing MoUs that quietly lapse, of incentive schemes whose disbursement crawls, and of headline numbers that outrun reality. An MoU is an intention, not a fab. The Micron shipments and the Dholera construction are encouraging precisely because they are evidence of execution — but one or two proof points do not make an ecosystem. The honest assessment is that India has finally assembled the right ingredients and the right urgency; whether it has the institutional stamina to see a decade-long build-out through is the question that will actually decide chip sovereignty.
For startups, the pragmatic read is this: take the subsidised compute, build for data residency, and watch the deep-tech surface area expand — but do not bet the company on capacity or chip access that exists only on paper. The chip-and-compute layer is becoming real in India. It is just not finished, and it will not be for some time.
