By Vishal Sirohi, Co-Founder and CEO, Island Computing
Indian engineers have built and operated the world’s largest distributed systems for two decades. Most of that work happened inside companies headquartered somewhere else. What is changing now is where that skill is exercised and who owns what it produces.
India built consumer-scale SaaS, payment platforms on top of the India Stack, logistics systems that clear millions of transactions a day, and product engineering organisations serving global customers. The global capability centres employ over 1.4 million people, roughly a quarter of India’s technology workforce, and engineering and research and development account for 56% of GCC revenue.
The layer beneath the application covers hypervisors and the isolation boundaries they enforce, distributed control and data planes that stay correct across data centres, identity and authorisation substrates that answer in under a millisecond, storage engines, scheduling, and the orchestration that turns racks and GPUs into something a builder consumes without assembling it first. It covers operations at production reliability, a harder problem than the build.
That skill set concentrated inside a small number of foreign cloud platforms over twenty years. Indian engineers hold senior positions across all of them. The platforms were built elsewhere.
Three changes have landed at once to move that work onshore. Tightening US immigration policy and a less predictable geopolitical environment are bringing back engineers who spent a decade operating infrastructure at the scale of the largest platforms in the world. AI-augmented engineering has compressed the cost and cycle time of building infrastructure by an order of magnitude, and a team of eight to ten engineers with the right depth now delivers what required fifty a decade ago. Infrastructure used to demand headcount and capital that only incumbents could raise.
On the demand side, the DPDP Act, RBI localisation circulars, SEBI’s cloud advisories, CERT-In’s six-hour incident-reporting directive, and the MeitY addendum of 20 March 2026 have moved infrastructure decisions from procurement to design. A compliance requirement that has to be evidenced cannot be satisfied by a contract clause. It has to sit inside the platform.
MeitY’s five-layer framing of the sovereign AI stack, articulated by Minister Ashwini Vaishnaw at the India AI Impact Summit 2026, covers compute, data, models, applications, and skilling as horizontal layers. Inside the compute and infrastructure layer sits a four-part sub-stack: semiconductor, operating system and driver, networking, and distributed systems and orchestration.
The India Semiconductor Mission and Semicon 2.0 are working the lower components, with a stated expectation of at least fifty deep-tech companies emerging from the programme. The IndiaAI Mission had 38,000 GPUs deployed as of February 2026, available to startups and researchers at ₹65 an hour, with 20,000 more coming under AI Mission 2.0.
Vaishnaw has put US$90 billion as committed against a possible US$400 billion across the five layers within two years. With the capital and the hardware arriving, the distributed systems and orchestration component is what turns both into a product, and it has the fewest Indian teams working on it today.
The engineering problems in that component are open. Cloud platforms designed in 2006 enforce isolation at the account boundary, which protects one cloud customer from another. It does not protect one tenant from another inside a single customer’s account, and every meaningful business on the internet is now multi-tenant by construction.
The industry response has been for each customer to rebuild that isolation independently through multi-account architectures, cross-account permissions, and tag-based partitioning. That is the same engineering problem being solved thousands of times over because the substrate never treated the tenant as a first-class primitive.
Agentic workloads add a second set. An autonomous agent needs scoped, time-bound identity rather than a long-lived credential. Observability has to describe execution trees rather than request logs. Looping agents need hard budgets and circuit breakers enforced at admission time, before a runaway workload reaches shared hardware. Schedulers have to place work across heterogeneous compute with budget and service-level awareness built in. Retrofitting these into a stack that did not assume autonomous initiators does not work.
A GPU is one component of a computing system, alongside CPU, memory, and storage, currently carrying the role the CPU carried in the 1970s, and it will commoditise on the same path. A single GPU serves a single node, and enterprise workloads run across nodes, racks, and data centres on fault-tolerant distributed systems.
The platform above the GPU decides whether the capacity is usable in production. Models do not store data. GPUs do not store data. The master of the data is the database, and large-scale database engines remain overwhelmingly foreign-operated. Sovereign compute without sovereign data is not sovereignty.
For engineers weighing the move, the shift is a change of optimisation target. Application engineering optimises for feature velocity, and the feedback loop is a user. Infrastructure engineering optimises for failure behaviour, and the feedback loop is a page at two in the morning.
The craft is static stability, blast radius containment, observable failure modes, and a recovery path for every failure that does not depend on heroics. Operational excellence is a cultural discipline built into how a team works, built from the first hire rather than added when the first outage arrives.
UPI, Aadhaar, and DigiLocker compounded because the substrate underneath them was designed and operated in India. The next decade will test whether the same holds for compute.
The generation that wrote India’s applications is qualified to build what runs beneath them, and whether it does decides if India in 2035 is a technology producer or a technology consumer.