In 1987 India asked to buy a Cray X-MP. It wanted the machine for monsoon forecasting. The United States refused the sale, under the export-control regime then known as COCOM, on the ground that a computer capable of modelling the atmosphere is also capable of modelling a warhead.
A PIB backgrounder of 4 October 2026 sets out where India's supercomputing has reached — 40 machines, 68 petaflops, 13 of them above a petaflop each. It traces the lineage back to PARAM 8000 in 1991. It does not mention the refusal, and the refusal is the reason the lineage exists.
What a denial produced
The Centre for Development of Advanced Computing (C-DAC) was set up at Pune in 1988 under Vijay Bhatkar, with a mandate to build what India had been prevented from buying. PARAM 8000 was unveiled in 1991 at a speed of 1 gigaflop — one billion floating-point operations per second — demonstrated at the Zurich supercomputing exhibition that year, and subsequently exported to buyers including Germany and Russia.
The interesting part is not the speed. It is the architecture, and the fact that the embargo chose it.
A Cray of that era was built around a small number of extremely fast custom vector processors — expensive, exotic silicon that India could neither buy nor make. C-DAC took the only route available: wire together a large number of commodity microprocessors and make them cooperate. PARAM is short for PARAllel Machine, and the whole engineering problem moved from the processor to the interconnect and the software that splits a calculation across many chips.
That is massively parallel processing, and within a decade it was how the entire industry built supercomputers. The vector-processor approach died; the many-cheap-chips approach won. An export control intended to keep India behind pushed it onto the side of an architectural argument that was about to be settled in its favour. This is worth knowing precisely because it is the opposite of the usual story about technology denial.
The denial also explains why high-performance computing remains an export-controlled category today. HPC is dual-use by nature: the machine that integrates the primitive equations of the atmosphere also runs computational fluid dynamics over an airframe, or hydrodynamics relevant to weapon physics. COCOM's successor regime, the Wassenaar Arrangement, which India joined in December 2017, still lists computing capability. A country that cannot buy the machine has to build it, and then — having built it — is admitted to the club that controls it.
FLOPS, and the scoreboard problem
FLOPS is Floating Point Operations Per Second. A teraflop (TF) is a trillion, a petaflop (PF) a quadrillion, an exaflop (EF) a quintillion operations per second.
Here is where the backgrounder requires arithmetic it does not perform. It states that the world's fastest supercomputer achieves around 2.19 exaflops in 2026. It also states that India's National Supercomputing Mission has deployed 40 machines totalling 68 PF.
2.19 exaflops is 2,190 petaflops. The single fastest computer in the world is therefore roughly 32 times the combined capacity of everything NSM has built — and India's national total is about 3% of one machine.
That comparison is not a criticism of NSM, but it does establish what NSM is and is not. It is not a programme to build the world's fastest computer. Its distribution tells you its actual purpose:
| Capacity band | Number of systems |
|---|---|
| Above 1 PF | 13 |
| 500 TF to 1 PF | 12 |
| Below 500 TF | 15 |
| Total | 40, combining to 68 PF |
Fifteen machines under half a petaflop are not there to win benchmarks. They are there so that a university in a state capital has a cluster its researchers can actually get time on. NSM is a distribution programme — access spread across institutions — and judging it by peak FLOPS is using the wrong scoreboard. The target is 50 machines exceeding 123 PF in aggregate, which is a statement about coverage, not about rank.
The mission's architecture
NSM was launched in April 2015 with an outlay of about ₹4,500 crore. It is jointly steered by the Department of Science and Technology and the Ministry of Electronics and Information Technology — an unusual two-ministry arrangement that reflects the mission's two halves, scientific application and electronics manufacturing. Implementation rests with C-DAC, Pune and the Indian Institute of Science, Bengaluru.
Its stated method is a "Build approach" in three concurrent phases: assembly, then manufacturing, then design and manufacturing. Read that as a deliberate climb up the value chain — first screwing together imported subsystems, then making the subsystems, then designing what you make. Running the phases concurrently rather than sequentially is how capacity kept growing while capability was still being acquired.
What is actually indigenous — and what is not
This is the part most often got wrong, and it is the part most worth getting right.
The real achievement of NSM is the Rudra server: a server board and system designed and developed by C-DAC, with the technology transferred to Indian Electronics Manufacturing Services (EMS) partners for production. As of September 2026, 6,000 Rudra servers had been deployed in PARAM Rudra supercomputers, with a further 1,500 under manufacture. Around it C-DAC has built an indigenous high-speed interconnect, tested at 100 Gbps and 200 Gbps, indigenous cooling technology now at deployment stage, and a complete HPC system software stack.
Three PARAM Rudra systems were dedicated in September 2024, at the Inter-University Accelerator Centre in New Delhi, the Giant Metrewave Radio Telescope facility near Pune operated by NCRA, and the S.N. Bose National Centre for Basic Sciences in Kolkata, at a cost of roughly ₹130 crore each, alongside a separate HPC system dedicated to weather and climate research.
Now the caveat. An indigenously designed server is not an indigenous processor. The Rudra server is an Indian-designed board, chassis, interconnect and software stack populated with commercially available central processing units. The CPU die itself is not Indian. India's own processor effort is a separate and much earlier-stage programme — C-DAC's VEGA family, which is where the work behind VIHAAN sits, and the fabrication and packaging capacity being built under the India Semiconductor Mission.
Conflating the two — "indigenous supercomputer" with "indigenous chip" — is the standard error on this topic, and it is exactly the distinction a well-set question will test. What India has achieved is system sovereignty: the ability to design, assemble, cool, network and operate a large machine without foreign integrators, and to repeat it 6,000 servers at a time. That is a supply-chain capability, and it is more durable than any FLOPS figure, because a FLOPS number is obsolete in three years while a manufacturing base is not.
At the other end of the range sits PARAM Shavak, a "supercomputing-in-a-box" system designed and made in India for engineering colleges and universities — the smallest unit of the same distribution logic.
What the machines are used for
The applications are specific enough to be examinable:
- Genomics and Drug Discovery Platform — screening large molecule sets; used during COVID-19 to screen existing drugs for repurposing and to predict side effects including cardiac risk
- Urban Environment Decision Support System — coupled weather and air-pollution modelling, forecasting heavy rain and pollution episodes for city authorities
- Seismic Imaging Suite — subsurface mapping for oil and gas exploration, with performance claimed to match commercial tools
- Flood early warning for river basins — forecasts up to 2 days ahead, in use for the Mahanadi basin
- Forest Fire Spread Model — satellite remote sensing coupled to a spread model, tested in the Sikkim Himalayas
- Materials science and computational chemistry suites for simulating atoms, molecules and alloys
The institutional users include the India Meteorological Department, the Central Water Commission, the Central Pollution Control Board and the Ministry of AYUSH. PIB records the mission as contributing to 11 UN Sustainable Development Goals.
On people, the figures are the ones that best describe a distribution programme: over 16,000 researchers, including more than 2,900 PhD scholars, across over 400 institutions; more than 1.5 crore compute jobs executed; over 1,990 research publications, as of September 2026. Training runs through AICTE-linked faculty development, an annual EduHPC workshop, a structured course on SWAYAM under NPTEL, the HPC Shiksha portal, and hackathons and bootcamps. The National Knowledge Network (NKN) connects the sites, which is what allows a researcher at one institution to use a machine at another.
One claim to hold at arm's length
The backgrounder states that India generates nearly 20% of the world's data. This figure circulates widely in Indian policy documents, and there is no agreed method for attributing global data generation to countries — no standard denominator, and no consistent definition of what counts as data generated. Treat it as an official assertion about scale, cite it as such if asked, and do not treat it as a measured statistic. The substantive argument for expanding computing capacity does not depend on it.
Separately, ESTIC-2026 — the second Emerging Science, Technology and Innovation Conclave, organised by CSIR — is scheduled for 27 to 29 October 2026 at Bharat Mandapam, New Delhi, across 12 thematic areas, with supercomputing among them.
🔑 Revision block
- The origin: the United States blocked the sale of a Cray X-MP to India in 1987-88 under COCOM, citing possible diversion to missile and nuclear design. India had sought it for weather forecasting.
- C-DAC: Centre for Development of Advanced Computing, set up at Pune, 1988, under Vijay Bhatkar.
- PARAM 8000: unveiled 1991 at 1 gigaflop; PARAM = PARAllel Machine; demonstrated at Zurich in 1991 and later exported, including to Germany and Russia. PARAM Yuva reached 54 teraflops.
- Architecture: denial forced massively parallel processing — many commodity microprocessors rather than a few custom vector processors. The industry converged on the same choice a decade later.
- Dual-use: HPC remains export-controlled; COCOM's successor, the Wassenaar Arrangement, was joined by India in December 2017.
- FLOPS: Floating Point Operations Per Second. TF = trillion, PF = quadrillion, EF = quintillion.
- NSM: launched April 2015, outlay about ₹4,500 crore. Jointly steered by DST and MeitY; implemented by C-DAC Pune and IISc Bengaluru. "Build approach" — assembly, manufacturing, design-and-manufacturing, run concurrently.
- Status, September 2026: 40 supercomputers, 68 PF combined — 13 above 1 PF, 12 between 500 TF and 1 PF, 15 below 500 TF. Target 50 machines exceeding 123 PF.
- Scale check: the world's fastest machine is about 2.19 EF = 2,190 PF, roughly 32 times all of NSM's capacity. NSM is a distribution programme, not a peak-performance one.
- Rudra server: designed and developed by C-DAC, technology transferred to Indian EMS partners. 6,000 deployed in PARAM Rudra systems, 1,500 under manufacture. Indigenous interconnect at 100 and 200 Gbps, indigenous cooling, full HPC software stack.
- PARAM Rudra deployments, September 2024: IUAC New Delhi, GMRT/NCRA Pune, S.N. Bose Centre Kolkata — about ₹130 crore each.
- The distinction to hold: indigenous server design is not an indigenous processor. The CPU die is not Indian; India's processor work is the VEGA family and the semiconductor mission.
- PARAM Shavak: "supercomputing-in-a-box" for colleges.
- Applications: genomics and drug discovery (COVID-19 repurposing and cardiac-risk prediction) · urban weather and pollution · seismic imaging · basin flood warning at 2 days, used on the Mahanadi · forest-fire spread, tested in the Sikkim Himalayas · materials and computational chemistry. Users include IMD, CWC, CPCB, AYUSH. Contributes to 11 SDGs.
- People: over 16,000 researchers, 2,900+ PhD scholars, 400+ institutions, 1.5 crore+ compute jobs, 1,990+ publications. Backbone network — NKN. Skilling via EduHPC, HPC Shiksha, NPTEL on SWAYAM, AICTE faculty programmes.
- Claim to flag: "India generates nearly 20% of the world's data" is an official assertion without an agreed measurement basis.
- ESTIC-2026: second edition, 27-29 October 2026, Bharat Mandapam, organised by CSIR, 12 thematic areas.
🎯 Practice MCQs
Q1. The immediate trigger for India's indigenous supercomputing programme in the late 1980s was: (a) The denial of a Cray X-MP supercomputer to India under the COCOM export-control regime (b) The failure of the Indian Meteorological Department's monsoon forecast of 1987 (c) A recommendation of the Scientific Advisory Committee to the Cabinet on parallel computing (d) The collapse of a joint venture with a Japanese computer manufacturer
→ (a) The United States blocked the sale on dual-use grounds, and C-DAC was created at Pune in 1988 to build what India had been refused.
Q2. PARAM 8000, unveiled in 1991, had a rated speed of about: (a) 1 megaflop (b) 1 teraflop (c) 1 gigaflop (d) 1 petaflop
→ (c) One gigaflop — a billion floating-point operations per second. PARAM Yuva later reached 54 teraflops, and today's entry-level NSM systems are below 500 teraflops.
Q3. The architectural choice forced on C-DAC by the export embargo was to: (a) Use a small number of custom vector processors (b) Connect a large number of commodity microprocessors in parallel (c) Build an analogue computing array (d) Lease computing time from foreign facilities
→ (b) Massively parallel processing, with the engineering difficulty shifted to the interconnect and the software — the approach the global industry itself adopted within a decade.
Q4. The National Supercomputing Mission is jointly steered by: (a) DRDO and the Department of Space (b) CSIR and the Department of Atomic Energy (c) MeitY and the Ministry of Education (d) The Department of Science and Technology and MeitY
→ (d) DST and MeitY steer it jointly; C-DAC Pune and IISc Bengaluru implement it. The two-ministry arrangement mirrors the mission's scientific and electronics-manufacturing halves.
Q5. As of September 2026, NSM had deployed 40 supercomputers with a combined capacity of 68 PF. Of these, the number exceeding 1 PF individually was: (a) 13 (b) 15 (c) 12 (d) 40
→ (a) Thirteen above 1 PF, twelve between 500 TF and 1 PF, and fifteen below 500 TF — a spread that identifies NSM as an access programme rather than a peak-performance one.
Q6. If the world's fastest supercomputer in 2026 achieves about 2.19 exaflops, India's entire NSM capacity of 68 petaflops is approximately: (a) Half of that single machine (b) About 3% of that single machine (c) About 30% of that single machine (d) Twice that single machine
→ (b) 2.19 EF equals 2,190 PF, so 68 PF is roughly 3% — the single machine is about 32 times India's national NSM total.
Q7. The Rudra server developed under NSM is best described as: (a) An indigenously fabricated central processing unit (b) An imported server rebranded for Indian deployment (c) A quantum co-processor for hybrid computing (d) An indigenously designed server whose technology has been transferred to Indian EMS partners for manufacture
→ (d) Indian design, Indian manufacture, commercially available CPUs. The processor die is not Indian — that is the separate VEGA and semiconductor-mission effort.
Q8. Three PARAM Rudra supercomputers were dedicated in September 2024 at: (a) IIT Madras, IIT Bombay and IISc Bengaluru (b) BARC Mumbai, IGCAR Kalpakkam and VECC Kolkata (c) IUAC New Delhi, the GMRT facility near Pune and the S.N. Bose Centre, Kolkata (d) ISRO Bengaluru, NRSC Hyderabad and SAC Ahmedabad
→ (c) Each cost about ₹130 crore, and the choice of hosts — accelerator physics, radio astronomy and basic sciences — reflects the research workloads the systems were sized for.
Q9. The NSM flood early-warning application provides forecasts up to two days ahead and has been applied to the: (a) Mahanadi basin (b) Brahmaputra basin (c) Godavari basin (d) Krishna basin
→ (a) The Mahanadi basin. The forest-fire spread model, by contrast, has been tested in the Sikkim Himalayas.
Q10. The claim that "India generates nearly 20% of the world's data" should be treated as: (a) A measurement published annually by the International Telecommunication Union (b) An official assertion for which no agreed measurement basis exists (c) An estimate derived from the Census of India (d) A figure certified by the Top500 project
→ (b) There is no standard method for attributing global data generation to countries. Cite it as an official claim about scale, not as a statistic — and note that the case for more computing capacity does not rest on it.
📋 How this gets asked (PYQ pattern)
Supercomputing is examined in four recognisable ways, and the hardest of them is a distinction rather than a fact.
The first is the indigenisation distinction. Questions frequently offer "indigenously developed processor" as a plausible option about PARAM or Rudra. The correct position is indigenous design, system software, interconnect, cooling and manufacture with commercially sourced CPUs. Candidates who have learnt the word "indigenous" without the noun it attaches to lose this mark.
The second is institutional mapping. C-DAC builds; DST and MeitY steer; IISc co-implements; CSIR runs ESTIC; NKN connects. A single question pairing one body with the wrong role is standard, and C-DAC against CSIR is the most common swap.
The third is the numbers series. 1991 and 1 gigaflop for PARAM 8000; 2015 and ₹4,500 crore for NSM; 40 machines and 68 PF for the current position; 50 machines and 123 PF for the target. Four pairs, each of which has appeared in one form or another.
The fourth, and the one worth most in a written paper, is the dual-use argument. Why was the Cray refused, why is HPC still export-controlled, and what does that imply for self-reliance in strategic technologies? The answer runs from COCOM through the Wassenaar Arrangement to the semiconductor and quantum missions, and it is the same argument every time: a capability you can only buy is a capability someone else can switch off.
Preparing for NDA? With any indigenisation claim, identify the exact layer being claimed — design, manufacture, materials, or the chip itself. Scheme questions are won on that precision far more often than on remembering a speed. Build the base with our NDA study material, follow the daily NDA current affairs, and prepare with our faculty in the upcoming Cavalier courses in Delhi.
✍️ Written by Col Vijyanat Thakur — Faculty, Science & Defence Studies, at The Cavalier. Reviewed by the Cavalier Faculty Desk.