The Three Letters That Decide India's AI Future. G P U

Almost every AI system in the world today runs on a special kind of chip called a GPU. ChatGPT runs on GPUs. Gemini runs on them. The Indian Akashteer air defence system uses them. The Sarvam and BharatGen sovereign AI models India launched in February 2026 were trained on them.

The world's most strategic technology is now more concentrated than oil ever was. One American company designs almost all the advanced ones. One Taiwanese factory manufactures over ninety per cent of the cutting-edge ones. India is currently allowed to import roughly 50,000 of them through the year 2027.

This article explains, in plain English, what a GPU is, why so few sources supply them, and where India stands.

What a GPU Actually Is

A normal computer chip, the kind in any laptop, is called a CPU. A CPU is built to do one complex task at a time, very fast. Think of it like a fighter pilot. One person, highly trained, making rapid decisions in sequence.

A GPU, or Graphics Processing Unit, was originally invented to draw graphics on a screen. Drawing a screen means changing the colour of millions of small dots at the same time. A GPU is built to do thousands of simple tasks all at once, in parallel. Think of it as a thousand technicians on a workshop floor, each handling one small calculation, all working at the same time.

For decades, GPUs were used mostly for video games. Around 2012, researchers discovered something that changed everything. Training an AI model also requires doing trillions of small calculations in parallel. The same chips that drew video game frames could train a language model. The same workshop, with the same thousand technicians, just doing AI maths instead of pixel maths.

This is why every AI breakthrough since 2012 has been built on GPUs. There is no other class of chip in the world that can train modern AI models at any reasonable speed or cost.

The Nvidia Chokepoint

Almost all of the advanced GPUs that train the world's AI are designed by a single American company called Nvidia.

Nvidia's most important product right now is a chip called the H100, and its newer successor the Blackwell B200. According to industry analyst reporting on Nvidia's 2025 sales, roughly 80 per cent of the company's chips sold that year were Blackwell models, and most of the rest were H100s.A single Blackwell GPU costs in the range of forty to forty-five thousand US dollars.

No other company competes meaningfully at this level. AMD makes GPUs but holds a small share of the AI market. Chinese companies like Huawei have been building alternatives but cannot yet match Nvidia's performance or its software ecosystem.

Source: Institute for Progress, B30A chip analysis

Because Nvidia is American, the United States government decides which countries are allowed to buy how many of these chips. In 2025 the US introduced a tiered export control regime. India was placed in a middle tier with a cumulative cap equivalent to roughly 50,000 H100-class GPUs through 2027, with possible expansion through bilateral agreements.

For comparison, a single American hyperscale data centre can use more than 100,000 H100s. India's entire two-year allocation is less than what one American technology company deploys for one project.

Source: Swarajya on US GPU export tiers

The TSMC Chokepoint

The second concentration is more uncomfortable.

Nvidia designs its chips, but does not physically manufacture them. The factories required to make a modern chip cost twenty to thirty billion dollars each, take five to seven years to build, and use equipment that only a handful of companies on earth can supply.

The world's leading chip manufacturer is a single Taiwanese company called TSMC, or Taiwan Semiconductor Manufacturing Company. As of early 2026, TSMC controls about 72 per cent of the global foundry market and more than 90 per cent of leading-edge chip production at the most advanced nodes. Every Nvidia GPU, every Apple processor, most AMD and Broadcom chips, and most of the world's other advanced semiconductors are made in TSMC's factories on the island of Taiwan.

Source: TSMC market data, Wikipedia

This is the part that should give any defence reader pause. The most strategic technology in the world is manufactured almost entirely on one island, ninety miles off the coast of mainland China, in a region with active military tension. A disruption of TSMC's production, for any reason, would cripple global AI development overnight. TSMC is building additional fabs in Arizona, Japan, and Germany, but these will not match the Taiwan facilities in capability for many years.

India's Position

India has, until recently, had no advanced chip manufacturing of its own. That is changing, but slowly and at older technology nodes.

The most significant project is the Tata Electronics partnership with Taiwan's PSMC at Dholera in Gujarat. Construction began in 2024. As of April 2026, the fab is roughly fifty per cent complete. First commercial chip production is targeted for late 2026. Total investment will reach approximately ninety-one thousand crore rupees. At full capacity it will produce 50,000 wafers per month.

The honest part of this picture is that Dholera will produce chips at 28 nanometre and older nodes. Useful for automotive, industrial, and many defence applications. But these are not the cutting-edge nodes used for high-end AI training. TSMC is currently shipping 3 nanometre and moving to 2 nanometre. India is starting where Taiwan was around 2014.

Source: Tata Electronics announcement

Meanwhile the IndiaAI Mission has assembled over 38,000 GPUs across 14 service providers and made them available at subsidised rates of around 65 rupees per hour, roughly one-third of the global market price.A sovereign GPU cluster of 3,000 next-generation chips is under construction for strategic government use. Every one of those imported GPUs is, today, an Nvidia chip manufactured by TSMC in Taiwan.

Five Things to Remember

One. Almost every modern AI system in the world runs on a class of chip called a GPU.

Two. Roughly eighty per cent of the world's advanced AI GPUs are designed by Nvidia, an American company.

Three. More than ninety per cent of leading-edge chips are manufactured by TSMC, a Taiwanese company.

Four. India is currently allowed to import approximately 50,000 H100-class GPUs through 2027 under US export controls.

Five. India's first semiconductor fab, the Tata-PSMC project at Dholera, is on track for first silicon by late 2026, at the 28 nanometre node. Cutting-edge AI chips are not yet within Indian manufacturing reach.

This is the foundation. Without GPUs, there is no AI. Without sovereign GPU access, there is no sovereign AI. Every conversation about India's AI ambitions, defence or civilian, runs through these three letters.

In the next piece, we will look at what India has actually built on top of this foundation, and where the defence ministry now plans to take it.