The AI Pipeline, Part 4: The Geography of Compute and the India Lens

The short answer: AI compute hubs exist in only about 30 countries, and three companies (Amazon, Microsoft, and Google) hold roughly 70 percent of global cloud infrastructure. India generates nearly 20 percent of the world's data but hosts about 3 percent of its data centers, and about 74 percent of Indian AI startups depend on proprietary Western closed models accessed through APIs.

This is part 4 of the AI Pipeline series. Parts 1 through 3 followed the stack: minerals, design software, the lithography machine, the foundry, the memory, the clouds, and then the governments weaponizing every layer of it. This part zooms out to a map, because the pipeline has a geography, and then zooms into the one country where the gap between using AI and owning AI is widest and most instructive. I am writing from India, so the India lens is not a foreign-country case study for me. It is the weather.

Compute has a political geography#

Research mapping AI infrastructure finds AI compute hubs in only about 30 countries. Not 195. About 30. And the concentration inside those countries is extreme: Amazon, Microsoft, and Google together hold roughly 70% of global cloud infrastructure, and the infrastructure itself clusters in a handful of territories. More than 90% of the AI data centers that other organizations use are operated by American or Chinese companies. The OECD has an entire research program on the geography of AI compute, mapping what compute exists, where, and crucially whether countries own it or rent it, because owning and renting are not the same right.

The scale question is only starting. The International Telecommunication Union projects AI data centers may need around 68 gigawatts by 2027 and 327 gigawatts by 2030. For scale, entire national grids in most countries are smaller than that second number. The capacity that matters is being built in a few countries, owned by a few companies, and access is granted or denied by a few governments. Everyone else lives downstream of decisions made in three time zones.

India: the biggest user, the smallest owner#

Now the India lens. India generates nearly 20% of the world's data and hosts about 3% of global data centers. Its total data-center capacity sits around 1,500MW, while single facilities elsewhere approach a gigawatt. The government's own parliament answer about where India's advanced GPUs come from states that they are primarily manufactured in one country. You do not need me to name it.

The demand side is enormous and growing faster than almost anywhere. India accounted for 16% of global generative-AI app downloads in 2025, roughly 600 million of 3.8 billion. Jio has 524.4 million mobile users and is giving eligible ones 18 months of Google AI Pro; its own MyJio.AI assistant and AI call agent are in the product line. Airtel bundles free Perplexity Pro, and its management told investors "a few million" users showed up in the first days after launch. OpenAI runs an "OpenAI for India" program listing PhonePe, Pine Labs and CRED as partners; PhonePe has a strategic collaboration to bring ChatGPT to Indian users at scale. Enterprise adoption is faster than almost anywhere: CRED's companion Cleo runs on GPT-4o, GPT-5 and o3 models, MakeMyTrip built its Myra experience on OpenAI APIs, Cars24's agents handle over a million conversation minutes a month, Federal Bank built its Feddy chatbot and generative search on Gemini and Vertex AI, Fynd's commerce agent has processed over 4.3 million customer interactions, Swiggy orders work directly inside ChatGPT and Claude, Air India uses Claude Code, Infosys and TCS signed enterprise-wide collaborations with OpenAI (TCS deploying ChatGPT Enterprise and Codex across its workforce), Anthropic launched in-country Claude inference via Amazon Bedrock for users including Axis Bank, NPCI and IndusInd Bank, Google localized Gemini processing through Indian cloud regions for regulated sectors, and Mahindra put a Gemini agent in control of over 200 functions of a car. IndiaAI, the national mission, has onboarded more than 38,000 GPUs across 14 cloud providers and supports 12 sovereign model teams; Microsoft has committed $17.5 billion to Indian cloud and AI infrastructure; Google is bringing Trillium TPU capacity to Indian regions.

The consumption is world-class. The ownership is not. A survey of Indian AI startups found about 74% rely on proprietary closed models accessed through Western APIs. Competition Commission of India research puts about 67% of Indian AI startups in the application layer, 20% in data, and 10% in infrastructure. India's AI spending is projected around $6 billion by 2027 growing at 33.7%, with $2.5 billion spent in FY25 and expected to more than double by FY28, and India makes up 10-12% of the global AI-model user base. India is among the largest users of AI and among the smallest owners of it. The Economic Survey's warning is one line long: the country must avoid fragile dependencies. The parallel dependencies are not subtle either. India imports more than 85% of its crude oil, imports the ingot and wafer stages of its own solar supply chain, and its 5G networks run substantially on Ericsson equipment. Each of those was once framed as a temporary arrangement.

The 1989 fire#

India set up its Semiconductor Complex Laboratory in Mohali in 1984, three years before TSMC was founded, and worked its way from 5-micron processes down to 0.8-micron CMOS. Then, in 1989, a fire destroyed the fab, and the facility only became operational again in 1997.

Here is the part of this story I find genuinely haunting, because India's position in the pipeline was not always this.

India set up its Semiconductor Complex Laboratory in Mohali in 1984. TSMC was founded in 1987. India was early. SCL worked its way from 5-micron processes down to 0.8-micron CMOS, which by the standards of the late 1980s was legitimate frontier-adjacent work, and India was positioned to be a genuine participant in the industry it now almost entirely imports from.

Then, in 1989, a fire destroyed the fab. The circumstances were never satisfactorily explained, and contemporaneous industry reporting in the year 2000 openly discussed suspicions of arson alongside government foot-dragging. The facility only became operational again in 1997. Eight years of lost time in a business where a single generation is a chasm, in exactly the years when Taiwan, South Korea, and the wider Asian cluster was compounding. The story has a genuinely strange epilogue: SCL still operates today, making radiation-hardened chips for space and strategic uses, one of the few state-owned fabs anywhere, a ghost of the 1984 ambition.

India's attention shifted after 1991 liberalization toward software services, a rational choice that traded owning technology for exporting labor on top of other people's technology. Taiwan built fabs. Japan built materials. The Netherlands built machines. Korea built memory. India built IT services, a genuinely successful industry, built entirely on top of the pipeline these posts describe. The lost decades compound brutally in semiconductors because the nodes compound: a two-decade lag in nodes is not twice as bad as a one-decade lag, it is a different sport.

Running to catch a curve#

The scale of the gap, measured honestly: TSMC hit 90nm volume production in the mid-2000s and 28nm in the early 2010s. The Tata Dholera fab, built on a technology transfer from Taiwan's PSMC, was announced at a 28-110nm plan and has since been reportedly pushed to a 90nm start, with commercial production around mid-2028. TSMC is ramping 2nm and drew 24% of its 2025 wafer revenue from 3nm alone. Many Indian tape-outs still sit at 180nm or get fabricated overseas, and no Indian company designs, fabricates, or packages HBM, or runs CoWoS-class packaging, the exact two bottlenecks part 2 identified as the AI chokepoints. NITI Aayog's own assessment says India imports about 90% of its chip demand and has been running the wrong race, volume over strategy. India ranks only 11th in silicon production and imports its silica feedstock for high-grade uses, which is the ingot problem in miniature: even the sand layer has an import bill. Over 90% of fab equipment and 85-90% of specialty chemicals for the new fabs are still imported. India's parallel dependencies rhyme with each other: crude, solar ingots, telecom equipment, GPUs, all imported at similar rates.

The counter-moves are real and worth taking seriously, because they are the first credible ones in decades, and they read differently from the 1980s effort:

• India Semiconductor Mission 2.0 broadened from fabs to equipment, materials, design IP, supply chains and R&D, and Semicon 2.0 approval openly states the first major silicon fab lands around 2028, starting at 28-110nm, honest about the starting point rather than inflating it
• The NITI Aayog roadmap pivots toward what part 2 showed is the actual bottleneck: advanced packaging, compound semiconductors, wide-bandgap materials and AI-native chip design, plus a national EDA tool access program so designers are not gated by foreign software licenses
• India employs nearly 20% of the world's semiconductor chip-design workforce, with Indian engineers working on 2nm-class designs, an enormous installed talent base whose value is currently captured elsewhere
• The SHAKTI processor family from IIT Madras has taped out, and ISRO built an actual chip (IRIS) on it, small proof that the design-to-tapeout path exists domestically
• Micron opened the country's first semiconductor assembly-and-test facility at Sanand, exactly the OSAT layer India lacked
• Tata signed the PSMC technology transfer, a strategic partnership with ASML, and an MoU with Merck for semiconductor materials; Carnegie's assessment is that the ecosystem is maturing enough that ASML is paying attention
• The IndiaAI mission funds 12 sovereign model teams. Sarvam AI open-sourced its 30B and 105B models, built the Indus model line, and has announced plans for a trillion-parameter model. Zoho built its own Zia LLM for enterprise. The minister for electronics and IT has publicly claimed India belongs in the same group as the US and China in AI development

Stanford's Global AI Vibrancy tool ranks India 3rd in the world on its composite index, 15.36 points behind China and 6.03 ahead of Canada. And the 2026 AI Index's chart of notable frontier models counts 59 American, 35 Chinese, 8 South Korean, with India absent. Both of those facts are true at the same time, and the honest summary is the space between them: 3rd-most-vibrant AI ecosystem on earth, 16% of global gen-AI downloads, nearly 20% of global chip design talent, zero notable frontier models, a fab starting two decades behind the curve, three-quarters of the startup ecosystem reselling Western closed models, and the HBM and packaging industries entirely absent.

There is a human layer to this too, and it is not comfortable. Reporting on Indian workers training AI systems includes people filming themselves doing housework as training data for AI companies, and photo essays of workers training robots that may take their own jobs. The country with 16% of the world's gen-AI downloads is also exporting the labels that teach the models. That is what being downstream of the pipeline looks like at ground level, and it deserves to be said plainly rather than buried in a footnote about the trade deficit.

The market has already done its own math on all of this. India became the least-favored Asian stock market in one major global fund manager survey in 2026, foreign investors broadly sold Indian stocks while raising stakes in AI-linked names, Indian IT stocks face pressure as AI reshapes the services model those stocks are priced on, and India was overtaken in market capitalization by South Korea and Taiwan as the AI wave re-priced who owns the pipeline. Markets are not sentiment here. They are reading the same map this series is reading.

Part 5 closes the series with the synthesis: what the pipeline actually is once you have walked all of it, who holds the kill switches, what the evidence says about whether any of this concentration is producing real value, and what a country that owns none of it can actually do.

The full series#

• Part 1: Dirt, Sand and the Software That Draws Chips
• Part 2: One Machine and One Island
• Part 3: The Buyers and the Export Control Era
• Part 4: The Geography of Compute and the India Lens (you are here)
• Part 5: Nobody Owns the Stack, Everyone Owns a Switch