The GPU Queue Is the New License Raj

The CSR Journal Magazine

I keep asking myself an uncomfortable question. India has one of the largest engineering workforces on earth, some of the cheapest mobile data anywhere, and a startup scene that has produced 122 unicorns. So where is the Indian frontier model? Where is our answer to OpenAI, Google, Anthropic or DeepSeek? We have summits, missions, portals and acronyms in abundance. What we do not have is a globally competitive model that developers abroad choose by default.

The comfortable explanation is that we started late. I do not buy it. DeepSeek did not start early either. What it had was capital, compute and room to experiment. Indian entrepreneurs are offered a different bundle: thin capital, rationed compute, regulation that arrives before revenue, and a government that wants to be banker, landlord, referee and cheerleader at the same time. Let me name the problems one by one.

The money is not there

Stanford’s 2026 AI Index puts US private AI investment in 2025 at $285.9 billion, about 23 times China’s $12.4 billion, and Europe sits far behind at roughly $20.9 billion. India does not even feature in that headline comparison. Indian AI startups raised just over $643 million across 100 deals in 2025, while American AI startups pulled in $121 billion. One year of American AI funding is roughly 188 times what our entire AI startup sector raised.

Indian AI money went mostly to application businesses rather than capital-intensive model development, and seed funding fell 30 percent to $1.1 billion. Of course, founders build wrappers on top of foreign models. Wrappers are what the cheques can afford. One recent analysis claim that Indian AI startups raised about $3.94 billion in the first quarter of this year alone. I hope that holds, but one good quarter is a data point, not a trend.

The deeper problem is who pays for research. A Carnegie analysis finds private industry funds only about 36 percent of India’s R&D, against more than 70% in the US, China and South Korea. Our biggest companies grew rich on services and outsourcing.

Why would they gamble billions on frontier research when selling engineer hours already pays? Until corporate India behaves like a research funder, no government mission can fill that hole.

Compute: the promise and the queue

The IndiaAI Mission was approved by the Union Cabinet in March 2024 with an outlay of Rs 10,372 crore, which is roughly $1.2 billion, spread over five years. Set that beside the $285.9 billion Americans privately invested in a single year. Our national ambition is about 0.4% of 1 year of their spending.

The flagship idea is a common compute pool; subsidised GPUs rented to startups and researchers. On paper it is a success. Officials say 38,231 GPUs have been onboarded and 190 projects approved, and credit where due, the original target was only 10,000. In practice, founders lived in the gap between brochure and server room. The mission earlier targeted about 34,333 GPUs, yet only about 17,374 were installed and operational, creating a supply crunch. Analytics India Magazine reported that three months after Sarvam was chosen, only Sarvam had received about 1,500 GPUs, while Gnani, Soket and Gan had received nothing. A build plan of six to ten months does not survive a three month wait for hardware

Then there is the question of ownership. The bids for the mission’s GPUs covered the H100, H200, MI325 and MI300X, and every one of those chips is designed abroad. The largest new infrastructure pledges come from Microsoft at $17.5 billion and Amazon at over $35 billion by 2030. I welcome the investment. But can we call it sovereign AI when the sovereign is renting? The US hosts 5,427 data centres, more than ten times any other country. Rented time on a shared pool, allocated by a committee, is not a compute strategy. It is a waiting room.

The state as kingmaker

Venture capital is brutal but fast. It makes many bets, kills bad ones early and learns quickly. The IndiaAI model works the other way. A ministry shortlists a handful of teams, 12 in the first phase, and hands them GPUs, funding and a launch stage. Some are genuine startups such as Sarvam, Soket, Gan and Gnani. Others are not: Business Standard listed Tech Mahindra among the startups selected to build indigenous models. Why does an IT giant need a public subsidy to build a model? And what happens to the thirteenth team, with a better idea and no contact in Delhi?

Committees are poor at picking winners in a field that changes every quarter. They are excellent at producing launches. Sovereign models from Sarvam, BharatGen and Gnani were unveiled at the IndiaAI Impact Summit 2026, and the early models were modest, with Sarvam-M at 24 billion parameters and BharatGen at 3 billion. They are respectable first steps and nowhere near the frontier. I am not mocking the engineers, who work with what they are given. I am questioning a system that measures success by announcements instead of customers and revenue.

Look at the rest of the scheme menu. A global programme sent a second cohort of 10 AI startups abroad, with Station F in Paris and HEC Paris as partners. There are plans for 500 AI data labs and an AI governance framework. The Union Budget allocated Rs 2,000 crore, and the plan is now to double the mission’s corpus of Rs 20,000 crore.

Bernstein has warned that India’s AI push risks being sidelined by global tech giants. Every scheme has a logo and a launch event. I would like to see the published customer outcomes. Meanwhile we pay to ship promising founders to Europe while founders at home wait for GPUs. Is that the right order of priorities?

Rules before revenue

Nothing says “we want you to win” like a compliance calendar. India has no standalone AI law, yet the DPDP Act, the IT Act and the Copyright Act already bind AI companies. Full compliance with the data protection rules is due by May 2027. The industry body IAMAI warned that the treatment of publicly available personal data poses practical challenges for companies training models and asked for an exemption for AI training as an interim step.

Then came the synthetic content rules. The IT Amendment Rules 2026 were notified on 10 February and came into force on 20 February. The October 2025 draft wanted a watermark covering 10% of the surface of synthetic content, an idea later dropped for a qualitative standard. Industry groups warned that labelling and metadata duties would burden startups and SMEs. A ten-person lab in Pune cannot absorb a compliance regime drafted with Big Tech’s legal departments in mind.

Copyright is the real landmine. A DPIIT committee is still examining how copyright law should treat AI training, and ANI’s case against OpenAI over training data is the sign of litigation to come. So, a founder today must guess whether scraping Hindi newspapers will invite an injunction. Uncertainty is a tax, and it is levied before the first rupee of revenue.

Why are we writing rules for the deepfake panic of the month before we have built the models that could help detect deepfakes?

Talent we have, careers we do not

Talent is not the shortage. Zeki counted 50,460 AI authors and inventors in India in 2025, second only to the United States at 220,520 and ahead of Germany. LinkedIn’s index shows AI skills appearing in Indian profiles at three times the global average. ThePrint’s reading of the Stanford report is that talent may be staying home, but not because India has resolved its weaknesses, only because the US has become less accessible. That is luck, not policy.

Our best engineers still face the same choice. They can take a stable job at a services firm or a global capability centre with a good salary, or join a risky lab with no compute, no seed cheque, no clear law and no anchor customer. Rational people choose the salary. Our problem is not producing engineers. It is giving them something to build.

The founder flip and the missing customer

Ask around and you will find that many ambitious Indian founders incorporate abroad, in Delaware or Singapore, because capital, courts and tax rules feel more predictable there. Every flip is a vote of no confidence in our own system. We then spend years debating how to bring them back, instead of asking why they left.

A model is only sovereign if someone buys it. Indian enterprises and government departments default to the safest purchase, a global brand with a global reference list. Procurement that rewards the lowest bidder and the longest track record punishes a young lab by design.

How can a startup show three years of government clients when the government will not be its first client? No GPU subsidy fixes that. I would rather see the state become an anchor buyer, with real contracts for Indian language tools in courts, agriculture advisory and public health, awarded through open, benchmark-driven tenders.

Language is supposed to be our moat. India has 22 scheduled languages and hundreds of dialects, and the digital text available in most of them is thin compared with English. The usual response is another mission to collect datasets. Yet data collection without a clear legal basis is exactly what the copyright and DPDP uncertainty blocks. We announce a language moat and then build a legal moat around it.

What China shows, and what it does not?

By March 2026 the top US modelled its nearest Chinese rival by just 2.7%, despite the vast gap in private investment. Money is clearly not everything. But the same report cautions that private figures understate China’s spending, since government guidance funds deployed an estimated $184 billion into AI firms between 2000 and 2023. The state did spend heavily.

I am not saying Beijing’s method should be copied. I am saying that state money deployed as patient capital is a very different animal from a ministry choosing a dozen names and supervising their GPU queue. Is India building a fund or a gatekeeper?

What I would change

First, turn compute into vouchers. Any registered startup with a credible plan gets credits it can spend at any empanelled provider, so the money follows the founder, not the committee.

Second, publish a monthly delivery dashboard showing GPUs promised, installed, allocated and used. A mission that cannot show its own queue should not ask for a doubled corpus.

Third, settle the law. Give training a clear copyright rule and a clean path under the DPDP framework so founders can price risk before they build.

Fourth, make the state a real customer through open tenders, so Indian models earn revenue before they earn applause.

Fifth, draw long-term domestic capital such as pension and insurance money into deep tech funds, because a private research culture will not grow from public grants alone.

Sixth, replace the summit economy with annual audits that anyone can read.

The verdict

Late is not fatal. The frontier has been closer to a sprint of freedom and capital than a marathon of subsidies, and the gap between the top two nations has nearly closed within a few years.

India can still matter. But not if we confuse activity with progress, and not if every rupee of ambition must pass through a ministry before it reaches a founder.

How many of the 38,000 GPUs are running a paying customer’s workload today?

Who audits the mission, and where is the report?

Why does a founder need permission to experiment but no permission to be failed by delays?

India does not lack brains, data or hunger. It lacks a system that trusts its entrepreneurs enough to get out of their way.

Views of the author are personal and do not necessarily represent the website’s views.

Dr. Jaimine Vaishnav is a faculty of geopolitics and world economy and other liberal arts subjects, a researcher with publications in SCI and ABDC journals, and an author of 6 books specializing in informal economies, mass media, and street entrepreneurship. With over a decade of experience as an academic and options trader, he is keen on bridging the grassroots business practices with global economic thought. His work emphasizes resilience, innovation, and human action in everyday human life. He can be contacted on jaiminism@hotmail.co.in for further communication.

Long or Short, get news the way you like. No ads. No redirections. Download Newspin and Stay Alert, The CSR Journal Mobile app, for fast, crisp, clean updates!

App Store – https://apps.apple.com/in/app/newspin/id6746449540

Google Play Store – https://play.google.com/store/apps/details?id=com.inventifweb.newspin&pcampaignid=web_share

Latest News

Popular Videos