US export restrictions on AI chips: Implications for India
Business Standard, 20 January 2025
The US has proposed a group of restrictions on export of top end AI chips. AI chips are being treated, by the US government, as defence equipment and many elements of defence restrictions are being brought into this space. The use of these chips in India faces new restrictions as compared with before. Export of software and software-enabled services is central to Indian growth. From an Indian economic strategy viewpoint, however, the new restrictions do not make an important difference. The path to doubling of services exports in the next decade remains fully available. The new environment does have some interesting implications for some actors in India.
The Indian growth strategy
The one successful piece of the Indian economy is exports of IT and IT-enabled services. From 2017-18 to 2023-24, merchandise exports grew at 2.8% per year in inflation-indexed dollars (doubling every quarter century). IT services exports grew at 9.17% real (doubling every 7.5 years) and `other business services' grew at 11.86% real (doubling every 5.8 years). Services exports were 341 billion USD in 2023-24. The present growth rates suggest a doubling in a decade!
These facts are central for the Indian growth strategy. Merchandise exports has not worked, and we should wonder what the far-reaching changes are, which would enable them to start growing well. But services exports growth has been strong. The Indian growth agenda requires nurturing this field, and getting that hoped-for additional 341 billion USD of exports per year over a decade.
Services exports from India
What does this entail? At heart, Indian services personnel is doing work for top global firms who are headquartered in places like the US, the UK, Japan, Germany and France. This is done partly through GCCs (FDI into India) and partly through arms-length contracts with Indian services firms. The puzzle of Indian economic policy is to keep this engine purring.
The new US restrictions [Prasad and Sahu, 2025; Framework for Artificial Intelligence Diffusion: Interim final rule; request for comments from the Bureau of Industry and Security, Department of Commerce, US Federal government; Rand corporation primer; Chinatalk podcast episode] are part of the global tech wars [EiE Ep64, The global tech wars] and the third globalisation [EiE Ep17, The three globalisations], aimed at hindering China's access to AI technologies. They do not interfere with growth and success in India in services exports. At heart, a buyer like J P Morgan has to figure out how it will get its AI data centres humming e.g. in the US. After that, Indian genius (employees of J P Morgan in India or employees of Indian firms like Infosys) will connect into these global systems and do complex and well-remunerated work.
Some people are excited about Indian global leadership in AI technology. The fact that persons like Sundar Pichai and Satya Nadella are of Indian origin should not make us think that Indian firms or Indian universities are contenders in global AI innovation. Some of the rhetoric on India's place in the AI game is a bit exaggerated. We in India should be focused on India's growth strategy, on improving prosperity in a poor country, and not on pride.
For an analogy, look back at how Indian software services exports got to $163 billion/year in 30 years. This was done with a negligible role of India in development of the hardware (e.g. Intel CPUs), the operating systems (e.g. Unix) or foundational software (e.g. databases) which were all done in the West. India got positive externalities from the array of technologies developed abroad, from the transistor to Unix and the Internet. Western firms supplied the technologies (while navigating export controls in the US), the capital, and the markets for the Indian services export revolution. The Indian state contributed deregulation (e.g. telecom and capital flows) and fostered high end knowledge in India. No CPUs were given to firms by the government. These ways will work for the next doubling.
AI applications by domestic firms
Turning away from exports, what restrictions will Indian firms face for domestic AI applications? There are three pathways: (a) Indian firms can freely rent server capacity from cloud vendors operating abroad, as before; (b) Annual imports per firm, below 1700 Nvidia H100 chips (which cost about $25,000 each), are unregulated, which takes care of the needs of almost any end-user firm in India and (c) AI applications that require `inference' can be done by cheaper chips like the Nvidia H20 or consumer GPUs, where Indian purchases are unrestricted.
The restrictions are focused on data centres that will be used for training AI models: a field where India is largely absent. Large data centres are being built by domestic players like Tata Communications, L&T along with E2E Networks, Yotta, CtrlS and Jio. All of them, put together, now face a limit of 100,000 chips per year (going up to 320,000 in 2027). These are large values compared with present usage. For a sense of scale, the biggest single computer in the US which is used for nuclear weapon development has 50,000 chips.
Indian firms that want above 1700 chips/year will require `National Verified End User' ("NVEU") authorisation. The American side has been alarmed by events in India, e.g. reports of a company in India which imported 1100 Nvidia chips from Malaysia and re-exported them to Russia for $300 million. Actions by the Indian state which block leakage of high technology to Russia, China and Iran will help more Indian firms get to this NVEU status.
AI chips and high skill functions placed in India by global firms
More AI knowledge will come to India when companies headquartered in countries like the US obtain `Universal Verified End User' authorisation. They need to prove, to the US federal government, that they have adequate systems to protect their equipment in India. Once this is done, upto 7% of their global chip count can be placed in India. The top global firms are adding over 100,000 chips per year, so this opens up a path to over 7,000 chips (and associated knowledge flows) to India per year per firm. The top global firms have many choices where they can place their equipment. The Indian government will have to engage in give and take with them, to improve their prioritisation of placing high end work in India.
Back up to the media page for the year
Back up to Ajay Shah's home page