AMD Will Train 10,000 Delhi University Students On Its ROCm AI Stack

In this storyAMD

AMD and the University of Delhi have signed a memorandum of understanding to train up to 10,000 students in artificial intelligence over the next year.

The agreement brings AMD's AI Engage Developer Program to the university, giving students and faculty hands-on access to ROCm, the company's open-source GPU software platform, along with learning resources and developer GPU cloud credits. The programme will integrate AMD curriculum, technical workshops and cloud-based GPU resources into student learning.

Jaya Jagadish, senior vice president and country head at AMD India, said the collaboration is meant to move students from learning AI concepts to building AI solutions. Vice-Chancellor Yogesh Singh framed it as part of a growing industry focus on giving graduates practical AI skills beyond classroom teaching.

The MoU sits inside a larger commitment. In September 2025, AMD said it would train 100,000 STEM graduates in India in AI and GPU programming over three years, and provide 100,000 hours of free access to its Developer Cloud to Indian researchers and startups. That cloud offers on-demand access to AMD Instinct MI300X accelerators without any hardware purchase. The Delhi University agreement is the first large university tranche of that pledge to be announced.

India already carries unusual weight inside AMD. The company opened its India Design Centre in Bengaluru in 2004 and inaugurated its largest global design centre there in November 2023. Roughly a quarter of AMD's global workforce is based in the country, working across data centre, gaming, PC and embedded products.

Neither party has disclosed the financial terms of the arrangement, the specific courses affected, or how the 10,000 figure will be counted.

ROCm's problem has never been the hardware. AMD's Instinct accelerators compete on specifications. What they lack is the software gravity Nvidia accumulated over roughly two decades of CUDA being the thing universities taught, textbooks assumed and graduate students learned first.

Every engineer who learns CUDA before ROCm becomes a switching cost. Retraining them later is expensive and rarely happens voluntarily. Reaching students before they have committed is the cheapest available attack on that moat, and the only one that compounds.

Which makes Delhi University a more interesting choice than an IIT would have been. DU is a general university with hundreds of thousands of students across its colleges, most of them not in computer engineering. IIT graduates are contested by every employer on earth and would have made a better press release. DU offers volume, and volume is what a hundred-thousand-graduate target actually requires.

The GPU cloud credits are the part doing the real work. A student gets hands-on time on Instinct hardware they could never buy or otherwise access. That is precisely how CUDA became the default: not by being sold, but by being the only thing available when people were learning. AMD is running the same play a generation later, on the one campus system in India large enough to move the numbers.

The limitation is worth stating plainly, because skilling announcements are made far more often than they are audited.

Up to 10,000 students in a year, through curriculum integration, workshops and cloud credits, is a wide and shallow programme by construction. Exposure to ROCm in a workshop is not proficiency in it. Nothing published sets out how many of the 10,000 will write code that runs on an Instinct GPU rather than sitting through a session about one. And the students who go on to serious machine learning work will still meet CUDA at their first job, because that is what production infrastructure runs on today.

What makes it worth taking seriously anyway is where it sits in India's actual problem. Indian chip-design companies have raised $162 million in total since 2023. The value in AI silicon is accruing not in the compute die but in the layer beneath it: interconnect fabric, memory architecture, packaging, rack integration. That was the whole logic of Nvidia putting $3.5 billion into MediaTek this week. Those are systems problems, and they need engineers who understand GPU software stacks at the level ROCm exposes, not engineers who can call an inference API.

On that measure, a generation with open GPU stack literacy is more useful to India than another cohort trained on model fine-tuning. ROCm being open source matters here in a way CUDA never could, because the students can read it.

The open question is who captures the benefit. AMD already employs about a quarter of its global workforce in India. A programme that produces thousands of ROCm-fluent graduates produces, among other things, a deeper and cheaper hiring pool for AMD's own Bengaluru design centre. That is a legitimate commercial motive rather than a criticism. But India has been the world's engineering back office before, and the difference between building capability and supplying it is whether any of these graduates end up starting companies rather than joining one.

For daily, sharp analysis of the biggest moves in the Indian business and startup ecosystem, follow StartupFox.