#122 Satellite Internet Wars
In this edition of Technopolitik, Ashwin Prasad talks about sattelite internet and SpaceX’s recent partnerships with Airtel and Jio. In the coming weeks Ashwin will be writing a larger series on everything satellite internet. Rijesh Panicker follows, with a piece that looks at India’s AI strategy.
This newsletter is curated by Adya Madhavan.
Antarikshmatters: Starlink Lands in India?
— Ashwin Prasad
Satellite internet made headlines in India last week when Airtel and Jio announced partnerships with SpaceX. Although details are scarce, both telecom giants will tap into their extensive retail networks to distribute Starlink terminals and assist customers with installation and setup. These arrangements provide the telcos with a new revenue stream through distribution commissions, while SpaceX gains access to the world's most populous country with many underserved regions. The partnerships also lend Starlink credibility by association with India's most well-known telecom providers. These agreements are contingent upon Starlink receiving authorisation from the Indian government to operate in the country.
Massive satellite constellations orbiting Earth enable global satellite internet coverage. SpaceX's Starlink leads this space race with over 7,000 operational satellites as of last month. This number continues to grow rapidly, giving Starlink an overwhelming advantage over competitors in both operational capacity and future potential. This explains why Jio and Airtel, despite having plans of their own to provide satellite internet in the future are making use of this opportunity with Starlink now.
These deals appear to be imbalanced, leaning in favour of Starlink in the short term. However, given the duopolistic nature of the Indian telecom market, this makes business sense. Once Airtel committed to working with SpaceX, Jio had little choice but to follow suit or risk ceding ground to its competitor in a potentially transformative market segment.
These partnerships likely represent just the first step toward broader collaboration in satellite internet, where opportunities abound. Currently, Airtel and Reliance will sell Starlink terminals—equipment similar to direct-to-home satellite TV systems, consisting of an antenna and WiFi router. The future will bring a lot more possibilities. SpaceX has already tested direct-to-cell services that could become operational within months.
As this technology evolves, its potential reach and penetration are extraordinary. Internet connectivity is fundamental to our modern world, enabling countless integrated capabilities and supporting global commerce and economies. Demand will only increase. Policymakers and stakeholders must understand satellite internet's importance and potential.
In upcoming articles, I will explain how this technology works and why it matters.
Technopolitik: Empanelment, Subsidies, and AI Compute: Is India’s Strategy Future-Proof?
— Rijesh Panicker
The Ministry of Electronics and Information Technology (MeitY) recently announced the launch of a continuous empanelment process for AI cloud service providers, which allows firms to apply on an ongoing basis to supply GPUs and related AI compute services. This follows the announced launch of the GPU access portal, after all selected bidders, including L2 and L3 bidders, agreed to match the L1 bid price for compute services.
Against the originally planned 10,000 GPUs, now close to 19,000 GPUs will be made available, of which nearly 14,000 will be Nvidia (12,800 H100 and 1480 H200) and the rest a mix of chips from AMD, Intel, and AWS.The bidder’s have also offered significant discounts to market prices, ranging from 11% - 80% of market price, depending on the chip and module.
The government’s move to use private providers of GPU and cloud compute, instead of procuring GPUs directly, supports the development of India based GPU and compute service providers. In addition, the continuous empanellment process (if done well), could prevent cartelisation amongst the service providers and allow newer providers to come into the market.
At the same time, it is worth asking if the L1 pricing is sensible and sustainable in the long run? First,a pure lowest cost (L1) structure, that does not account for actual operational performance, creates incentives for the providers to manage compute performance downwards in an attempt to reduce replacement capital expenditure.
Second, the fact that all of the providers agreed to the lowest cost (at significant discount to market) could be a sign of current low private market demand for GPU compute in India, which may seem a good reason for the government to step in. In the long run however, the ongoing subsidies being offered by the government (upto 40%) may perversely cap private demand, as consumers restrict their use of GPUs to the subsidised amounts. Back of the envelope calculations show that the cost of training a Llama3.3 (about 39.3 million GPU hours; Rs. 390 Cr) or a Grok3 (45.6 million GPU hours; Rs. 456 Cr) will cost the government about Rs.157 Cr and Rs.180 Cr respectively in subsidy, while training a true reasoning model like OpenAI’s o3 may cost much larger amounts. Should we really subsidise private companies to build foundational AI models at these costs?
Finally, demand for GPUs globally is likely to go up, driven by large AI infrastructure projects like InvestAI in the EU, which is looking to invest 20 billion euros in 4 AI gigafactories (100,000 chips each) or OpenAI’s Stargate project, which is looking to invest $100 billion this year and a total of $500 billion to build AI datacenters across the US. In addition, with the restrictions under the AI chip diffusion framework, supply of advanced AI chips are likely to go down. As a result, we are either going to see a gap in supply or significantly higher prices in the future.
India’s choice of building it’s sovereign compute infrastructure may seem like a good move in the short term - it provides ownership of our own compute and supports the development of an India based AI IaaS (Infra as a Service) market. However, an interesting and perhaps more robust alternative, would have been to use the 4,500 Cr allocated to compute as a co-financing mechanism, for building compute infrastructure.
Private markets should be allowed to function freely and democratising access to academia, researchers and others can be done through other means. In the long run, this will give us the best means to keeping pace with the future. This would also provide a simple and elegant solution for what happens after the IndiaAI mission sunsets after four years.