The Moving Bottleneck: Networking, Power, Memory, and the Race to Win AI

Artificial intelligence is often discussed through models and GPUs. This episode looks beneath that surface, at the power delivery and networking required to make AI work at scale.Host Sriram Viswanathan speaks with Rajiv Khemani, a serial deep tech entrepreneur whose career has tracked several major infrastructure cycles: internet networking, cloud switching, blockchain compute and now AI networking. Khemani reflects on his early work at NetBoost and Intel, his operating role at Cavium, and the founding of Innovium, which Marvell agreed to acquire for $1.1 billion in 2021. He also explains how work on low-power blockchain silicon led his team toward the infrastructure demands created by generative AI.The discussion examines why incumbents often overlook emerging markets, why purpose-built hardware can outperform systems inherited from an earlier technology cycle, and how founders decide whether to keep financing a company or sell while the outcome remains attractive. Khemani describes the concentration risk of selling to a small number of hyperscalers, the fragility of semiconductor supply chains, and why leading-edge chip development now demands much larger balance sheets.The conversation then turns to AI’s emerging bottlenecks. Large models require many accelerators to operate as one computer, making low-latency scale-up and scale-out networks central to performance. The episode explores heterogeneous compute, open networking standards, memory scarcity, AI’s growing electricity demand, and the competition between AI and Bitcoin mining for energy. Speaker Profiles and LinksSriram Viswanathan: Founding Managing Partner, Celesta Capital — https://www.linkedin.com/in/onesriram/ Rajiv Khemani: Co-founder and Executive Chairman, Upscale AI; deep-tech entrepreneur and IIT Delhi alumnus LinkedIn: https://www.linkedin.com/in/rajivkhemani/ Profile and contribution to the IIT, Delhi, Yardi School of Artificial Intelligence : https://scai.iitd.ac.in/rajiv-khemani References Mentioned and Further ReadingUpscale AI : https://upscaleai.com/ Upscale AI Launch Announcement : https://upscaleai.com/press-release/ Velaura AI : https://velaura.ai/ Acquisition of Innovium and cloud data-centre switching rationale, Marvell: https://www.marvell.com/company/newsroom/marvell-to-acquire-innovium-accelerates-cloud-growth-with-expanded-ethernet-switching-portfolio.html Cavium combination and infrastructure semiconductor strategy, Marvell:  https://www.marvell.com/company/newsroom/marvell-and-cavium-to-combine-creating-an-infrastructure-solutions-powerhouse.html Energy and AI, International Energy Agency: https://www.iea.org/reports/energy-and-aiEnergy demand from AI, International Energy Agency: https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai Tokenisation in the context of money and other assets, Bank for International Settlements : https://www.bis.org/cpmi/publ/d225.pdf Leveraging tokenisation for payments and financial transactions, Bank for International Settlements : https://www.bis.org/publ/othp92.pdf Collective communication for clusters exceeding 100,000 GPUs, Meta researchers : https://arxiv.org/abs/2510.20171 Load balancing for AI training workloads, UC Berkeley researchers : https://arxiv.org/abs/2507.21372 Reliability in large-scale machine-learning clusters : https://arxiv.org/abs/2410.21680 Bitcoin: A Peer-to-Peer Electronic Cash System : https://bitcoin.org/bitcoin.pdf Timestamps:</