The rapid adoption of AI is putting immense pressure on data center infrastructure, especially when it comes to solving connectivity bottlenecks that hinder efficient GPU utilization. As AI, particularly generative AI (GenAI), grows in scale and sophistication, hyperscalers like Amazon, Microsoft, Google, and Meta are investing billions to address these challenges.
A key figure in this transformation is Sanjay Gajendra, COO and co-founder of Astera Labs, a semiconductor company pioneering solutions to optimize AI infrastructure. He highlights that hyperscalers will invest around $300 billion next year to upgrade their infrastructure—a 25% increase from this year and nearly double the investment of 2023. Gajendra notes that this “arms race” is necessary to keep up with the increasing demand for AI. One of the primary issues faced by AI systems is the underutilization of GPUs, the critical AI processing units. Current reports show that GPUs are only about 51-52% utilized, with nearly half the time spent waiting for data or memory due to slow or inefficient connectivity. This underperformance is holding back the full potential of AI workloads, especially in large-scale applications that require enormous clusters of GPUs. Addressing this requires advanced connectivity solutions. As Shivananda Koteshwar, MD at Astera’s India site, explains, AI servers now feature a mix of GPUs, CPUs, NIC cards, and accelerator cards, all of which require robust and reliable connectivity across various protocols like PCI, Ethernet, and CXL. The challenge is ensuring high-speed communication between these components without introducing errors or inefficiencies. The need for low-latency, high-throughput, and low-power solutions is more critical than ever.
Astera is stepping up to this challenge by providing connectivity solutions that eliminate data, networking, and memory bottlenecks in AI systems. Their work is vital in powering hyperscaler AI infrastructure around the globe. In addition to these efforts, silicon photonics is another emerging solution for improving connectivity. Jitendra Chaddah of GlobalFoundries highlights how integrating optical components into semiconductor chips can drastically increase data throughput while consuming less power than traditional electrical interconnects. As AI expands into edge computing, Chaddah also points out the rising importance of efficient edge connectivity. The future of AI hinges on overcoming connectivity limitations. With advancements like Astera’s solutions and silicon photonics, the infrastructure needed to support AI’s exponential growth is quickly becoming a reality.
