Storage Is the Spinal Cord of AI: Inside Singtel’s Sovereign AI Cloud

This is Val Berkovici, chief AI officer at WEKA. This is Manoj Prasanna Kumar, CTO of Singtel Digital InfraCo. So, Manoj, we've been working together recently on really cool projects with HTX. And often, we can't talk about what they do, but how did we discover each other? Why did you choose WEKA? What's been most interesting aspect, business, technology, culture, otherwise, of our relationship? We have been building a sovereign AI cloud called Re AI, in short, Re imagine AI. And we have been working with a couple sovereign AI customers who care about storing and processing data within Singapore. And we have been deploying chips early on from h one hundreds to g b two hundreds, all the way to the more latest family of GPUs. So storage has been, in our view, the spinal cord of our AI cluster because the more faster data can be stored and retrieved, the more faster the training jobs can complete, the more faster inferencing jobs can turn back the responses to incoming queries. So we believe in deploying parallel file systems, which are very fast in enabling high performance AI in our sovereign AI cloud. So we picked WEKA as a parallel file system offering, enabling high performance storage for our AI clusters for our sovereign customers on the high throughput of read and write and even leveraging technologies like direct attaching the parallel file system blocks all the way directly to the GPU chips in enabling some cutting edge AI workloads for a mix of enterprises, sovereign AI customers, and so on. I love that you brought that up because, again, that's the product that I'm leading right now, Augmented Memory Grid. So it's fascinating that the Linux kernel can boot up a server and see a file system, but then the GPU kernel can start up with a model and inference server and see the same SSDs as memory. So what are your thoughts going forward, whether it's that particular detector or other roadmap items going forward? What's most exciting to you about our collaboration in the future for HTX and otherwise? The way we see storage is not just the throughput and it's not just the top of the class parallel file system offering, but we also see supporting storage infrastructure. For example, object storage, which is extremely important to store checkpoints, to store weights, to archive models, and so on. And it is also extremely important to balance cost of storage as customers scale, adopting petabytes of storage capacity. And the third aspect is beyond the types of storage, the parallel file system and the object storage. We also believe in the storage data platform layer in terms of multi protocol support, being able to convert data from one format to the other format. And that is where the real value is in offloading the overhead of managing the different formats of the data and managing the different sinks in which the data is deposited and retrieved, which is very critical to ensure that a customer has a very seamless experience deploying AI in our cloud. Fantastic. I really appreciate this deep dive you've given us today on our collaboration. Looking forward to coming back soon and talking even more. Thank you so much, man. You're welcome, Neural. Thank you.