Telecom companies that have spent decades and billions of dollars engineering and constructing purpose-built wireline and then mobile networks and all of the hardware and software that goes with them may not take it kindly when you tell them someone else has come up with a better approach to doing all that, but that is essentially what Nvidia and Softbank are set to do.
“5G can now be run as a software-defined workload in an AI factory [data center],” said Ronnie Vasishta, senior vice president of telecom at Nvidia, during a recent media and analyst briefing..
The partners, who not so long ago were at opposite ends of a failed Arm acquisition (Softbank wanted to sell Arm; Nvidia wanted to buy it) recently announced they are collaborating on a new platform for generative AI and 5G/6G applications that leverages Nvidia’s GH200 Grace Hopper Superchip, MGX reference architecture, Bluefield-3 DPUs, and Aerial software for creating cloud-native 5G networks, and will transform 5G and future mobile network technologies into virtualized, software-defined overlays that can be hosted at next-generation AI data centers. SoftBank said it plans to roll out a series of new, distributed AI data centers across Japan that will do just that.
By making 5G radio access networks (RANs) and even more modern virtualized RANs into overlays in an AI data center, Nvidia and Softbank say they can solve a persistent problem for network operators in the 5G era–that networks are “over-provisioned” but “under-utilized.” Hosting 5G as an overlay in an AI data center will reduce the capacity disparity, make managing 5G networks less costly and more dynamic in their ability to allocate capacity when needed.
Traditional telecommunications [mobile] networks are built for a single purpose,” Vasishta said. “They're built to deploy and compute the signal processing and the RAN infrastructure, but as built they are over-provisioned for peak demand. With the growth of generative AI applications, the compute requirements for those networks is going to grow, but because you're over provisioning and the average utilization of those networks is about on average 25% of peak. We see a significant under utilization of the network that's being built and the RoI on the 5G investments has been relatively low.”
The answer to this, the partners believe, lies in migrating the traditional CPU-centric, single-stack architecture of legacy telecom data centers to the type of full-stack, GPU-accelerated architecture that Nvidia has been championing for enterprises in so many other industries.
The big challenge lie in convincing telecom companies that have been gradually virtualizing RANs little by little to think bigger and do it faster as AI transforms everything around them. Softbank’s experience will provide a model for them to do that.
Vasishta acknowledged the challenge, but added that the new architecture will allow for more revenue generation and monetization of network assets with lower investment going forward. For example, as operators migrate to 6G and whatever comes next, these new network technologies can be deployed for less costs and on the same hardware. That is a compelling argument for telecom companies to bury their pride about past network investments if they can save a few more dollars on future plans.