The rapid rise of artificial intelligence workloads has reshaped the data center industry, vaulting companies into the limelight, and pushing sector investment ever closer to the trillion-dollar mark.
For Nvidia, the designer of the key GPU architecture that has made this all possible, the rise has been even more meteoric. Once primarily concerned with churning out gaming chips, the company has been remade as the on-again-off-again most valuable enterprise in humanity’s history, all thanks to a data center boom.
The Telecoms Supplement
Innovation across the networks
For the telecoms giants – most of whom are used to thinking in decades and slowly investing in long-return infrastructure between expensive and misguided forays into media ventures – this has all moved at a speed and scale that they’re not used to.
Nvidia’s Ronnie Vasishta, a softly-spoken executive at a company led by grand orators, is there to help them find their place in this AI world, preaching the gospel of the GPU.
“My responsibility at Nvidia is all things telco,” Vashista tells DCD in a 2024 interview. “It's a pretty broad brush of responsibility - it ranges from telcos that are using or looking to use AI and platforms and accelerated compute for things like customer care, customer experience, internal use of large language models for productivity improvements such as the way you route technicians and trucks, and then extends all the way into the virtualization of the RAN network.”
Telcos are also starting to look at deploying “AI factories around the world,” he says, using Nvidia’s nomenclature for AI data centers. “Not all telcos have experience with data centers, and how they deploy them, and how they build them, so we create the reference architecture.”
The Western telecoms giants have a long but troubled history with data centers. After surviving the telecoms winter of the post-Dotcom boom and loading up on debt for various misadventures, most sold off their data center assets only a few years before Covid-19 and AI dramatically increased the value of the facilities.
Experiments to turn central offices into Edge data centers have mostly failed to meet the hype, and telecoms have been forced to sit back and watch hyperscalers become the new infrastructure barons of the modern age.
“They have burnt fingers and they felt that that was not a business model for them,” Vashista admits. “But things have changed - you're training foundation models, you're deploying GPUs for inference or deployment of those models. And so the task is somewhat different than it would have been in the past.”
Elsewhere in the world, particularly for Asian telcos, the experience hasn’t been so traumatic. “There are other telcos around the world that just haven't gone through that experience,” Vashista says. “They see this rapid demand for AI, and they're like ‘we need to be in there.’”
SoftBank, the Japanese telco bankrolling Stargate, is deploying AI clusters across its homeland. India's Reliance Industries, a conglomerate that does everything from telecoms to petrochemicals, has promised to build the "world's largest" data center in Gujarat, filled with Nvidia Blackwell GPUs.
Earlier this year, South Korea's SK Telecom began offering GPU-as-a-Service (GPUaaS) with Nvidia chips, in partnership with US neocloud Lambda - itself an Nvidia investment.
Outside of the West, where AI models covering English, French, and Spanish are plentiful, there is also an opportunity for telecom operators to be more invested in the development of AIs themselves.
“Some telcos want to train in the local language, and local dialects,” Vashista says. “I was talking to a telco just yesterday, that services a very large indigenous population, and they want to keep that language alive.”
Way back in 2023, Reliance Industries’ Jio Platforms partnered with Nvidia to develop a large language model for the various languages spoken in the country. It's not clear how ambitious Reliance plans to be, but it has ample scope - while the Indian constitution recognizes 22 official languages, the 2001 Census recorded 122 major languages and 1599 other languages.
The next part that Vashista’s small team hopes to disrupt is radio access network (RAN), the part of a telecom network that connects mobile devices to the core network using radio signals, a combination of hardware such as base stations and software.
“The approach we're taking that is unique is that we want it completely software-defined, no custom hardware in there, no custom PCIe cards,” Vashista says.
From 2005 to 2018, Vashista was the CEO of eASIC which, among other things, developed custom RAN hardware. Now, he argues, most of the software can be run on GPUs.
“That we can ride the wave of Nvidia's hardware platform, and we can ride the wave of deployments of infrastructure,” he argues.
“When a company or telco decides to deploy AI infrastructure, we can say, ‘RAN is a software overlay, nothing different, just orchestrate it the same as with any other workload - you can run RAN by day, when everyone's using their phones; by night, you start running AI training, and you're generating revenue from that.”
As for where the base stations and antennas should be placed, Nvidia thinks it has a solution for that. Building upon its broader industrial metaverse effort Omniverse, the company’s Aerial Omniverse Digital Twin hopes to allow telecom companies to replicate deployment environments.
It’s a digital twin platform that is radio frequency (RF) aware,” he says. “You can place antennas in the digital twin and place hundreds of mobile users moving around the scene.”
The claim is that the model is physics accurate, allowing you to simulate the electromagnetic waves in the environment with “buildings that are glass or concrete, or you can have foliage or different weather types, and you can see the effects of this on the RF propagation.”
Vashista adds: “Not only are the simulated antennae distributing the RF signals, but you're now collecting data as well with software-defined base stations. So you can say, more output power. Or you can tune the beam a different way.
“You can optimize it. Maybe 1,000 users exit the scene all of a sudden in downtown New York. What would happen? Maybe I've got a bunch of drones flying around. I want to make sure those drones are always connected as they're going through the suburbs of New York or Tokyo.”
The digital twin can be small-scale, he says, based on LiDAR data or other sources, or it can span entire cities.
The effort is still in the earlier stages, and is part of Nvidia’s 6G Developer Program. “I know, it sounds like a long way off,” Vashista says, picking up on DCD’s skepticism. “In fact, some telcos don't like to talk about 6G, they're still trying to monetize 5G.”
But, he says, “we took an approach that, if you're going to deploy in 2030, you need to start researching now. Some of the tools from the 6G research effort will still be usable for 5G and 5G Advanced, he notes - reiterating that the general use of GPUs means that they can transition across generations.
“The operators don't need to invest in 6G-specific equipment to enable them to run 6G and service 6G on the compute side, which means that that big investment cycle can be targeted more towards the compute platforms, which are more generic and can generate revenue as well,” he says.
Convincing all the telecoms companies that have struggled to adapt to take this leap is still an uphill battle, however. But Vashista, now in his fourth year at Nvidia, believes that momentum is on his side.
“At Nvidia, we like to say that we move at the speed of light. Maybe telcos historically have moved at the speed of sound. But what we're starting to see is they're really leaning into the use of generative AI, because it offers them a business model variation that they haven't had before.”