Rakuten is not buying the Nvidia pitch for AI-RAN in JapanRakuten is not buying the Nvidia pitch for AI-RAN in Japan

Stock up on some artificial intelligence (AI) chips, lodge them in your telco facilities and get rich from new services while the chips simultaneously do a more efficient job of supporting your radio access network (RAN). In a nutshell, it's the latest pitch from Nvidia, which has pivoted from hyperscalers to network operators as it seeks new customers for its graphical processing units (GPUs), the AI whizzkids of the semiconductor world. But Rakuten's Sharad Sriwastawa is not buying it.

The boss of Rakuten Mobile and Rakuten Symphony – the telco subsidiaries of the Japanese e-commerce player – examined the economics of putting Nvidia's GPUs at the network edge and evidently did not like what he saw. In Japan, a telco would not be able to link enough mobile sites to a distributed unit (DU) equipped with a GPU for the economics to make sense, Sriwastawa told reporters on a call this week.

"We evaluated that and at least from our perspective – unless we talk about GPU sharing for other services – the cost-benefit analysis that we did was not good enough to go down that route," said Sriwastawa in response to a Light Reading question. "In Japan, we have a limitation of about 20 kilometers where the DU workload can be, and that limits the number of sites that we can connect to these edge data centers. If we start deploying a GPU, which has huge capacity, not everything will be working to full capacity."

But the criticism comes days after Nvidia announced plans for an AI-RAN deployment by SoftBank and will prompt further scrutiny of the move. The domestic telco rival to Rakuten has already connected 20 mobile cells to a single Nvidia-powered GB200 NVL2 appliance. Nvidia expects it to connect "hundreds" more sites next year, while SoftBank says a commercial deployment will start in 2026.

To Nvidia and beyond

In a technical blog, Nvidia claimed its technology is about 40% more energy efficient than a custom RAN product and brings a 60% saving compared with an x86-based virtual RAN. "The comparison is done for Watt/Gbps and assumes the same number of cells for each type of solution," said Nvidia when asked where those savings come from. "The gains for GB200 NVL2 are possible due to the very high density of cells achieved per server, which would otherwise require multiple basestations."

Nvidia's product combines one of its GPUs with a more humdrum central processing unit (CPU) based on the architecture of Arm, a SoftBank-owned but UK-headquartered chip designer. The GPU handles Layer 1, an especially demanding part of the RAN software, leaving other functions to the CPU. Nvidia also contributes a data processing unit for the fronthaul connection to the site.

But a centralized RAN, substituting a DU at an aggregation point for multiple appliances or servers at mast sites, is doable without Nvidia. And Rakuten appears to have done it. "In terms of edge data centers, we have more than 1,900 now where the DU workloads are," Sriwastawa revealed this week.

Those edge data centers today support a network comprising more than 65,000 4G sites and around 34,000 5G ones, up sharply from the 11,600 5G sites Rakuten reported at the end of last year. Broadly, it implies a ratio of one edge data center to every 34 4G sites.

Sriwastawa's comments about DU distance limitations, however, raise doubts about the real number of sites Rakuten can support from a single edge data center. It's still unclear if deployment constraints have forced Rakuten to maintain DUs at mast sites across large parts of its footprint. The reason for the distance limitation is also not clear, but fiber is widely available in Japan compared with most other countries.

In place of Nvidia's GPUs and CPUs, Rakuten has been hosting RAN software on Intel's x86-based chips. It has also been working to integrate Qualcomm's QDU-branded "accelerator," a custom chip supported on a PCIe card, with its portfolio. "We are almost finalizing the QDU solution," said Sriwastawa this week. The Qualcomm chip would play a similar RAN role to an Nvidia GPU, looking after the Layer 1 functions while other software is left to the Intel CPU.

Nvidia's big pitch, of course, is that its GPUs can double up as RAN accelerators while they also play their more usual role of supporting AI applications. And the sales message accompanying the SoftBank update is that a telco can earn $5 in new AI-related revenues over a five-year period for every $1 it invests in AI-RAN infrastructure. Without endorsing those lofty forecasts, Sriwastawa acknowledges the attractions. "If we can do slicing of GPUs and provide other services, that is where we see it will be useful," he said.

Profitability creeps into sight

The sharp increase in 5G sites over the last year is one of the most notable network changes at Rakuten Mobile, which has previously appeared far behind rival telcos NTT Docomo, KDDI and SoftBank on 5G rollout. And despite the 5G investment, Rakuten remains confident its network-related capital expenditure will not exceed 100 billion Japanese yen (US$650 million) this year, down from JPY167 billion ($1.1 billion) in 2023.

The company has been under pressure to slash costs after racking up cumulative losses of nearly $7.3 billion at group level since 2019, when it launched its project to build Japan's fourth mobile network. Besides reducing capex, Rakuten last year said it had slashed monthly network costs from about JPY39 billion ($250 million) in late 2022 to just JPY23 billion ($150 million) by the end of 2023. Its network operations team employs fewer than 260 people compared with "thousands," it has previously claimed, at each of its rivals.

While the mobile venture has underperformed early expectations, there have been signs of improvement in the last year. Revenues grew 30% year-over-year for the recent third quarter, to JPY73 billion ($470 million), and earnings losses (before interest, tax, depreciation and amortization) shrank from JPY34 billion ($220 million) to JPY10 billion ($65 million) over the same period. Sriwastawa expects to break even based on monthly earnings by the end of this year. Customer numbers, meanwhile, have risen from about 6 million at the end of 2023 to more than 8 million today.

"We want to grow to 10 million plus and then probably 12 million, 15 million," said Sriwastawa. "The word of mouth is growing in Japan, and we see lots of acquisitions coming from just reference." Anything that boosts customer numbers while helping Rakuten to avoid additional spending will look very welcome.