GPU-as-a-Service offered by mobile operators (Analyst Angle)

Mobile operators are investing in GPUs! SKT, DOCOMO, Verizon, T-Mobile, Singtel, Swisscom, Telenor, Softbank, and many others have set up data centers with edge AI services. For a market analyst like me, this gets my attention. On the surface, it seems that the NVIDIA strategy of GPUs in the RAN may be getting some traction. Is this just “walled gardens” all over again?
Let’s dig a little deeper.
First of all, the opportunity for AI “Inference as a service” in the RAN doesn’t really exist. After interviewing many different companies, I was able to break down the revenue for “GPU-as-a-service” by latency requirements, ranging from milliseconds up to 10 seconds or more. There is significant revenue out there, but very little needs network latency below 100 ms. (We have been able to identify some very interesting Augmented Reality, automotive, and remote machinery applications that could drive AI revenue in the sub-100 ms range. My personal opinion is that these will mature in the 2030-2035 timeframe.)
Secondly, these operators are not investing in GPUs distributed across the country. They’re investing in centralized AI data centers for each country. This is a tip-off that the goal is not to achieve low latency, but to achieve “sovereign AI services”. Swisscom, Telenor, and Singtel are great examples of operators that are likely to succeed by providing customers with a nationally sovereign service.
Third, I looked into the economics of placing GPUs throughout the RAN. Calculating the ROI consists of tabulating the costs of Grace Blackwell servers, upgrading the fiber, and overhauling central office sites with power and big air conditioners or even liquid cooling. On the other side, there are energy savings that come from centralized RAN pooling, plus the new revenue from localized services.
The detailed results are available in our published report. My overall conclusion is nuanced: Yes, this looks like the failure of “walled gardens” and “Edge Computing” all over again. But there is a place for national GPU-as-a-service offerings, and mobile operators have a natural way of providing that service precisely within the borders of each country.