On the Virtues of RAN Virtualization with COTS Hardware [or Lack Thereof]

People who are really serious about software should make their own hardware.

Alan Kay said this in 1982 and the quote is still used to capture the essence of Apple's success 35 years later.

Since virtualization has a whole lot to do with software, is the mobile network virtualization community really serious about it when it comes to the radio access network (RAN), and if they are, is there any particular reason why it should be exempt from this assertion about the need for specialized hardware that is attuned to the exact needs of the software it runs? I don't see one. Also, I don't think virtualization is the panacea that will have mobile network operators (MNOs) rising like a phoenix from the ashes of spiraling ARPUs in the bottomless pit of consumption that is the unlimited data plan. Virtualization helps to scale a big portion of a mobile network but it remains to be seen if the cost base is suitable for an 80% CAGR in mobile data traffic with reducing ARPU. Unlimited data has quantity and ubiquitous access intrinsically prioritized over quality - a problem that is better addressed by mass-manufactured simplicity than over-engineered virtualized complexity, at least for user traffic in the RAN, if not the control plane as well.

Operators who virtualize their RAN may achieve short term reductions in CAPEX and OPEX, but only from the traditional telecom vendor's share of the pie and the reductions would materialize more because vendors are aware of the table stakes and less due to any fundamental reduction in the cost base. However, there is also the rest of the pie - this not so small matter of new CAPEX for servers, high-capacity and low-latency routers, and software for enterprise-grade operating systems, cloud execution and orchestration, and new OPEX for different software subscriptions, more rack space, power, cooling, cabling, etc. It might be the accountants who get paid the most for delivering a positive narrative at the end of all this.

When it comes to commercial-off-the-shelf (COTS) hardware, the writing is already on the wall and the graffiti is by courtesy of the machine learning community, not David Choe this time round. Machine learning forerunners like Google, Facebook and Baidu's Silicon Valley AI Lab have their own cloud TPU, Big Basin, and HPC cluster with GPUs respectively to extract the highest performance for a particular application. Had these companies limited themselves to COTS products, they would've fallen behind competitors in the race to deliver tangible improvements to customer experience sooner. Even if all MNOs that believe in salvation-by-COTS virtualize their RAN, they will eventually find themselves in a similar predicament as to the one they're in today - one of not being able to differentiate oneself from the competition.

Virtualization forerunners could embrace it all by designing their own network hardware for delivering differentiated services at jaw-dropping efficiencies due to a carefully architected synergy of hardware and software. This, of course, cannot be achieved in a single attempt and is more of an iterative process for the identification and elimination of bottlenecks in an architecture purpose-built for a mobile networking application. Many will be quick to point out there isn't enough scale for a MNO to do this. While this may be true for the typical MNO today, if Amazon were to jump into the mobile networking fray tomorrow, the same naysayers would agree in unison that it would make perfect sense for the company to go global by making custom networking platforms which scale efficiently. So why are the MNOs waiting to be disrupted? If MNOs can't achieve the necessary scale within their regional networks then perhaps they should work with global vendors who can? Perhaps the pursuit of the right synergy of hardware and software, and at prices the industry can sustain despite the traffic growth and ARPU loss would be time better spent than the pursuit of virtualization and COTS hardware just for the sake of it?

Hardware aside, with the operators' ultimate goal supposedly being to further fragment the virtualized protocol stack into smaller blocks to attract more competitors in the software space, would the fragmented software business even be a viable one? If there is a stable software company today with the financial wherewithal to partner with tier-1 MNOs that sells one layer of a protocol stack, I'd like to know more about it.

It almost seems more plausible that MNOs are using virtualization as an imperative to get vendors to perhaps compromise their own futures by recoding functions into standardized building blocks and with standardized interfaces between them so it becomes easier and less risky for the operator to progressively replace the vendors' blocks with their own blocks at a later date. If true, such a strategy could prove to be too clever by half.

Once the MNO becomes deeply familiar with the inner workings of each functional element, it might come to the realization that many of the functions could be accumulated into one process and the removal of unnecessary internal interfaces could be beneficial for efficiency. Furthermore, they may realize that if some functions were to be etched into an ASIC, efficiency would be higher than when running on a COTS processor - if only they had the scale necessary to make building an ASIC feasible. Since it would be unlikely that there would be another vendor out there with that exact ASIC available commercially, they might then find themselves purchasing a proprietary system with a higher level of system integration but with standard interfaces as well which happens to include a similar ASIC that performs efficiently, as expected. They would've then found themselves at square one, where they last bid farewell to the vendor with proprietary hardware. Who knew square one could seem innovative and high-tech if approached via the scenic route?