According to a Light Reading article, one driverless robotaxi generates 20 gigabytes of traffic per day. I don’t know if this is true, but I think a collateral point the story makes is an important one. We’ve been conditioned by the consumer Internet to a traffic model dominated by sending to the edge, and there’s at least a risk that developments will reverse that trend. The claims the article cites, of course, link this to AI, but I think AI is only an indirect villain in this. Is there really a significant shift in traffic patterns, and if so would it expose telcos to new risks, open new revenue opportunities, or both?
Many consumer broadband connections, as most consumers know, offer higher download than upload capacity. Even when the interface is uniformly clocked (200 Mbps symmetrical, for example) the networks may be designed with the almost-always-valid assumption that traffic will flow out to the user much more than in.
One follow-up question is whether robotaxis or AI actually generate that kind of WAN data pattern. The way almost every AI user works with the tools doesn’t. I write a prompt that might be two lines long, and get a ten-page report in return. Many have argued that AI search results actually reduce traffic in the WAN by reducing the number of times a user clicks on a result link. And I personally have questions on whether robo-vehicles of any sort have a reason to push a bunch of inward-heading bits; to rely on network connections for public-safety-related behavior would be reckless, IMHO. Insurance issues? Maybe, which might justify continuous video monitoring, but that wouldn’t involve AI.
So is all this bunk? Maybe some of it is, but there are some points to be considered, mostly points relating to the role that tech in general and AI in particular might play in real-time activity. Robotaxis may not be current examples of this, but they are perhaps advance warnings of a network challenge.
Tech-supported real-time systems seem to depend on “world models” or “digital twins” to create a computer-accessible view of what’s going on in at least a part of the physical world. To generate this view and keep it aligned with real-world processes, you need to obtain some sort of sensor data. For that, your choice is perhaps a whole bunch of limited sensors, infrared, ultrasonic, switches, and other familiar things, or a form of real-time video analysis. For simple real-time systems of the sort used already in factories and warehouses, simple stuff would likely serve. For real-time missions that involved more autonomous elements (including people) a better strategy would be to analyze video.
Doesn’t this then validate the notion of robotaxis pushing gigabits? Not really, because a robotaxi is big, expensive, and so could likely justify on-board technology. The problem arises when you try to incorporate things that are numerous, small, and have to be relatively cheap. If everyone carried XR glasses linked to on-board AI (in their phones, likely), that might give us enough analysis power to handle human-inclusive real-time systems, but we aren’t there yet, which means that some real-time traffic carrying video for analysis is likely to come along. That, clearly, would have network impacts and potentially costs that would impact the ability of these applications to make a business case.
Real-time applications overall need three things to be true about their network services in order to work. First, they have to be reliable because their failure would break the real-time system and likely generate the same sort of real-world impacts the Light Reading piece described, but on a larger and more dangerous scale. Second, you’d need to have low enough latency to maintain the correspondence between your world model and the real world, to the level that the pace of action/reaction was accommodated. Third, you’d need the cost to be low enough not to compromise the business case or make widespread deployment of the services unprofitable. Would these be possible?
Truth be told, we don’t know. Enterprises who have commented on this to me so far are almost 100% convinced that few if any of our three requirements are met today, and by a rough 4:1 majority think that they’re not in the cards for the next three years. But by a 6:1 majority, they think that by 2030 this sort of service set will be offered in most developed markets. To me, that suggests that in the transition period beyond Year Three, there will be massive change.
For the record, I think enterprises are more hopeful than realistic in their views here. However, I do think some comments they’ve made offer some suggestions on how things could evolve.
Most optimistic enterprises think that some public-policy support, meaning governments at some level, would have to be involved. There’s often a specific suggestion that public safety personnel would drive progress. A body camera on police, fire, and rescue personnel is already used in some areas, and it’s reasonable to suggest that making some of the cameras generate real-time feeds could be possible. AI analysis of these feeds might generate enough immediate value to drive a network opportunity.
The real value here, though, seems likely to depend on some form of hierarchy of analysis. Some things need immediate, reliable, actioning, so you need some on-body analysis. If this were provided, it could reduce the burden on network connections that link the camera systems to deeper analysis points. Hierarchical AI of some sort is definitely a potential source of more bidirectional traffic flows.
Are there other means? One that’s been suggested is some sort of populist resource pool created by using a connection to share unused resources. There have been many suggestions, and a few implementations, that have done this with access point technology. Personally, I do not believe that sharing access point resources, or local compute resources, to make up an ad hoc pool is a viable mission for reasons of profit and security/governance. Enterprises, at least, would need guarantees of both security and governance, and since even cloud provider promises here aren’t enough to support business-critical applications, it’s unlikely that consumeristic pools would be accessible. Could someone set up a commercial sharing deal, where companies deployed custom server/storage contributions to export through the network? I think security and governance issues would make this difficult too.
In all, I think that there are opportunities in real-time “uplink” services, but they depend on the development of applications that would exploit them. The needs in this space are clear, the technologies are available, but the ecosystem needed is yet to be realized. Still, it may be the best hope for new telco revenue.
