Most factory-floor ‘physical AI’ runs on lightweight (non-GPU) edge compute, reckons NTT Data – but 5G sensing, in the form of ISAC via SRS, could unlock a new wave of demand for private 5G, tying edge IoT and AI into a single enterprise stack.
In sum – what to know:
Edge AI mechanics – Heavy model-driven IT workloads need GPUs; most industrial-grade physical AI cases run on lightweight, task-specific models at the edge.
Breakout use case – SRS (sounding reference signal) in 5G enables network-based ISAC positioning for indoor robotics and digital twins – an “iPhone moment”, maybe.
Integration game – private 5G is part of a broader SI-led edge solution, reckons NTT Data – which is why Nokia is getting out, and it is just getting going, it says..
As promised, here is the rest of the discussion about IoT sensing and AI sense-making at the enterprise edge, plus the role of 5G in between, with Shahid Ahmed, global head of edge services at NTT Data – who RCR caught up with three weeks ago and wrote about two weeks ago in a brief corrective about the likely use of GPUs in hard-nosed Industry 4.0 scenarios in the name of physical AI, as well as about real momentum in the private 5G space. The point was really just to ground this trendy physical AI concept in real-world industrial sensor systems inside factories and plants – and make it distinct from the cloud-hype around general-purpose GPU-geared generative AI.
Just to pick up where we left off, briefly: does a big automotive maker, say, need GPUs in its factories for OT-based physical AI workloads? Does BMW, for example – which NTT Data is working with variously in some capacity in Germany, China, the US, and globally – want this? And does it ask NTT Data to supply them – with whatever else it is delivering? The BMW question is speculative, clearly; the references, here, are a couple of years old, and the German firm is characteristically tight-lipped about the nuts-and-bolts of its digital-change efforts. Ahmed responds: “No, it doesn’t need GPUs for that. And I can’t comment on the other question because I don’t know.”
But he also draws a line, here, between two very different kinds of AI in factory environments – and their different compute needs. Because BMW, or whoever, requires on-prem GPUs, of course, for heavy IT-style AI workloads – for large agentic and generative models, plus for integration with enterprise platforms (“like ServiceNow”), or other centralised analytics and orchestration layers. These are model-heavy, and less latency-critical, and will run in the cloud anyway, or else in a beefy edge server – versus physical AI with camera sensors and environmental monitors, set on narrow task-specific logic problems (fault detection, security checks, safety monitoring) with real-time results.
Which are all classic private 5G cases, of course. He says: “IT workloads will absolutely require that kind of resource. But edge AI use cases don’t need heavy local compute power – or even power, by the way. Like I said, we could run some of these on a Raspberry Pi, no problem. And you’ve got Nvidia’s Neutrino chip too – which is tiny, and works great. But it’s not a GPU.” On prompting (about proposed AI-grid and AI-RAN cases), he suggests the GPU-hype in the public-5G metro edge will tend towards network optimization, rather than physical AI in enterprise properties. “AI RAN is a whole different story. That requires some GPU level compute because 3GPP protocols are very inefficient.”
Ahmed explains: “All kinds of stuff is happening there, all kinds of handshakes. We talked about the uplink SRS (see below; the original conversation has been reordered) – I mean, that requires a huge handshake before you even connect – several times. So that kind of AI performance will require GPU at the edge. And so AI RAN, for sure; but anything else at the edge – again, math-based models – [does not need a GPU].” As we were then; actually, the first question to Ahmed is a rambling one, about all of this. Talk about how this fits together – IoT and 5G; physical AI and generative AI, at the enterprise edge, metro edge, cloud edge. I mean, that’s the whole AI infrastructure story, right?
Positional play
It’s a fair question, actually; because NTT Data, global system integrator, is dealing with all of this – top-10 global IT provider by revenue; number three for infrastructure implementation / management; a ‘full-stack’ portfolio across cloud/edge compute, networking, apps; $30 billion of annual revenue; $3.6 billion of annual R&D; 200,000 staff; 70-odd markets; 75 percent of Fortune Global 100 companies. “Yeah, it’s a long one,” responds Ahmed. “There’s no short answer to that.” But he sums up his own remit within the group – which is at the edge, at the “intersection of connectivity, mobility, and physical intelligence”, covering IoT, 5G, and AI, plus its Transatel IoT MVNO business.
“We’ve combined our edge 5G, IoT, compute, and AI, plus our MVNO, into one offering. There are sub-offerings, as well, but it is one entity, one architecture. So we have contracts with major auto makers, where cars roll off their lines with our [Transatel] eSIM.” Is this BMW, then – or someone else? ransatel has a deal with BMW, but the example goes beyond its scope, Ahmed won’t say, and the sense is there are others, anyway. “You turn it on/off, lock it, know where it is – and so on. Some have a really solid private/public case – for when a car comes off the line into a staging area, to be collected by a locomotive to go to a dealer, right? They’ve got, I don’t know, 1,500 cars in the lot, and the last thing is to flash them with firmware.”
He goes on: “It used to be that they’d use Wi-Fi, very patchy, or go manually to every car in the lot with a handheld device with a pseudo-USB cable. And it took too long, and was fraught with errors. With this, they just flash them all over-the-air in one shot. No one’s opening a door, no one’s plugging in; the process has gone from hours to minutes – 20 minutes in some cases. And that’s outside the factory. Inside the factory, we have private 5G, with the same eSIM, for two main use cases: one, just for access to the internet, which is your meat-and-potatoes – ‘just give me the connectivity’ – and two, more complicated, for where AGVs are running around, and need consistent coverage.”
All of which is well-understood, of course: connectivity for comms and safety; connectivity for mobility and automation. “People look for complicated use cases inside factories and warehouses, but it starts with connectivity,” Ahmed says. More interesting, he suggests a new use case – and dares even to call it an ‘iPhone moment” for the private 5G sector. “There’s a third use case, just beginning, with the uplink SRS signaling protocol – which is maybe the physical AI part.” SRS stands for ‘sounding reference signal’, and defines that an uplink reference signal can be transmitted by a device (UE) to a base station, and measured for time and angle.
It was introduced in (LTE) Release 8 for channel sounding, extended in Release 9 for positioning, and further developed in (5G NR) Release 15 for more flexible configuration, and also as a potential enabler for sensing in emerging ISAC frameworks. It requires a “huge handshake”, as referenced above – presumably not big enough for a GPU – to compute when to issue the UE signal, and how to coordinate it across cells, schedule resources, and collect measurements. It is “interesting” for indoor set-ups, particularly, where GPS is unavailable. “People talk about digital twins but you need good positioning to do that. It is a critical case – especially for robotics and kinematics,” says Ahmed.
Where camera-based systems “describe” location and cause, based on constrained line-of-sight and frame rates, and interpretive inference, the SRS calculation is network-coordinated, probabilistic, and error-bounded – and knows them. It shows position in live-time, with centimeter-level accuracy. Ahmed comments: “That’s where the private network comes in. It’s the iPhone moment. It will start to drive demand for private networks indoors.” It’s such a cliché that its use, here and now, makes it an even bigger statement, potentially – and one the private 5G market might like to hear. So what about the state of the private 5G market? This was discussed last time, but there’s more here.
Integration problem
There’s some perceived jeopardy in the market, suggests RCR. We’ve just got off the phone with AT&T, which said it’s in, but that it’s also hard – and not what everyone thought it was going to be a couple of years back. Nokia is quitting – again, because it’s just too hard. There are mixed messages in the market about Verizon – going gangbusters, or else getting cold feet. (It’s hearsay.) European telcos are on the case, but hardly saying very much. So what gives, Shahid? Why Nokia is quitting and you’re still in – if it is all too difficult? Is it because this is an integration business, as Nokia has said – and integration is NTT Data’s own meat-and-potatoes?
Ahmed responds: “I don’t know. Check with Nokia; it has its own challenges. But yes, for us, this is an SI challenge – completely. It’s a system integration playbook; 100 percent, end to end. Because even before we show up, we’ve got to have a consulting engagement just to define the TCO and ROI – so we can sell it to the local factory manager, who runs the airport and the factory; so we can say, ‘this is what it looks like in five years’. And that stuff takes time; and then we send our design team. I mean, it is not just like, ‘here’s the BOM, and you can use Google Maps to figure out where the radios go’. It doesn’t work that way; we have an army of people on this.”
Is there a rush of business away from Nokia? Are people asking you to manage their networks? “There are a lot of worried customers,” he says simply, and he won’t be drawn. “It will be interesting to see how it all plays out.” NTT Data has always worked with Nokia tactically – on certain opportunities in certain markets; rather than officially, as a preferred go-to-market partner. Those roles are reserved for Celona, historically, and Ericsson, more recently. Ahmed says of the timing with the latter: “It has worked on its pricing structure, and scaled down its solution. Enterprise IT managers don’t want to configure millions of radio parameters; they just want a dozen of them.”
So Ericsson has right-sized its private 5G solution for enterprises, then. Plus, it is “still an Ericsson system”, he says. “It’s robust; it’s the best radio in the world.” NTT Data is using the Swedish vendor for bigger environments, he says; most of the rest go with Celona. And the mood in the camp – because discussion about the hard yards of system integration make it sound like a tough gig, still? “Three things,” he responds – and he summarizes the discussion in reverse order. “One: this is beginning to see scale, like we expected two years ago. Yes, we missed the mark, but we’re turning the corner. Two: private 5G enables geo-positioning indoors; that’s coming, and it’s real physical AI.
“Three, public networks and private networks, edge AI and physical AI – these things are all connected as one. No enterprise is going to buy physical AI, like they don’t buy IoT or private 5G. It’s about the solution, and a physical AI case has to be connected in the enterprise. So you have to have a single offering – a simpler way to go to market. That is the name of the game.”
