AT&T has clarified its emerging AI “grid” and IoT strategy, combining regional inference, cloud platforms, and private 5G to target enterprise use cases while testing where edge AI delivers most value.
In sum – what to know:
Regional edge – AT&T is using Nvidia GPUs at six regional Cisco data centers in the US for inference workloads, initially attached to security cameras; the service will expand internationally.
Cloud platform – AT&T is offering a Microsoft Azure-based platform for manufacturers to share and analyze machine data across sites, using generative AI for diagnostics and optimization.
Private network – AT&T has twin private 5G offerings, big and small, with unnamed vendor partners, and contracts with municipalities and car makers; the tech is strategically “important”, it says.
Helpfully, AT&T has put its recent ‘AI grid’ and AI factory announcements, from GTC and MWC in March, into some kind of context for us, and also spoken quite clearly (not exactly freely, but candidly, and for the first time) about its business with private networks. Where to start? Let’s take them in (reverse-chronological) order, and summarize: AT&T is using Nvidia GPUs at six regional Cisco data centers in the US for inference workloads, initially attached to security cameras, as part of a managed IoT offer for enterprises (GTC, March 19); it has a new Microsoft-based edge-to-cloud IoT platform for multinational manufacturers to share machine data between plants (MWC, March 5); and it is busy selling two private 5G systems, one big and one small, to US-based enterprises of all sizes.
These projects might be considered among a rush of related releases from the firm, geared variously to put AI to work in the network and on the network: a new preview of AWS Interconnect with AWS, embedding fiber and 5G FWA directly into AWS workflows to support latency-sensitive AI; a recent demo with Ericsson of ‘AI-native’ link adaptation on an Intel-based cloud RAN stack, claiming efficiency gains of up to 20 percent over legacy systems; a broad fiber target to reach over 40 million US customer locations by the end of 2026, boosted by its Gigapower FFTH joint-venture with Blackrock and acquisition of Lumen’s consumer business. There has been lots more from AT&T lately, as well, but its discussion with RCR is about IoT and AI, mostly, as sold by its AT&T Business function.
Cameron Coursey, vice president of connected solutions at AT&T, is on the end of the line, and says more besides. “We are working at providing AI inference at the switch locations for 4G, and eventually 5G – in the data centers. This is not about putting [AI] in cell sites – which some other operators are looking at. I’m also not looking at AI for the RAN or AI in the network itself. This is for customers in the enterprise segment in IoT and how we can provide a capability there,” he explains. There is a suggestion here, maybe, that AT&T doubts the placement of AI accelerators in radio (RAN) and core network assets inside the metro network, but Coursey’s remarks are better considered just as a way to frame the conversation, and the remit of its new IoT projects with Cisco and Microsoft.
AI grid concept
The work with Cisco, in particular, is trendy – insofar as it occupies the telco-end of this new ‘AI grid’ concept, effectively popularized by Nvidia at GTC, but a portfolio item in Cisco’s enterprise proposition.
“Think about it as Cisco putting Nvidia GPUs inside its data centers, and AT&T using them for customers,” says Coursey. The set-up is offered as an extension of AT&T’s IoT network proposition, which uses a “dedicated” converged 4G/5G core network, separate from its “big smartphone network” (albeit using the same public RAN and carrier backhaul) – which is hosted as managed service by Cisco in the same Nvidia-linked locations.
“It allows us to move quickly; it needs to stay very constant. We do customization and send traffic to enterprises instead of to the internet.”
He explains: “We have six of these (Cisco) data centers in the US, which we used in a dynamic approach – so the (IoT) traffic gets routed to the closest one. We could have more, but that’s been adequate up to now for the IoT cases we have… Eventually Cisco will put [the same GPU facility] in some of the data centers we use with it in other parts of the world as well, outside the US.” To be clear, because RCR wants to know, the work with Microsoft, plus Nvidia (again) and edge-AI/ML specialist MicroAI, is separate, but also might conceivably be connected as AI cases multiply and morph at the enterprise edge. “We’ll consider stitching them together as needs arise,” says Coursey. Really, it is less interesting, or just less vogueish, than the grid stuff – an Azure-based analytics solution for industrial IoT assets.
But, then, the product, called Connected AI for Manufacturing, was flagged at MWC and presented at GTC, like the beating heart chambers of AI hype machine, and it has some smart additions: notably, the firm’s generative AI product, Ask AT&T, billed as an adapted version of Chat GPT, to provide a language interface for IT/OT staff to interrogate live machine maintenance and performance data from matching models at sister-sites across the world. Coursey says: “Data can be shared in Azure across different factories, and factory operators can ask questions to get to the bottom of what’s happening with their machines, and how to optimize them. Information you’re gathering in one factory can help to know what’s going to happen in another.” Hence the idea of ‘connected AI for manufacturing’.
Edge AI platform
There is some accelerated computing from Nvidia in there somewhere, of course; plus “edge AI” capabilities from Microsoft. It might not be connected, or integrated, with the grid proposition, as yet, but it is really separated only by the use-case profile – predictive machine maintenance at the edge (or LLM-enabled diagnostic maintenance in the cloud) versus video inference, where anomalies are flagged in near real time in regional data centers rather than directly in the production loop. You get the point – same customers, often, same edge IoT cases, mostly. Video is the kick-off case, says Coursey. “Even a couple years ago, we could see that video was the next bastion – taking the place of a lot of sensors; video security was coming up. It was the next frontier, and it is the first use case for this.”
AT&T already had a “standalone” machine vision product, Video Intelligence – which was “not AI-based at the edge”, he says; the shift has been to embed inference into its core network in Cisco’s data centers, and also to compress the video at the enterprise edge – ahead of its transmission to the regional network edge. “The beauty of it is we’re working with a partner with really good video compression technology. It doesn’t require a lot of processing on the camera itself… We’re combining the two – the compression tech and the inference engine – to provide the most efficient mechanism possible.” Interesting; it opens up a discussion, for another day, about how data compression tech might just slow this perceived edge-wards migration of AI workloads, including further into the telco network.
It is also working with Linker Vision (LinkerVision), introduced by Nvidia, which has a machine vision platform for transforming raw RTSP data from CCTV in multi-camera streams into actionable insights for industrial AI and smart city applications; it has reference cases with a number of smart-city projects, notably in Taiwan. The pair ran a six-month pilot at AT&T’s mixed-use Discovery District area in downtown Dallas. “We took the video feeds off the cameras, routed them to the Cisco data center, and ran the inference on the Nvidia GPUs. LinkerVision tweaked its algorithms to identify the things most important to the security staff. We’re expanding now with public sector customers – cities, municipalities, departments of transportation – and also with enterprise customers.”
Machine vision cases
For the latter, he suggests firms that are storing and renting high-value equipment – to “make sure nobody runs off with it”. He comments: “So the first use case is video. It fits the bill for the kind of low latency you get, and the security you can offer by peeling [the AI] off at a data center location, and not going back to a cloud service provider.” Coursey suggests certain other AI workloads, which might justify a round-trip to a Cisco data center somewhere between the city and sticks. “We’re looking at audio chatbots in vehicles, say, or AI DJs at the edge of the network – with which you can have a conversation with in a way that it feels like a real person, versus if you do it another way. We’re also looking at how to expand into the autonomous vehicle domain. Video is only the start.”
AI DJs? Hmm. We’ll park that one for now. But latency is a contested terrain, of course. Coursey is careful not to overstate the value of Cisco setup as a catch-all for AI-geared IoT – however that is defined. The point is not to replace the cloud, but to complement it, he says. “Many video use-cases can be served out of cloud service provider locations, and we’re not trying to compete.” The advantage, rather, is reliability – where reliability matters more. He goes on: “We know this provides more consistent video latency and jitter… where you don’t have the same quality of service [if you are] going all the way back to a location outside the network.” So it sounds, really, like a speculative exercise, to an extent – where the game is afoot, and you want to optimize infrastructure for whatever goes on top?
The question is for Coursey; it sounds like you have your existing IoT business, ticking over, but this is a new piece of work, which may go any number of ways. “That’s exactly right,” he responds. “We’re working with Cisco and Nvidia to experiment around where we want to go in the future – on this managed network for IoT, without having to go huge on the big AT&T network. This is a proving ground, which will expand into other parts of AT&T’s network as it makes sense… We have a funnel, based on the announcement at GTC, and we’re working that funnel. Certain customers are very interested, and others are just trying to learn. The plan is to [engage] Nvidia’s developer community, as well – to have a sandbox with Cisco to try different things to drive new use cases.”
He adds: “We’re in the very early days, so people are still doing a lot of learning.”
RCR will cover the private networks piece with AT&T in a separate post tomorrow.
