How consumer networks could ease the AI workload burden
The AI story is, in part, now one of constraints. Platforms, power and processors are all in greater demand than there is supply, and networks that are designed for AI workload characteristics are in short supply too.
A network designed for AI infers a range of advanced capabilities that should be familiar to TM Forum members by now: Level 4 autonomous networking, intent-based controls, real-time granular charging and metering, and dynamic self-optimization of network resources, to name a few. As any telecoms veteran knows, however, “the network” is many things, and not many networks do all the things AI workloads need them to do.
As a result, a common question emerges: in which networks should a telco invest in order to capitalize on transporting AI workloads? One answer may be local, and perhaps even consumer, networks.
AI developer and blockchain patent-holder Steve Ruben argues that because “there are highly capable models you can run on new AI PCs with multiple coding and human languages”, it opens an untapped opportunity.
Ruben already sees developers asking whether they “need to engage a model that knows everything, or that knows how to do just what I need to do”. He envisions a solution that would allow developers like himself to “distill a model that meets your specific needs”.
Ruben adds that emerging AI tools and economics, like those with which Deepseek surprised the AI market, makes AI architectures that use specialized local models more realistic and reliable. “With my $3,500 AI server, I’m not paying for my cloud model and, even more important, my model is always running and no one is messing with it,” Ruben says.
To support the growth of this approach to AI architectures, Ruben suggests, telcos could bring in an “AI infrastructure solution that borrows on the connected system” so that “when the customer is offline, their local resources can be applied to the collective”.
Some communications service providers (CSPs) may be exploring similar approaches to determine whether local networks serving consumers can be transitioned to provide AI capacity dynamically.
“I am seeing customers investigate all their existing compute assets enabling pervasive distributed AI,” says BT Business CTO Colin Bannon. “There is massive latent compute idle resource untapped. I call that ‘dark compute’. Imagine if you could share a model across multiple idle PCs in a network. Small bespoke models could be run locally by sweating existing assets but the network becomes the computer. The last time this happened was SETI, and 75 million PCs with screen savers installed would download tokens from Berkeley,” he explains, referring to research conducted by the Berkeley SETI Research Center, which used Internet-connected computers in volunteers’ homes to help process telescope data to search for extraterrestrial life.
Though SETI (short for Search for Extraterrestrial Intelligence) ceased collecting data in this manner in 2020, it demonstrated that it’s possible to create a long-term, low-cost cloud computing solution that flexibly utilizes local, and even personal, resources to process computing workloads.
A primary hurdle that those trying to achieve this vision are likely to encounter comes in shifting from a download-heavy network model to a symmetrical one. “The entire industry is download-driven. The model was always you upload a search query and get a video. But now, you upload a video and get an explanation – it reverses all the flows,” says Guy Lupo, TM Forum’s EVP of AI & Data Innovation.
But the concept of turning local networks into cloud computers for AI remains compelling because the trend in AI usage is moving beyond simple text queries to ChatGPT, says BT’s Bannon. “When you go agentic or voice or multimode, latency becomes important,” he explains, which makes having more dynamic capacity closer to customers a looming requirement for AI connectivity.
Ruben echoes Bannon’s point that AI usage continues to become more sophisticated. He explains that AI is already helping developers like him and his team to triple their coding output thanks to automation that learns how they like to code and auto-generates quality code to match. He adds that this capability becomes even more potent with agentic AI because of its potential to change how systems integration is done.
“I tell the AI, ‘Here are my models, here’s the API I want to engage with, now make me a client for my business domain so I have a fancy wrapper to use to call this API’ and it has done all the mapping. It does all the [gritty] data transformation and all the cookie cutter code you fill in. So as a developer, it frees you up to focus on the creative development and not on these integration and mapping operations,” Ruben says.
The integration advantages agentic AI may offer can also create opportunities for CSPs, explains Dave Milham, Chief Architect for TM Forum. Though hyperscalers and AI platform providers “don’t want to build the network”, says Milham, they do want networks to evolve to support their AI traffic.
Making this a reality means solving disparities across vendor domains, such as when intent-based systems aren’t interoperable out of the box. “There’s an opportunity to integrate all the different vendor (systems) together, especially as they boil their capabilities up to just these intent interfaces,” says Milham. The CSP value proposition, he says, can be based on integrating these intents to add value on top of vendors’ technology offerings.
Milham says integrating intents isn’t trivial but is becoming easier because of agentic AI. “Agentic AI software agents will prove critical in making all the inter-domain execution and negotiation work,” which is necessary to deliver the level 4 autonomous network capabilities AI workloads require, Milham explains.
Milham says that while software agents have been in use since the late 1990s, “AI has made agents much more feasible, practical and cost effective.” He explains that some industry standards are being re-aligned for AI and that TM Forum is helping organizations like IETF with this effort. For example, the TM Forum ODA Production team, which Milham leads, is working to define controllers that connect TM Forum APIs with different vendors’ intent-based interfaces across technology and supplier domains.
One result of simplifying and automating these key aspects of software development and integration is that it will spawn more workloads, many autonomously, to carry out all the automation tasks assigned to AIs by people and other AIs. It is from this projected escalation in AI workloads that many telcos hope to capture new sources of much needed connectivity revenue growth.
We will take a closer look at the revenue opportunities related to connectivity for AI workloads in our upcoming research report: Monetizing connectivity for AI workloads.