There has been endless rhetoric from most major telcos about becoming the best AI-powered network for AI workloads, but rhetoric isn’t reality. Palantir CEO Alex Karp pointed this out clearly when announcing his company’s new partnership with Lumen Technologies.
Why not troll Lumen too?
In his Fox Business interview with Lumen CEO Kate Johnson, Karp called out every elephant in the phone booth, saying:
You have a business that if it can be transformed in a year as opposed to five years could be the backbone of essentially every AI use case in the country. You have these somewhat older telco infrastructures that really can’t support the burden that AI places on the network.
In one statement, Karp trolled the entire telco industry (he even glanced apologetically toward Johnson when he said it). Telco network transformation has been too slow, and when it has happened, the results are usually insufficient to serve AI and benefit from its growth.
No telco wants to hear its infrastructure can’t handle AI, especially after taking on massive debts and burning through capital on 5G, fiber deployment, and digital transformation.
Yet here we are.
It takes a real straight shooter to dig into these issues because most telco executives will downplay Karp’s sometimes halting statements. But Geoff Hollingworth, Chief Marketing Officer for Rakuten Symphony, has guts and is one of my favorite sources. He was ready to oblige a spirited conversation, and I’ve shared key takeaways for your information and enjoyment.
Meeting hyperscalers’ needs directly
Lumen’s past financial woes were well publicized. With its back to a wall, the company shed debt, restructured and consolidated its business, and transformed its network aggressively.
Hollingworth explained this began with having a different mindset than most telcos. It built a different culture from the ground up by embracing modern approaches to problem solving and value creation that were focused on what their main target customers – hyperscalers – said they needed and required.
For example, Hollingworth says hyperscalers need fiber to connect data centers, but “Lumen was struggling because hyperscalers needed deliberate SLAs and Lumen could not deliver on them with their existing operation.”
One fundamental problem was an inability to fix outages fast, in part because the company could not identify where all its infrastructure and gear was located. Trucks would roll to root cause locations only to find the gear wasn’t there or wasn’t the root cause.
Lumen’s response to this problem, Hollingworth explains, was to replace its legacy inventory management systems with a modern one from Ciena’s Blue Planet. It also re-engineered its processes around this system to ensure it could meet hyperscalers’ specific SLA requirements and win their trust and their business.
Telcos are undisciplined with data
Lumen’s inventory management glow-up was one example of how a telco could become better at using data to make decisions and solve problems for customers.
Hollingworth says that “every successful modern company that has been born in the last 30 years has succeeded because they are built on software and have a great data discipline. They understand their data.”
AI is only as good as the data that feeds it, but most telcos have struggled to improve their data discipline. “Telecoms have to master being disciplined around the data they have and make it truly consistent and reliable,” Hollingworth says, especially if they want to profit from AI.
Without a real data strategy, Hollingworth explains, AI-based solutions end up being “bolted-on, limited, and restricted.” In contrast, a strong data strategy opens opportunities for “efficiency through automation, new experiences, and new service offerings” – particularly those that use software and data to create the experiences target customers want.
Transforming inventory systems to meet hyperscalers’ SLA requirements is a prime example of this. It requires having a “consistent and living inventory of all assets, both physical and virtual or software-based,” Hollingworth explains, adding that this is how Rakuten Mobile, the first software-based, cloud-native mobile operator approaches its operations.
With consistent data in hand, telcos “can choose to do many things with it” Hollingworth says, including employing specialized, small language models (SLMs) and context engineering to improve the quality of AI’s outputs and decisioning while lowering costs and keeping proprietary data private – more on that later.
Telcos can miss on AI like they missed on APIs
You might disagree with Hollingworth on telcos’ lack of software skills and data discipline, but history will prove you wrong. Telcos invested billions in new networks (remember LTE and 5G?) at the same time the API economy took off, yet they hardly benefited with revenue or profitability growth in the same timeframe.
This, Hollingworth explains, is because offerings like network-as-a-service (NaaS) and network APIs have been “designed like a standards body, not like a software company.” He points to companies that have been successful in the API business. They are “software companies first,” Hollingworth says. As a result, “they don’t think ‘I’m going to roll out 10 APIs and find customers for them.’ They think, ‘I want to do X and I need an API so I don’t have to involve a complex, third-party interaction to make business happen.”
Meta was once bad at APIs too
Facebook, now Meta Platforms, “built its business on APIs and software, and now you have a platform with data that is valuable to others who want to interact with it.”
But Facebook wasn’t always good at APIs. It took time to evolve them. “I was a customer of the first Facebook APIs that rolled out, but the APIs were awful and they changed every week in ways that were not backward compatible,” Hollingworth says. Facebook continued to iterate on its APIs and created the market-leading data platform they offer today. That has made Meta one of the magnificent seven.
Contrast the customer-solution, and software platform-led approach with most telcos’ approach to API exposure. “Trying to expose APIs on a network academically is bizarre to me,” Hollingworth says. “I don’t think it ever attaches because you don’t know who you are building the APIs for or what is valuable to them.”
Going back to the Lumen example, Hollingworth explains “they knew they needed to cater to the hyperscalers without paper and emails, so that was a driver to build their APIs out and have the data layer and networking underneath it.”
Why aren’t telcos filling the $1T compute backlog?
Most telcos are not software companies, but they are network infrastructure companies that serve hyperscalers with co-locations, points of presence, and inter-data center connectivity. There is no way telco leaders are unaware of the massive and growing demand for compute resources, so why aren’t more trying to fill it?
New deals seem to be announced every day between compute infrastructure providers like Nvidia and AMD and hyperscalers like Google or platform providers like Palantir. It has been widely reported that there is now at least a $1 trillion backlog in demand for compute resources fueling these deals. But telcos have rarely been in that deal mix.
Hollingworth lists off several fundamental reasons why most telcos simply can’t get after demand for GPU cluster-sized compute offerings. These bolster Karp’s claim that telco infrastructure — not just networks — are not suited to support AI’s demands.
> Insufficient power and cooling
“The power consumption of GPU density and cooling requirements are insane” and “telco facilities are not engineered that way at all,” Hollingworth explains. Because telco data centers and network sites weren’t designed for the power and cooling GPU clusters demand, it’s “a massive lift or impossible,” Hollingworth says, for that physical infrastructure to be re-purposed to serve up compute.
> Locations not chosen for AI
Not only were telco infrastructure sites not designed for AI power requirements, they also weren’t placed geographically to serve AI customer needs. “They’re not sure who will use them in those distributed locations,” Hollingworth says. As a result, it doesn’t make sense to burn capital putting GPUs in them, even if the power and cooling issues could be solved.
> Architecture not intended for AI
In addition, the networks that run through those sites are often centralized on a hub and spoke architecture. This also doesn’t meet hyperscale or AI needs, like robust inter-datacenter or customer edge connectivity. Telco broadband networks also tend to be engineered for asymmetry where downstream legs have more capacity than upstream by a multiple. AI workloads need far more symmetry in connectivity, making telco engineered networks too slow on the uptake in a literal sense.
Can the SLM wave provide telco infrastructure a reprieve?
Amidst the rising demand for compute and initial focus on massive, even trillion-parameter LLMs, developers have more recently favored SLMs coupled with agentic AI tools. SLMs trained for specialized purposes may have smaller parameter footprints, but they are extremely capable of performing specific tasks more accurately than LLMs. There are more than 2 million open-source AI models in the market today, and hundreds of thousands of them are smaller than 2 billion parameters. Many can run on desktop GPUs, new PC chipsets, and even mobile devices in some cases.
There are also new tools, Hollingworth cites DSPy, that are designed to optimize AI language models (LMs) to run more efficiently and improve the quality and accuracy of their outputs.
Specialized LMs can be trained for purpose and run locally or at the customer edge. This means far less need for power and cooling like the massive LLMs that serve millions of subscribers’ many different needs on-demand.
This begs the question – is the SLM wave a better fit for telco infrastructure?
Frameworks like DSPy, Hollingworth explains, make it easier for developers to optimize context engineering and improve prompt engineering. “This is important,” Hollingworth says, “because you take a much smaller language model and get a higher quality output than an LLM…and that opens up a totally different cost model.”
Practitioners like Hollingworth have found that “if you put LLMs into workflows at scale, it becomes very expensive very quickly, but if you can get the quality output on a smaller model, you can get a massive advantage from a cost point of view.”
As a result, telcos may have a “chance to go down that optimization path and unlock some real value.”
Is Nvidia’s $1B stake in Nokia really about 6G?
Speaking of GPU deals, Nvidia announced on October 28 that it had invested $1 billion in legacy telecom equipment giant Nokia to “pioneer the AI platform for 6G.” But something about this statement doesn’t quite make sense in the context of the rest of this AI discussion.
Hollingworth explained that Nvidia has been struggling to sell GPUs in telecom, for the reasons stated previously – “the industry just isn’t designed to run on GPUs.”
Plus, 6G still seems a long way off. “6G is not going to happen… if it happens (it) is probably 5 to 10 years out and it’ll take that length of time for people to find ways to do it, if at all,” Hollingworth says.
The 6G rhetoric may be necessary because there is a successive G economy, from researchers and equipment suppliers with telco operators at the end being served a meal of which 80% was not wanted or requested.
“Currently the 6G conversations are the same and as delusional as 5G,” Hollingworth says. For example, 6G talk often focuses on sensing, yet “telecom couldn’t even do location, how will they do sensing?”, he quips.
The more likely driver is that “the human-machine interface is moving from a smartphone screen to a contextualized conversation between machines or agents” which becomes the interface to apps. This, Hollingworth points out, is what OpenAI is driving with its apps in ChatGPT.
This shift is “massive” and at “day zero” Hollingworth says. He adds that it impacts every aspect of how humans interact with software. Even an OSS provider must ask itself what the operations dashboard should look like, and how people should interact with it.
Hollingworth predicts: “Apps changed everything and new leaders emerged. AI will do exactly the same.”
Can telcos be among those leaders? Not if their networks aren’t truly ready to monetize AI.
