Telcos are being asked to fund the next wave of AI and 5G innovation while already burdened by enormous debt and relentless infrastructure costs, and Vish Nandlall’s argument is that the constraint is “basic economics,” not a lack of vision.
He lays out a simple test for whether new network spending can clear the bar once interest and capex are taken into account, then applies it to “bright shiny toys” including edge AI—asking what those projects would need to earn in operating income, not just revenue.
The lemonade-stand analogy for telecom operating economics
Nandlall frames telecom finance as a borrowed-stand scenario: the stand costs $1,000 and borrows $700 at 6% interest, which results in $42 of annual interest owed regardless of performance.
After paying lemons and sugar, the stand throws off $100 of operating profit, but the owner is left with $3 once the $42 interest and $55 upkeep (for small upgrades like replacing a blender and keeping the stand competitive) are accounted for.
In the analogy, when a better stand opens across the street, upkeep rises from $55 to $60 and the owner’s $3 becomes negative—an illustration of how competition can consume what looks like available profit.
2025 operator numbers: interest and capex consume operating income
Nandlall says he uses operating income as the reference point to evaluate carrier finances, and compares interest expense and capex against that figure to gauge each cost intensity.
For T-Mobile (2025), Nandlall cites operating income of $18.3 billion and shows interest of $3.8 billion (equal to $21 per $100 of operating income) and capex of $10.0 billion (equal to $55 per $100 of operating income).
For Verizon (2025), he cites operating income of $29.5 billion with interest of $6.6 billion (equal to $22 per $100 of operating income) and capex of $17.0 billion (equal to $58 per $100 of operating income).
For AT&T (2025), he cites operating income of $24.2 billion with interest of $6.8 billion (equal to $28 per $100 of operating income) and capex of $20.8 billion (equal to $86 per $100 of operating income). He then argues, based on this row of figures, that the network business by itself spends roughly as much rebuilding itself as it earns.
Edge AI break-even: locality has to pay back utilization penalties
Nandlall positions edge AI as another spending pitch that has to pass a hard test: he states a principle that when interest plus capex eats most of operating income, every new investment must clear a high bar.
In a hypothetical AT&T example where a network team says it needs $1 billion for edge AI, he assumes new debt at approximately 6% costs $60 million per year in interest “forever,” assumes GPU-heavy infrastructure depreciates over about 4 years at $250 million per year, and estimates power, cooling, maintenance, and software operations at about $100 million per year.
On that basis, he calculates the edge AI project would need about $410 million per year in new operating income to break even, not revenue. Using AT&T’s stated 19% operating margin, he estimates this would equate to roughly $2.2 billion of new annual revenue, and says he does not see any edge AI use case sold through a telecom go-to-market motion plausibly generating that level of new revenue in the next three years.
Why 5G didn’t create the “smartphone wave” for AI
Nandlall argues that the standard pitch is that 5G is like LTE and that apps will arrive if people wait, but he claims LTE made the smartphone into the internet by changing what people do with their phones, while 5G added capacity and enabled fixed wireless access solutions without changing phone behavior.
He says promised 5G applications such as AR/VR, network slicing for industry, and ultra-low-latency consumer cases either did not materialize at scale or did not need 5G specifically, and therefore frames 5G as an asset-utilization story rather than a growth story.
He then extends the same logic to edge compute: for edge compute to make sense at a carrier, locality must create more economic value than the cost of distributing infrastructure close to users, and he says the edge AI use cases where that holds are real but narrow—ultra-low-latency industrial control, certain regulated data-residency workloads, and some real-time vision applications—described as vertical, contract-driven, and slow to scale.
Sources