AI Bandwidth Commitments Are the New Migration Circuits
By Bill Henrichs
Founder & President, Bearstone LLC | Former Head of Telecommunications, Simon Property Group
When I ran telecommunications at Simon Property Group, we provisioned a great deal of bandwidth we never fully used. Migration circuits were sized for peak load and never rightsized after the cutover. SD-WAN tiers were committed at deployment and never validated against what the sites actually pulled. The carrier billed the commitment every month, and the variance between what we committed and what we consumed sat quietly inside the line items — because no one on our side had been assigned to reconcile it. That was telecom a decade ago. I am now watching enterprise IT set up to repeat it with AI.
Networking teams are provisioning bandwidth right now — carrier circuits, cloud interconnects, data center cross-connects — for AI workloads that do not yet exist at production scale. The provisioning is happening before the workloads are deployed. The contracts are being signed before the real usage patterns are known. And in most enterprises, no one has been designated to reconcile the committed bandwidth against actual AI traffic once the deployment goes live.
The forecasts driving these decisions are real. Goldman Sachs projects global data center power demand will rise 50 percent by 2027 — and as much as 165 percent by 2030 against 2023 levels — driven by the buildout of AI infrastructure. Dell’Oro projects rising demand for near-edge connectivity and high-speed private interconnects to carry those workloads. What the forecasts do not capture is the accountability gap behind them. The category is new. The governance gap is not.
The pattern enterprises have already lived through
Every prior generation of network architecture produced the same gap. MPLS migration circuits, over-provisioned and never rightsized. SD-WAN bandwidth tiers, committed at deployment and never validated. Cloud interconnect commitments, sized for projected growth and never reconciled when the projection missed. In each case the carrier had no incentive to surface the variance — the committed bandwidth was billed monthly, and in the absence of a designated owner inside the enterprise, the gap compounded quarter after quarter, renewal after renewal, until someone finally pulled the contracts and the invoices apart and read them line by line.
Three complications unique to AI
AI is now entering that pattern with three additional complications. First, the committed bandwidth almost always exceeds actual production usage — AI workloads are over-provisioned at deployment because the cost of an under-provisioned circuit is a model timeout, while the cost of an over-provisioned one stays invisible until someone validates the bill. Second, the traffic profile rarely matches the contract. Inference workloads are bursty; they look nothing like the sustained throughput most carrier circuits are priced against, so enterprises end up paying for a sustained-throughput commitment to support a workload whose real exposure is latency, packet loss, and the occasional burst. Third, the interconnect and cross-connect fees compound alongside the circuit. An AI workload typically traverses a carrier circuit, a cloud interconnect, and a data center cross-connect — three separate billing relationships, each with its own commitment terms and its own variance pattern — and most enterprise teams own the reconciliation on only one of the three.
The accountability question
Inside most enterprises, the accountability question for AI bandwidth has not been answered. Networking owns provisioning. Procurement owns the contract. Finance owns the budget. None of them owns the monthly reconciliation against the contracted commitment. In the absence of that ownership, the carrier bills for the commitment and the enterprise pays it — not because anyone agreed to overpay, but because no one was assigned to validate the difference between what the contract permits and what the traffic actually requires.
The proof, from prior generations
I have watched this exact gap produce the largest exposures in enterprise telecom — not in AI yet, but in every prior generation of bandwidth commitment that followed the same pattern. In a 31-month governance engagement with a multi-brand retail portfolio — hundreds of locations, two carriers — validating every invoice against contract terms identified $2,122,436 in unauthorized charges. Of that, $1,852,314 was credited back to the client: an 87 percent collection rate. The portfolio’s approved telecom budget fell 16.9 percent from FY2024 to FY2025 — not because anyone cut services, but because charges that should never have been there came off the invoice. None of it came from renegotiating. It came from someone validating the commitment against the actual requirement, every month. AI bandwidth is the next category where that gap will open.
The governance principle
The principle does not change with the workload. Validation has to be assigned; it does not happen by default. Contracted commitments require monthly reconciliation against actual usage, line by line, against the contract terms. Variance gets disputed when it appears, not after it compounds for thirty-six months. The enterprises that get this right designate the owner before the deployment goes live — not after the first invoice arrives that nobody recognizes.
AI bandwidth commitments are about to become the largest unmanaged contracted spend category in enterprise telecom. The provisioning is happening now. The reconciliation is not. The exposure compounds in the gap between the two.
