A telecom operator runs two supply chains with one team, and stretching a single method across both damages each of them at the same time. Network rollout demand is generated by a construction programme whose pace is set by permits, landlords, grid connections and civil contractors, so it should be planned as a project schedule rather than forecast as a demand series. Device demand behaves like fast moving consumer goods on a subsidy and contract renewal calendar, with a commercial life measured in a few quarters. Returns and refurbished units are a second inflow nobody ordered, and ignoring them means buying new stock while serviceable units sit inside a repair loop.
The Monday planning review runs two agendas that never quite meet. One is a site list: how many are on air, how many are waiting on a permit, how many have a power connection date from the utility and how many are still guessing. The other is a handset and router list, with a promotion starting in three weeks and an allocation from the vendor that arrived smaller than the plan assumed. The same team owns both, usually with one forecasting method and one monthly cadence, which is how a regional warehouse ends up holding radios for sites that will not be built this year while the retail channel is short of the one memory variant everybody wanted.
These are two businesses. Demand on the network side is generated by a construction programme whose pace is set by councils, landlords, grid connection queues and civil contractors. Demand on the device side is generated by consumers, subsidy calendars and contract renewal cycles, on products with a useful commercial life of a few quarters. Telecom supply chain planning goes wrong most often at the point where one method is stretched across both, and it goes wrong in both directions at once.
Rollout demand is a project schedule, and should be planned as one
A site moves through stages: candidate, search ring, lease signed, planning permission, structural approval, civil works, power connection, equipment delivery, installation, integration, on air. Material demand attaches to a stage transition with a lag rather than to a month, since the antennas and radios are needed at civil complete minus a couple of weeks, and the cabinet and power kit earlier than that.
Statistical forecasting on the shipment history of remote radio units carries almost no information about next quarter, because the information lives in the site tracker rather than in the history. The demand object is a site, with a bill of materials by site type: greenfield macro, rooftop, street pole, in-building system, each with its own kit list and its own civil profile.
What is genuinely forecastable is the behaviour of the pipeline. Conversion rates between stages, cycle times between stages, and the distribution of those cycle times by region and by construction partner. Those numbers are stable enough to model and they are what turns a list of dated intentions into a material plan. Forecast the pipeline mechanics and drive the material off them, rather than forecasting the equipment.
What slippage does to equipment you already bought
Sites slip for reasons no planning system can see. A landlord renegotiates. A council refuses and the appeal takes eight months. The distribution network operator gives a connection date twelve months out and then moves it. Fibre backhaul is not where the design assumed. A structural survey says the tower needs strengthening before the new antennas can go up.
The equipment, meanwhile, was ordered against the original dates, often early because the vendor frame agreement carries volume commitments and price steps that reward buying ahead. It now sits in a regional warehouse as project stock, which is where several unpleasant things happen to it. Hardware revisions move on and stored units drift out of the current build standard. Warranty clocks that start at delivery burn while the box sits. A technology refresh redefines the target configuration and equipment bought for the old design becomes a swap-out candidate before it is ever installed. Capex is committed and sitting in construction in progress rather than earning.
Two planning habits help more than a better forecast. Stage the material commitment so that long lead items are triggered at permission granted while commodity items wait for civil start, which keeps the expensive exposure attached to the milestone that predicts installation best. And split the exception reporting into material waiting for site and site waiting for material, because they look similar on a shortage report and have opposite remedies.
The device side is fast-moving consumer goods with a subsidy calendar
Handsets, tablets, routers and set-top boxes behave nothing like network equipment. Volume is front-loaded into a launch window, the product has a commercial life measured in quarters, and the residual value of unsold stock falls on a schedule. Variant mix is its own forecasting problem, since colour and memory splits move with pricing and with what the vendor allocates, and mix error empties one bin while another ages.
Demand is also partly manufactured by the operator. A subsidy change, a tariff refresh, a trade-in promotion or a retention campaign creates a spike with a known date and an unknown size. Channel structure adds the familiar gap between what you ship into stores and dealers and what actually sells through, which has its own method covered elsewhere.
The cadence this needs is weekly, informed by sell-out data, with short-horizon methods and a fast reforecast after launch. The cadence the network side needs is monthly or quarterly, tied to capex governance and a construction programme. Running the device business on the network's calendar means the first reforecast after a launch lands after the launch window has closed.
Returns and refurbishment are a second inflow you did not order
Devices come back. Dead on arrival units, cancellations inside the statutory withdrawal window, warranty swaps, trade-ins against an upgrade, and customer premises equipment returned at the end of a contract. Each stream has a different rate, a different lag from the original shipment, and a different grading outcome.
The planning consequence is that supply has two inflows. New units arrive when you order them. Refurbished units arrive because of something that happened weeks earlier, frequently the same launch or promotion that created the demand you are trying to serve, and you cannot switch that flow off. The refurbishment line itself has finite capacity and a throughput time, so a trade-in surge can leave a month of value stranded in a grading queue.
Router and set-top fleets go further, since in most markets they are rented rather than sold. That makes the estate a rotable pool: units come back, get tested, cleaned, reflashed and reissued, and the number you need to buy depends on how many you recover and how fast you turn them. Operators that plan new purchases without modelling the recovery loop systematically overbuy. The forecasting method for return volumes is covered elsewhere; the point here is that the supply plan has to accept an inflow it does not control.
Why one team with one method damages both sides
The two failure directions are easy to recognise once named.
Apply statistical demand planning to rollout material and you get a smoothed ramp that no site plan ever matches. The plan orders steadily, the programme delivers in bursts, and the gap accumulates as project stock. Exception reports fill with items that are neither early nor late in any meaningful sense, because the baseline they are measured against was never real.
Apply project thinking to devices and the quarterly bucket hides the launch curve entirely. Allocation decisions get made in a governance forum that meets after the decision window, promotional lift is treated as a variance to explain rather than a signal to plan, and ageing stock is discovered at the quarterly review.
The structures worth separating are the demand signal, the planning cadence, the horizon, and the measures. On air sites, capex phasing and schedule adherence belong to one side. Fill rate, sell-through, ageing and gross margin on device sales belong to the other.
The structures worth sharing are the ones where they genuinely compete: warehouse space and receiving labour, field installation and logistics capacity, freight and customs, cash, and any vendor that supplies both. That shared layer is the reason a single planning forum still makes sense, provided the forum arbitrates between two plans rather than merging them into one.
Spares sit wherever the network is, which is everywhere
Field maintenance stock is the third thing in the building, and it obeys neither of the first two logics. Coverage has to follow a network that is geographically dispersed by design, with regional hubs, forward stocking locations, and van stock carried by engineers that in many operators is not counted at all.
Criticality comes from the network topology rather than from the part. The same line card is a different planning problem in a redundant access site than on a transport ring where its failure takes out a region, and the response time commitment differs accordingly. That means the planning system needs the topology as an input, and most implementations discover this late.
Two further wrinkles are specific to the industry. Repair and return loops with vendors mean a faulty unit goes back for repair and rejoins the pool forty or sixty days later, so what you are planning is a circulating population rather than a consumption stream. And decommissioned sites are a supply source, since equipment harvested from a swap-out programme can feed the spares pool for the generation still in service. The stocking method for service parts is covered elsewhere. What is worth insisting on here is that the network design tells you the criticality, and criticality drives everything after it.
Supporting a technology after you stop deploying it
Every operator is running a generation it no longer buys. Copper and legacy voice platforms, 2G and 3G radio access, older transport gear, set-top boxes for a platform that has been superseded. Regulatory obligations, wholesale contracts and a stubborn population of devices keep the service alive long after the last new deployment.
Vendors move first. A last time buy notice arrives, and sizing the response means projecting failure rates against a shrinking installed base over a shutdown timeline that is itself a moving political target. Buy too much and you scrap it; buy too little and you are on the third-party maintenance market at whatever it charges.
Decommissioning is a materials plan of its own, and it is usually nobody's job. A swap-out programme generates a reverse flow of removed equipment that has to be transported, tested, sorted into spares, resale and recycling, and disposed of under waste electrical regulations. It is simultaneously a demand plan for new equipment and a supply plan for used, and treating the two halves separately is how operators end up buying spares for a generation while skipping them into a container on the other side of the country.
Where this stops
Rollout demand is bounded by permitting and civil works timelines that no planning system can see and no forecasting method can shorten. Anyone selling a model that predicts council decisions is selling something else.
The honest treatment is to stop forecasting the dates and start measuring their stability. Take each site's required-on-site date across successive plan versions and compute how far it moves at four weeks, eight weeks and twelve weeks out. That distribution is the real input, and it is what tells you how late to commit material and how much of a buffer the programme needs.
The same measurement exposes the data limit. Site trackers are maintained by construction partners whose incentives favour optimism, so dates carry a bias that varies by partner and region. Measure the bias per partner, apply it to their dates, and share the number with them.
This week, pull the last two quarters of site milestone dates from the tracker, compare each site's plan date against the actual, and produce one slippage distribution per construction partner. It takes an afternoon, and it usually changes the ordering policy before it changes anything else.
More on this in industry on the DemandLabs platform.
