The strongest case against pricing the network
Here is the objection at full strength, because it deserves to be heard before it is answered. Expected value of information is a computation, not a metaphor. It requires a decision model: a defined action set, a payoff function, a likelihood linking signal to state. A network planner deciding whether to commission a new fibre ring between two exchanges can, in principle, specify all three. But "continuous ingestion of traffic telemetry, fault alarms, spectrum filings and churn signals" is not one decision. It is an open commitment to observe everything a network might emit, indefinitely, against decisions that have not yet been named. Call that "priced information" and you have smuggled academic rigour into what is, honestly, just hoarding. A telemetry lake with no decision attached to most of what it holds is not an information asset. It is a cost centre with a spreadsheet justification bolted on afterwards.
This lands because it is often true. Operators keep years of packet counters, alarm logs and handset diagnostics that no capacity model has ever touched. Calling that "held as revisable belief with provenance" flatters what is, in a great many cases, unexamined accumulation.
Where the objection is right
Concede it cleanly: the expected value of information is only ever computed on a slice. A planner does not price "all knowledge of the network." She prices a specific question — should I add a third carrier to this cell cluster before the next tender window, given the last six weeks of PRB utilisation and the churn signal from the retail arm. That question has a payoff (avoided congestion penalties, deferred capex, retained subscriber revenue) and a likelihood (how well utilisation trend predicts breach probability). Everything outside that slice, however voluminously logged, is unpriced by definition. There is no honest sense in which a five-year alarm archive is "valued" until someone names the decision it would change.
The discipline this forces is real. If a network planner cannot name the drill/no-drill decision that a spectrum filing from a rival operator would alter, the filing has not been priced — it has been filed. The correct response to "we should ingest everything" is "which decision does this move, and by how much." Where no answer exists, the honest label for the ingestion is option value, or inertia, not information value. Large Universe Models, as a category, only earn the name where provenance lets someone compute that likelihood later. A stream with no schema, no timestamp discipline and no known error rate is not a belief. It is unlabelled inventory.
Where the objection runs out
But the objection proves too much if pushed further, because it would also rule out the seismic survey, the interim trial look, and the day-ahead forecast — all cases where the decision model is specified narrowly and the pricing holds exactly. The telecommunications analogue is the capacity plan. A regional network is dimensioned against a forecast traffic mix: video share, signalling overhead per session, average concurrent bearers per cell. The failure mode is well known to every planner who has lived through one — a messaging app ships a new default (auto-play video, background sync, a codec change) and the traffic mix that the capacity plan assumed six months ago no longer exists. The plan was correct against the state of the world it was built on. The world moved.
The value of information here is computable, and it is not abstract. The signal is early telemetry showing a shift in per-session bearer duration or upstream/downstream ratio in the two weeks after a major app release. The decision it might change is whether to bring forward a backhaul upgrade scheduled for next quarter or hold it. If the telemetry cannot move that decision — because the upgrade is already committed regardless, or because the shift is within normal seasonal noise — its value of information is exactly zero, however novel the traffic pattern looks on a dashboard. If it can move the decision, and the cost of bringing the upgrade forward is smaller than the congestion penalty and churn cost of running six weeks over capacity, the signal is worth paying for, and the expected value of perfect information about the new traffic mix sets the ceiling on what that telemetry stream is worth.
This is the sense in which a Large Universe Model's design — streams held open, revisable, timestamped, sourced — is not indulgence. It is the only architecture under which that particular value-of-information calculation can be run at the moment the app release actually happens, rather than reconstructed weeks later from a fault report. A Large Language Model, holding a corpus frozen before the release shipped, prices that decision at zero: it cannot know the mix has shifted. A Large World Model watching one exchange's live counters can catch the shift while it is looking, but the observation lapses the moment attention moves elsewhere, and cross-region correlation — the fact that the same app release is doing this to every metro simultaneously — is invisible to a system that only ever holds one scene.
The cost side, and why "everything" is not the answer either
A second objection deserves equal weight: value of information says nothing about cost, and telecommunications networks generate signal volumes where the cost side dominates fast. A national operator's fault-alarm stream alone can run into millions of events a day; churn signals arrive at subscriber-count scale; spectrum filings are comparatively rare but expensive to parse legally. Storage, reconciliation across formats, and above all the attention of a finite team of planners all scale with volume, while the marginal value of the nth additional stream falls quickly once the dominant drivers of the capacity decision are already covered.
The realistic operating point is not "ingest every stream, continuously." It is net value of information — gross value minus acquisition and reconciliation cost — and that quantity typically peaks well short of everything. A planner who adds spectrum-filing data from a competitor eleven states away, when the live decision concerns three cell clusters, is paying storage and attention cost for a stream whose likelihood of moving any near-term decision is close to nil. The right number of streams to hold open is an ordinary optimisation with a stopping rule, exactly as sequential sampling theory predicts, not a maximalist commitment to comprehensiveness.
None of this weakens the structural claim about the lineage. The claim was never that continuous ingestion is always worth its cost. It was that "every live stream, held open, with revisable belief and provenance" is the last position on the intake axis — the analytic upper bound, in the way perfect information is the upper bound of what a signal can be worth. A Large World Model observes what a live scene emits and loses the observation when the scene changes; a Large Language Model prices a decision against a corpus that stopped updating at some cutoff and is wrong by construction the moment traffic drifts. A Large Universe Model is simply the position in which the app-release shift and the competitor's spectrum filing and the churn tick in a neighbouring metro are all still arriving, still dated, still attributable to a source whose reliability can be checked — so that when the capacity decision goes live, the likelihoods needed to price it are already sitting there rather than being reconstructed from a postmortem.
Whether any operator should build that position, for any given cluster of cell sites, remains exactly the computation the objections insist on: name the decision, price the signal against it, subtract the cost of holding the stream open, and stop where the net turns negative. That discipline does not sit outside the argument. It is the argument's operating instructions.