Bandwidth, Latency, Diversity: The Three Prices of the Wire

To model interconnect as an asset class you need its price vocabulary. Unlike compute, which the market summarizes crudely as dollars per accelerator-hour, the wire has three distinct prices, and confusing them is the most common analytical error. They are bandwidth, latency, and route diversity. Each behaves differently, each is scarce under different conditions, and each supports a different investment.

Bandwidth is capacity: how many bits per second a route can carry. It is the price most people intuit, and it is the one that has fallen fastest over decades, driven by advances in optical transmission that pack more wavelengths onto each fiber and more bits onto each wavelength. Falling per-bit cost is not the same as falling total spend. As unit prices drop, demand for AI workloads rises faster, so aggregate capital committed to bandwidth keeps climbing. For the investor, the signal is not the price of a bit but the rate at which new capacity is being lit relative to the rate at which traffic is growing. When traffic outpaces lit capacity on a specific route, that route develops pricing power even as the global average price per bit keeps falling. Averages conceal the routes that matter.

Latency is delay: how long a bit takes to traverse the route, bounded ultimately by the speed of light in glass, roughly two-thirds of the speed of light in vacuum. Latency cannot be bought down below the physical floor set by distance; it can only be minimized by choosing shorter physical paths. This makes latency a real estate problem disguised as a networking problem. For inference workloads serving interactive users, and for financial and real-time systems, a few milliseconds of round-trip delay is the difference between a viable location and a dead one. This is why the map of valuable data center land is drawn not by cheap power alone but by proximity, in fiber-route terms, to population and to other compute. A site fifty milliseconds from its users is a different asset from one five milliseconds away, regardless of how cheap its electricity is.

Route diversity is the third price, and the least appreciated. It measures how many physically distinct paths connect two points. Diversity matters because a single cut, an anchor drag across a subsea cable, a backhoe through a terrestrial conduit, can sever a route entirely if there is no independent alternative. For any workload that cannot tolerate an outage, and increasingly that means the workloads that anchor the largest contracts, the relevant question is not the capacity of the best path but the number of independent paths. A campus served by one glorious high-capacity route and no backup is fragile; a campus served by three diverse routes commands a premium precisely because it can promise continuity. Diversity is where insurance-like value lives in the network.

These three prices interact in ways that reward careful thought. A route can be abundant in bandwidth but poor in diversity, making it cheap per bit and yet strategically risky. A location can have excellent latency to one metro but terrible diversity, because all its paths funnel through a single chokepoint. The analyst's job is to hold all three in view at once and to notice when a market is pricing only one. During the early cloud era, bandwidth dominated the conversation; latency and diversity were afterthoughts. The AI era is reweighting them, because training clusters are latency-and-diversity-sensitive at the fabric scale, and inference delivery is latency-sensitive at the metro scale.

Consider how the three prices map onto the three scales from the previous lesson. At the subsea scale, diversity dominates: the strategic value of a new transoceanic cable is often less about raw capacity, which is abundant, than about providing an independent path that does not share a chokepoint with existing cables. At the metro scale, latency dominates: the value of dark fiber between two campuses is set by the physical distance it saves. At the intra-cluster scale, bandwidth dominates, because a training run consumes enormous east-west traffic between accelerators, and any shortfall stalls the whole computation while chips wait for data.

The practical takeaway is a discipline. Whenever you evaluate an interconnect asset, price all three dimensions explicitly and ask which one the current market is underweighting. My read: the market reliably underprices diversity, because diversity produces no value until the day a cable is cut, and markets discount low-probability, high-consequence events. That systematic underpricing is a recurring source of mispriced assets, and the analyst who prices diversity correctly holds an edge that a bandwidth-only model can never see. Carry this vocabulary forward. In the modules ahead, every capital flow, ownership decision, and supply-chain constraint can be decoded as a bet on one or more of these three prices.