The Integrated Site Scorecard: Electrons and Photons Together
The purpose of this course has been to install a second axis of analysis alongside the one the market already runs. The market scores power with rigor. This lesson assembles the complementary discipline, scoring the network, into a single integrated frame so that a site is judged on both supply chains at once rather than on the one that happens to be visible.
Begin by restating the thesis in its operational form. Network architecture is the under-tracked constraint on AI compute delivery, which means that for any candidate site, deliverable compute equals the minimum of two independent readiness scores: an electron score and a photon score. A site is only as good as its weaker axis. A gigawatt of cheap, firm power behind a single fragile fiber trunk is not a prime site; it is a power asset with a delivery problem. The integrated scorecard forces both axes into view and refuses to let a strong power number paper over a weak network position.
The photon score has five components, drawn directly from the earlier modules. First, long-haul diversity: how many physically independent routes serve the site, verified by path separation rather than by counting commercial circuits that may share a conduit. Second, carrier independence: whether those routes are owned by genuinely separate parties, an independent Lumen path and an independent Zayo path, for example, rather than two products from one carrier on one corridor. Third, interconnection richness: how close the site sits to a meet-me point where many networks cross-connect, which determines pricing, optionality, and the ease of adding capacity. Fourth, latency geography: the fiber-kilometer distance, and resulting latency, to the target user population and to any partner campuses in a distributed cluster. Fifth, for workloads with international exposure, subsea access: how many independent subsea systems, through how many separate maritime geographies, reach the region, and how far the landing stations sit from the site. These five convert the vague instinct that fiber matters into a scored, comparable assessment.
Apply the frame and the site taxonomy sorts itself. A prime site clears both axes: abundant, firm power and a high photon score, meaning diverse independent long-haul, a nearby rich interconnection point, short latency to its markets, and, where relevant, resilient multi-geography subsea reach. A power-rich, network-poor site, the classic trap of this cycle, has cheap megawatts in an empty county but a single trunk, one carrier, distant interconnection, and long backhaul; it looks attractive on a power-only scorecard and carries stranded-asset risk the moment its single path is tested. A network-rich, power-poor site, common in and around established metros, has superb connectivity but cannot secure enough firm power; it is constrained on the other axis. The integrated view is what separates these cases, and a power-only view structurally cannot.
The distributed-training wrinkle sharpens the point. Because frontier training runs are outgrowing single-site power envelopes, operators increasingly link multiple campuses into one logical cluster, exchanging gradients and checkpoints between sites. That architecture makes inter-site latency and inter-site path diversity first-order design constraints, not afterthoughts. Two campuses that individually pass a power screen may be useless as a paired cluster if the fiber between them is high-latency or rides a single trunk that a backhoe can sever. My read: the rise of multi-site training elevates the photon axis further, because it turns the network between campuses into part of the computer itself, and a slow or fragile link between them degrades the whole cluster, not just the connection.
The method in practice is disciplined and repeatable. For each candidate site, produce two scores, not one. Run the established power and grid-interconnection analysis, then run the five-component photon analysis with equal rigor, and take the site's grade as the weaker of the two, adjusted for how easily the weaker axis can be remediated. Remediation asymmetry matters and is worth pricing explicitly: sometimes a fiber deficiency can be cured by financing a new lateral to a nearby diverse route, a soluble capital problem with a known cost and a short timeline, while a power deficiency in a constrained grid may be effectively unsolvable for years. Scoring both axes lets you see not just the current grade but the cost and feasibility of improving it, which is often where the real value or the real trap hides.
My read to close the lesson: the market's power-only scorecard systematically overvalues power-rich, network-poor sites and undervalues both the fiber remediation that could rescue them and the fiber fragility that should disqualify them. That asymmetry is the analyst's edge. The frame here, deliverable compute as the minimum of an electron score and a five-part photon score, is not merely descriptive; it is a pricing tool that surfaces value and risk the consensus is still rounding away. Use it and you evaluate sites the way the constraint actually behaves, on both supply chains at once, rather than the way the market currently habits itself to look.