Every answer and its explanation appears here once you have finished the path. Each one then links to the matching glossary entry, where the concept is set out in full with its worked example.
1. A conventional server rack dissipates a few kilowatts. A rack dedicated to training large models can exceed a hundred. What does that factor change for an insurer?
Severity: a thermal failure destroys far greater value within minutes
Rack density measures the electrical, and therefore thermal, power concentrated in a single rack. It was historically a few kilowatts and changed scale with artificial intelligence workloads: racks housing graphics processor clusters for training reach several tens of kilowatts, and sometimes exceed a hundred. What an insurer should read in that figure is not a probability but a severity, and the distinction is the heart of the question. An equipment's failure frequency does not grow mechanically with its power; what grows is what a failure carries away. A thermal failure on a very high density rack can damage equipment of considerable value within minutes, and the physical proximity of components concentrates the exposure instead of spreading it: the same floor area now bears a multiple of the value it once bore. It should be added that this density is not an operator choice that could be moderated, it follows from the workload: training a large model requires processors that talk to each other, hence physically close, which rules out diluting the value across more square metres.
Glossary entry · densite-kw-rack2. Beyond a certain density air no longer suffices, and a coolant is circulated close to the components, or directly onto the chips. What new peril does that introduce?
Leakage: coolant a few centimetres from very high value electronics
Liquid cooling is far more effective than air at dissipating high thermal densities, which is what made it necessary as soon as power per rack exceeded what ventilation can absorb. It is therefore not an optimisation choice but a consequence of the workload, and one cannot avoid it by going back to air. The peril it introduces belongs to a different family from the one it solves, which is what makes it interesting to underwrite: fluid is installed immediately beside, sometimes in contact with, electronics of considerable unit value. A leak on a direct loop can damage an entire cluster, combining heavy physical damage with business interruption, where the classic thermal risk showed itself as a stoppage without destruction. So this is a displacement rather than an addition: overheating recedes, internal water damage appears, and a site rated highly on thermal control may rate poorly on this. For examining a site, the useful question is not whether it cools by liquid, which density dictates, but how it detects a leak and what it shuts off automatically when it does.
Glossary entry · refroidissement-liquide3. An operator reports a fleet average PUE of 1.1, where a traditional centre sits between 1.5 and 2.0. How should an underwriter read that figure?
As real efficiency, but also as the sign of cooling whose failure weighs heavily
PUE compares the installation's total energy consumption with the energy actually doing computation: a PUE of 1.0 would describe a centre where nothing is lost to cooling, conversion or lighting. Traditional centres sit between 1.5 and 2.0, optimised hyperscale installations approach 1.1. It is a good efficiency measure, used to compare operators and set regulatory targets, but reading it as a risk score would be a mistake, and that is the question's trap. A low PUE is obtained by cutting everything that is not computation, hence by cooling as sparingly as possible, often by liquid or immersion, and sometimes by reducing margins. It therefore signals heightened dependence on systems whose failure causes critical overheating and major business interruption. What the indicator does not capture matters just as much: neither the energy's source, nor water consumption, nor power redundancy. Two sites with the same PUE can present very different risk profiles, and the figure excuses none of the usual questions.
Glossary entry · pue-power-usage-effectiveness4. A Nordic centre is designed to reject its heat using outside air, with little or no mechanical cooling. That choice improves its PUE. What does it tie its operations to?
The local climate, hence physical climate risk and its drift
Free cooling uses outside conditions, cold air or spring water, to reject heat without resorting, or resorting as little as possible, to mechanical cooling. When the ambient temperature is low enough, the energy saving is considerable and PUE improves accordingly, which explains siting many facilities in cold regions. The reasoning is sound and the saving real; what must be seen is what it makes the site depend on. The arrangement's performance follows the local climate directly, so it is seasonal by construction and vulnerable to extreme heat episodes. A site sized on old climate norms sees its thermal margin shrink as summers warm, and a heatwave beyond the design thresholds can saturate the system and force load shedding or server shutdown. This is exactly the mechanics of non-stationarity: the fault lies not in the equipment but in the reference used to size it. The underwriting question is therefore the site's design date and whether its margin has been revised since, more than the technology used.
Glossary entry · free-cooling5. An evaporatively cooled centre sits in a zone classified as high water stress. Its WUE, in litres per kilowatt-hour, is unfavourable. What is the twofold nature of the risk?
Operational through shortage, and regulatory through use restrictions
WUE compares a centre's water consumption with the energy consumed by its computing equipment, in litres per kilowatt-hour, and it complements PUE by illuminating a long-neglected externality. Evaporative cooling's needs are intense, and particularly so in the hot dry regions where such sites are sometimes built, precisely because power is available and land affordable there. The rise of artificial intelligence has sharpened the issue, since training and inference concentrate high thermal density, hence heat to reject. For the insurer the risk is twofold and the two sides are not handled together. The first is operational: a water shortage can constrain operations and trigger business interruption, without any equipment having failed. The second is regulatory and reputational: a heavy consumer in a water-stressed zone is exposed to use restrictions, and a restriction is a political event whose occurrence does not model like a natural hazard. It is that second side that makes the exposure hard to price, more than the first.
Glossary entry · wue-water-usage-effectiveness6. Several American grid operators, Dominion Energy in Virginia among them, have reported connection delays exceeding several years for new data centres. How does that change the nature of project risk?
It becomes a schedule risk depending on a third party the operator does not control
The queue is the delay between a site's request for grid connection and its actual energisation, and that delay has stretched to several years in the highest-demand zones. The cause is arithmetic: the sector's growth in power demand, driven notably by training large models, outpaces operators' ability to reinforce lines and substations. What this changes fits in one sentence, and it is what the question seeks: a site complete as a building can stay unusable for months or years for want of sufficient power. The risk therefore stops being mainly technical and financial, two things the developer controls, and becomes a schedule risk hanging on a third party's decision and works. For insurance the practical consequence is a shift in the cover sought: what worries is no longer damage to the works but delay in start-up, whose triggering event is external both to the insured and to the site. Some projects meanwhile use on-site generators while awaiting definitive connection, which shifts the risk again, and that is the subject of the next two questions.
Glossary entry · file-attente-raccordement-electrique7. An operator installs a dedicated gas plant on the consumer side of the meter, to power its servers without waiting for the grid. What has it just added to its risk profile?
Industrial generation perils, explosion and mechanical failure, which it handles poorly
Behind-the-meter generation means generating assets installed on the consumer side, supplying a site directly without passing through the public distribution grid. Faced with saturated networks and connection delays of several years, it allows bypassing the queue by relying on dedicated sources, gas plants, fuel cells, even small nuclear projects. The consequence is heavier than a change of supplier, and this is what to retain: the digital operator becomes an energy producer, with the responsibilities that follow. For insurance this shifts a substantial share of risk onto generating assets that operator historically handles poorly, since it is not its trade, and it adds to a single site industrial perils that IT risk did not carry: explosion, mechanical failure, environmental liability. The result is accumulation rather than substitution, at the same address and often in the same policy: dependence on the grid disappears, industrial risk arrives, and the second is of a kind the operator has never had to insure.
Glossary entry · production-derriere-compteur8. A lithium battery storage system supports a centre's critical loads during an outage, complementing or replacing traditional UPS and generator sets. Why has it become an underwriting subject in its own right?
A cell's thermal runaway propagates and gives a fire that is hard to extinguish
A battery storage system stores energy to release it when needed, ensuring supply continuity, smoothing peaks and allowing intermittent renewable sources to be integrated. In a data centre it progressively complements or replaces UPS and generator sets to support critical loads during an outage. The peril it brings is specific and must be named precisely, because it is not handled like an ordinary fire: a cell's thermal runaway can propagate from cell to cell and trigger a long-duration loss, hard to extinguish since the reaction feeds itself, and releasing toxic gases that complicate intervention. This places the subject at the crossing of fire risk, environmental risk and business interruption, three lines that do not share reflexes. A further difficulty is not technical: the relative novelty of these installations at scale leaves little loss experience, hence little basis for pricing, putting underwriting in the position familiar from cyber, that of a real risk whose history says almost nothing yet.
Glossary entry · bess-stockage-batterie9. A centre equips itself with a microgrid combining generation, storage and consumption, able to island itself from the public grid. How does its power risk profile change?
Grid-caused interruption recedes, and generating assets enter the balance sheet
A microgrid is a local electrical network combining generation, storage and consumption, able to run connected to the main grid or islanded from it. For a data centre it brings real resilience: if the public grid fails, the site keeps drawing on its own sources, panels, fuel cells, batteries, generator sets. That autonomy answers the tension between the explosion in artificial intelligence power demand and public grids' ability to meet it, a tension whose most visible expression is the connection queue. What an insurer should read is a trade rather than a gain, and it is the thread running through this path's last four questions: dependence on the external grid falls, so grid-caused interruption risk falls with it, but generating and storage assets enter the balance sheet with their own perils, which must be insured and in which the operator is no specialist. The net outcome therefore depends entirely on how those assets are designed, tested and operated, not on having a microgrid. A poorly proven autonomous site is less reliable than a site depending on a solid grid.
Glossary entry · microreseau