Distributed data centers are more secure

A gigawatt AI data center feels strong. A fortress: fences, guards, cameras, backup generators, thick walls. Mathematically, it can be the most fragile thing you have ever built.

The cloud is a building. A very expensive building, wired to a substation, fed by fiber, filled with GPUs, and cooled by industrial water and power systems.

For about twenty years, "data center security" mostly meant cybersecurity: firewalls, identity, encryption, zero trust. That work still matters. Physical exposure is now the larger gap. AI turns a data center into a machine that converts electricity into intelligence, and those machines are now among the most capital-dense objects on Earth.

The numbers are public. OpenAI's Stargate was announced as $500B over four years, later framed as a $500B, 10GW program. Meta announced a $10B, 4 million square foot AI campus in Louisiana. TSMC put its US investment at $165B.

The bigger the machine, the more dangerous it is to put it all in one place.

What a data center actually is

Strip the marketing away and a data center is a machine for two jobs: push power into racks, pull heat out of them, safely, for years. Most of the cost and most of the risk lives in that plumbing, not in the walls.

Power comes in high and gets stepped down stage by stage. The grid delivers high voltage, often 138, 230, or 345 kV. An on-site substation steps it to medium voltage, around 11 to 33 kV. Near the hall, transformers drop it to about 415 V, then the rack takes it lower, and the chip runs at under a volt.

Why keep voltage high for as long as possible? Because power lost as heat in a wire rises with the square of the current, P = I squared times R. Higher voltage means lower current for the same power, which means less heat and less copper. That single equation shapes the entire building.

The hardware is heavier than it sounds. The big step-down transformers run 50 to 100 MVA each (an MVA is roughly a megawatt of capacity), they are custom-built, and lead times can pass a year because they depend on a special grain-oriented steel with few suppliers. You cannot order one next-day.

So operators buy redundancy and name it precisely. N means exactly enough gear to run. N+1 means one spare. 2N means a full second set. The Uptime Institute turns this into Tiers. A Tier III site is concurrently maintainable: you can service any part without going dark. A Tier IV site is fault tolerant: a single failure should not take it down.

Inside, the building is already broken into pieces on purpose. A hall is split into pods of roughly 1.6 to 2.5 MW, each with its own transformers and switchgear. Each pod also gets a diesel generator the size of a locomotive engine, around 3 MW and 4,000 horsepower, holding a day or two of fuel. Batteries in the UPS carry the load for the sixty or so seconds it takes a generator to start, reacting in under ten milliseconds. Even one building is built around dividing risk. The whole industry already talks this way: engineers will tell you an electrical fault has a smaller "blast radius" than a cooling failure. Blast radius is shop language, borrowed here.

Compute is heat

Every watt you put into a GPU comes back out as about a watt of heat. A 100 MW hall is a 100 MW heater. Cooling is half the machine.

Efficiency gets one number: PUE, total facility power divided by IT power. A PUE of 1.5 means you burn an extra half watt of overhead for every watt of compute. The rough industry average is around 1.6; the best hyperscalers run near 1.1. Water gets its own number, WUE, in liters per kWh, and it adds up fast. A mid-size site can drink hundreds of millions of liters a year if it leans on evaporative cooling.

Chillers are the biggest non-IT draw, so the cheapest cooling is the cooling you skip. "Free cooling" uses the outside air directly, and it works best where the air is dry, because dry air cools further through evaporation. That is one quiet reason a desert is attractive. As racks get denser, air alone stops working, and direct-to-chip liquid cooling becomes mandatory rather than optional.

Why everything wants to be in one place

The case for concentration is real. Training a large model is synchronous. Tens of thousands of GPUs work in lockstep and wait on each other, so the links between them have to be fast and short. High-speed copper only reaches a couple of meters, and the in-rack NVLink fabric is far faster than the network between racks. The cheaper your tokens need to be, the tighter you pack the GPUs.

That is why rack power keeps climbing. Racks sat under 10 kW for years. Nvidia's GB200 NVL72 packs 72 GPUs into one liquid-cooled rack at 120 kW and up. The 2027 roadmap, built on a new 800 VDC power architecture, points at 600 kW to a megawatt per rack.

The power physics is just as blunt. Deliver 600 kW at the old 54 V and you need around 11,000 amps of current. Do it at 800 V and you need about 750. Since loss scales with current squared, that is roughly 200 times less heat lost in the same copper. Density is efficiency. That is exactly what makes it dangerous.

The grid is a failure domain too

Concentration does not stop at the fence. A training run is synchronized, so its power draw is jagged. Meta's Llama 3 paper describes tens of thousands of GPUs swinging power at the same instant, "on the order of tens of megawatts," enough to stress the grid. Engineers have literally run fake workloads just to keep the swings from hurting the power system.

At gigawatt scale, the failure domain extends to the grid itself. SemiAnalysis models these load swings and finds a real path to a cascade. On April 28, 2025, the Iberian grid lost about 2,200 MW in a span of seconds, and Spain and Portugal went fully dark in under half a minute. In July 2024, a single fault knocked roughly 1.5 GW of Virginia data centers off the grid at once. Put a gigawatt behind one interconnection and a local hiccup becomes a regional one.

The plan we didn't build

When we started planning gigawatt-scale AI capacity in Israel, the first instinct was the obvious one. Build the biggest possible bunker. One giant site, one campus, one national machine. One site looks clean on a slide, simplifies procurement, tightens the networking, and tells investors a story they recognize.

Then we asked one security question. How much of the country's AI capacity are we putting inside one failure domain?

That question killed the plan. In Israel, geography is an operating condition. Drones, rockets, grid events, and war-risk insurance are not consulting words here.

The UAE made it concrete. On March 1, 2026, Iranian drones struck two Amazon AWS data centers in Dubai directly. Two of three availability zones in the ME-CENTRAL-1 region went down. 109 services failed. Banks, payments platforms, and ride-hailing apps went dark. The buildings had every standard protection: fences, cameras, redundant power, backup generators. None of it was rated for a drone strike. AWS told customers to migrate workloads to another region and warned that recovery would take weeks, given the physical damage involved. BBC, Data Center Dynamics.

Russia's war in Ukraine pushed the same move at larger scale: a smaller target per site, spread across real failure domains.

The unit of security is the failure domain

A failure domain is a part of a system that can fail on its own without taking the rest down. Google Cloud uses exactly this language: a zone or region is a failure domain, and spreading across them improves availability. Google's own example: two things at 99.9% availability, placed in separate failure domains, drop the odds of both failing at once toward one in a million, if they are truly independent.

The same idea already exists in several serious fields. Apply it to GPUs.

Field	What they already call it
Cloud	Failure domains, zones, regions
Insurance	Probable maximum loss
Power grids	N-1 contingency planning
Defense	Mission assurance
Critical infrastructure	Consequence-based risk (CISA)
Data centers	Uptime Tier III and Tier IV

How much value, power, and national capability can be lost in one local event, and how fast can the rest keep running?

The math is just division

Take a $40B AI platform. There are a few ways to build it.

One giant site: one event can reach the whole $40B. Split it into four independent sites: each holds about $10B. Split it into twelve 100 MW modules: each holds about $3.3B. That is the entire trick: divide the asset.

Architecture	Failure domains	Value per domain	Concentration
One giant site	1	$40B	1.00
Four sites	4	$10B	0.25
Twelve modules	12	$3.3B	0.08

The concentration number is the same one regulators use for markets: square each site's share of the total and add them up. One site scores 1.0. Four equal sites score 0.25, which behaves like four real sites. Twelve modules score about 0.08, like twelve. The rule of thumb: distribution cuts your worst single-event exposure by roughly 1 minus 1 over n. Four sites is a 75% smaller worst case. Twelve modules is about 92% smaller.

To make value density concrete: Meta's Louisiana campus is roughly $10B across about 371,600 square meters, near $27,000 of investment per built square meter. Stargate is about $50M per megawatt at the program level. A boring GPU hall can be worth more per square meter than a famous tower: every rack is full of capital and every megawatt is monetized.

What distribution actually buys you

Four weak sites on a map are still four weak sites. Distribution only works when it creates real independence.

What distribution reliably changes is the blast radius. One giant site concentrates the full loss; four real failure domains split it to $10B; twelve modules to $3.3B. The worst single event shrinks, recovery time shrinks, and the leverage anyone could gain by threatening one site shrinks.

Distribution can leave the simple long-run average loss unchanged. If every site has the same independent failure probability, the long-run math can look similar whether you have one site or four. The questions that matter are different: what is the worst single event? How much capacity survives? How fast can we recover? Can one local failure become a national failure?

On the probability: if each site fails independently with probability p, the chance that all n sites fail together is roughly pⁿ. That gets very small very fast. But only if the sites are truly independent.

The moment all sites share one common weakness, the math collapses. A shared substation, shared fiber route, shared cooling dependency, shared control plane, or shared security gap can turn many sites back into one failure domain. The Iberian blackout and the Virginia event above are exactly that: common-mode failure in the wild, many buildings, one failure domain.

Distribution is valuable only when it removes common-mode failure. Independence is the control variable. Distance alone does not create it.

It costs more, and that is the point

Distribution is more expensive. Full stop. You duplicate substations, cooling plants, security, and staff, you lose economies of scale, and you pay a networking penalty when sites sit far apart. In a calm place you could call that waste.

In a high-risk place it is the floor, not a luxury. And the risk is getting worse, not better. Long-range ballistic missiles and cheap drones put fixed, high-value sites within reach from much further away, and AI compute itself is becoming a weapon. It runs the targeting, the intelligence, and the decision loop in a real conflict. The more strategic the compute, the more a single concentrated site reads as a single point of national failure. So I model single-event disruption scenarios the way insurers, cloud architects, and grid planners already do: failure domains, probable maximum loss, blast radius, continuity of operations. These are box sizes for risk planning, general enough that they never double as a target list. The smaller the box, the easier it is to insure, defend, finance, recover, and explain to a government.

The same lesson is now in orbit

The strange part is that the industry is about to make the identical mistake one altitude higher. I wrote about it in They're Made Out of Compute: the plan to fly data centers into space, packed into tight clusters, fragile, exposed, and so far undefended. Same instinct, same blind spot. Whether the failure domain is a desert campus or an orbital cluster, the question never changes. How much can one event take?

The Israel lesson

Israel is a small country with real advantages for this: sun, talent, urgency, and a genuine need for its own compute. Small countries do not get to treat infrastructure risk as theoretical. If AI compute becomes critical to defense, the economy, science, and government, then data centers are national infrastructure, and national infrastructure should not hang on one building.

Gigawatt-scale AI is the goal. We build it as a network of smaller factories: modular, hardened, connected, spread across real failure domains. A GPU hall wired to a few hundred megawatts, replicated across sites that can each fail independently.

When infrastructure becomes strategic, ask one question: how much can any single event reach?

---

In April 2024 I stopped my bike near home and filmed interceptors and incoming fire lighting up the sky over Israel. I have thought about it ever since. Every site plan I look at now is a list of failure domains.

Shai Magzimof

---