A roof with blue cracks
Last night I dreamt of a hurricane. My mother, my brother and I were inside, and the roof was about to give way. Before it broke, the white ceiling turned blue along the cracks. It showed me where the pressure was, and which way to move to find shelter.
I woke up thinking about SCADA, power grids, hospitals and banks. We are quietly adding AI to the systems that hold up daily life. What is our blue paint? And when the power goes out, does the warning go dark with everything else?
What actually stops
A data center failure is never one failure. It is a chain, because every service assumes the one beneath it is up.
Power sits under all of it. Water pumps need electricity. Telecom towers run on batteries that last hours. SCADA needs power and a network. Operators go blind exactly when they most need to see.
The critical chain
Critical services sit on six layers. A failure anywhere travels up to the people at the top, and recovery has to climb back up in order.
- 6People and services
Hospitals, banks, water, trading: what the public feels - 5AI layer
Fraud scoring, imaging triage, alerts, summaries. Newest, least understood - 4Core software
SCADA, hospital records, payments, claims. Often one shared provider - 3Data centers and cloud
A few regions host many sectors - 2Network and cooling
Telecom on batteries for hours; cooling faults stop whole halls - 1Power
Under everything. Black-start plants restart the grid from zero
The recovery loop. Restarting power needs control systems and communications, which need power. Every new automated layer is one more thing that must come back before the rest can.
Where each chain breaks first
| Sector | First link to break | What people feel | Manual fallback |
|---|---|---|---|
| Hospitals | Records and imaging, then the AI triage on top | Scans read in plain order; notes go to paper | Paper charts, runners, reading in sequence |
| Banking | A cloud region or payment processor | Cards declined, transfers stuck | Stand-in approval rules, branch cash |
| Water | Power to pumps, then SCADA visibility | Pressure drops; boil-water notices | Generators, manual valves, crews on radio |
| Power grid | Control-room software and comms | Outages spread before anyone sees them | Black-start plants, radio dispatch |
Where AI already holds weight
AI is now in almost every hospital and bank. Most of it is a helpful layer, but a few beams are already load-bearing.
Load-bearing today: card fraud screening on every transaction, anti-money-laundering monitoring, and imaging triage that moves likely strokes up the queue. Doctors themselves name deskilling as the top risk.
The deepest layer is still not AI. In 2024 a cyberattack on one claims processor, Change Healthcare, disrupted pharmacy and billing across much of the US for weeks. That July, one faulty CrowdStrike update grounded flights and sent hospitals to paper. AI is being stacked on top of exactly these single points of failure.
Built to be always online
This month Anthropic described Scion, the system that replaced Git for its own code. Daily merges grew from about 2,500 in May 2026 to over 30,000 by September, mostly written by agents. Scion assumes its clients are agents and automated jobs, online by default. For an internal code repository that is sensible, and Anthropic says Scion is not a general Git replacement.
As a premise for load-bearing infrastructure, it is the assumption a hurricane breaks.
Who can read the code when it breaks?
Speed is the problem, however we sugar-coat it. At tens of thousands of changes a day, nobody holds the whole system in their head, and systems balloon faster than anyone can read them.
We have seen a slow version of this. In 2020, US state unemployment systems buckled and governors asked for COBOL programmers, because the people who understood that code had retired. The code still ran. Nobody could read it under pressure. AI compresses decades of that drift into months.
The next generation is not less intelligent. The path to understanding is being skipped. And in a disaster, asking an AI to explain the code fails too: the AI sits in a data center that is also dark.
Gaps, not lies
The danger in AI research summaries is less invention than omission: a confident summary of half a document. What the model did not read does not exist for it, and a tidy list of citations hides that.
Omitted never means unimportant.
Verification needs expertise, and the people who lean on summaries most are the least able to see what is missing. That is why I built EasyRead for our own AI systems: it keeps the full source intact and tells the model exactly what it left out of view. Know thy thinking partner, its strengths and its gaps, and build around them.
Blue Paint as a system
The proposal
A layer that makes pressure visible along the whole chain, and keeps working when the chain breaks.
Foundation: read-only, honest about uncertainty, humans decide.
It starts with an offline readiness audit: map each critical service down through the six layers, mark the single points of failure, and score how long people could run it by hand. Then paint the strain, rank the safe moves, and leave behind the printed and physical versions that still work in the dark.
Warnings that work in the dark
- The warning must survive the disaster it warns about. Physical indicators, battery-backed displays, printed dependency maps.
- Show the safe path, not just the alarm. Where, how fast, and what to do next.
- Keep a slow, boring core. The part that keeps people alive should change rarely and be understood fully.
- Measure offline readiness. Not how many systems use AI, but how many could keep working without it.
- Practise by hand. Drills, radios, paper procedures, and people who remember how.
When the data center goes down, who can still read the cracks?
Sources
Izza Masud is the founder and primary architect of Distilligent AI.
ORCID: 0009-0001-0647-7577