13 November 2026. Since Thursday afternoon a good part of the working world has rediscovered the telephone. Outlook, Teams and Copilot are down or degraded globally, Microsoft points to elevated load after servers became unexpectedly unavailable, and no, it was not a cyber attack1.

The interesting line is not the outage, it is the recovery. Servers were put back into service, then fell over again as soon as real traffic arrived2. The system handled failing better than it handled coming back.

Anyone who has commissioned a building knows that feeling. The way back is the part nobody rehearses, and it is designed years before anyone needs it.

Power and procedures, not components

Uptime Institute’s Annual Outage Analysis 2026 finds power still the leading cause of impactful outages, dominated by uninterruptible power supplies, transfer switches and generators3. In its 2026 resiliency survey, 92% of respondents say human error contributed to their most recent impactful outage, 31% of them calling it a major contributor, with failure to follow established procedures the leading driver4. More than half of operators put their last major outage above 100,000 dollars, and one in five above a million5.

So the money goes into redundant components, and the sites fall over in the gaps between them, usually with somebody standing in front of a console deciding what to press.

Do this: write down your restoration sequences, then put a date next to each one for the last time it was tested in a degraded state, under load, by people who actually work nights. The blanks are the risk register.

The defects live between the packages

A facility programme is three projects wearing one hard hat: civil works, electromechanical, and the systems layer that starts last because it waits on the other two. Each package is run by competent people and arrives compliant with its own specification. Then they meet.

The classics: a generator start sequence specified at the generator and not across the transfer chain. A chilled water setpoint agreed verbally with the mechanical contractor and never typed into the points list. Alarm priorities that disagree between the electrical package and the console, so the console beeps at everything and the night shift learns to ignore it.

These surface at integration, which is where a requirements error costs 21 to 78 times its original price, and between 29 and more than 1,500 times once the site is live6.

Do this: give one person the interfaces, across every package, with the authority to settle an argument between two contractors. Hold the interface review before the long-lead orders go out. A coordinator without arbitration power produces excellent minutes.

What commissioning can and cannot catch

Integrated testing does find interface defects, which is the argument for doing it properly, on the full transfer sequence, with load banks rather than optimism. Two things limit it.

It sits at the end of the programme, competing with a move-in date, which is a fight it rarely wins. And it proves the nominal case, whereas sites fail with a unit in bypass, a tired battery string, or a chiller locked out for maintenance on the one week it mattered.

On an infrastructure relocation for a broadcaster, the thing that paid off was asking the people who would be on shift which failures they expected, during design, and using their list as the acceptance basis. Their list was better than ours. It changed what the contractors quoted, and the operations team turned up at commissioning already knowing what the building was meant to do.

Do this: put the degraded-mode scenarios in the tender. Put them in the commissioning plan instead and you will be asking for money that was spent eighteen months earlier.

Procedures are a design output

Procedures decide availability, and they are usually written last, by whoever knows least about how the site will run, then delivered in a manual that lives in a cupboard in another building.

Three cheap changes: write the sequences of operation with the operations team while the drawings are still open, train on the real installation during commissioning rather than on slides three months later, and design an alarm scheme that separates what needs action from what needs a glance.

People who helped write a procedure tend to follow it. People who were handed one tend to improvise, and improvisation shows up two years later in an incident report with a timestamp at 03:14.

The reporting rules are design requirements

Article 12 of Directive (EU) 2023/1791 applies to data centres from 500 kW of installed IT power, and Delegated Regulation (EU) 2024/1364 sets what goes to the European database each 15 May: energy, water, waste heat and ICT capacity7. On 21 September 2026 the Commission adopted a delegated regulation establishing a common EU rating scheme built on those indicators8. Germany sets a lower threshold, from 300 kW, and an earlier deadline of 31 March9.

Do this: specify sub-metering that separates IT load from mechanical and electrical losses, water metering on the cooling circuit, and the waste heat interfaces, then make the monitoring system export the indicators. Skip it and somebody will be rebuilding those numbers in a spreadsheet every spring until the building is demolished.

Before the second site

Standardising the control configuration across sites pays, within limits, and the limits are documented. Reuse does not start at zero: about 5% of new development cost goes to assessing and assimilating an existing component, understanding a configuration accounts for 47% of maintenance effort, and interface checks grow combinatorially, so modifying 2 modules out of 10 means revalidating 17 interfaces out of 4510. Building a reusable configuration costs around 50% more than a one-off, and reusing it unmodified around 20% of new development11.

Common configurationFirst siteEach later siteBreak-even
50%+25%−40%Between site 2 and 3
30%+15%−24%At site 3

Do this: fund the standardisation premium outside the first project’s fixed price, give the common configuration an owner between projects, and let a change control board rule on variant or special. Copying last site’s configuration and then tweaking it is the expensive option dressed up as the cheap one.

Four questions

Who owns the interfaces between packages, and can that person end an argument between two contractors?

Which restoration sequences have been tested in degraded state, under load, in the past year?

Who wrote the sequences of operation, and was anyone who works nights in the room?

Which reporting indicators come out of the monitoring system, and which come out of somebody’s spreadsheet?


Notes

  1. Live coverage of 13 November 2026: Microsoft attributes the incident to elevated load after servers became unexpectedly unavailable, and rules out a cyber-security cause: techradar.com. Scale, from Downdetector: close to 130,000 Outlook reports from United States users in the hour after 1 p.m. Thursday, across 74 countries and territories, The Washington Post, 13 November 2026: washingtonpost.com. ↩
  2. ABC News, 13 November 2026: Microsoft returned a portion of affected servers to service, then reported they could become unresponsive again once they received traffic: abcnews.com. ↩
  3. Uptime Institute, Annual Outage Analysis 2026, eighth edition, 13 May 2026: uptimeinstitute.com. ↩
  4. Uptime Institute, Data Center Resiliency Survey 2026, reported in the Annual Outage Analysis 2026: 92% say human error contributed to their most recent impactful outage, 31% as a major contributor; failure to follow established procedures is the leading driver of human-error outages: networkworld.com. ↩
  5. Uptime Institute, Annual Outage Analysis 2026, citing its 2025 annual survey: 57% put their most recent major outage above 100,000 dollars, one in five above one million: intelligence.uptimeinstitute.com. ↩
  6. Haskins, B., Stecklein, J., Dick, B., Moroney, G., Lovell, R. and Dabney, J., « Error Cost Escalation Through the Project Life Cycle », INCOSE International Symposium, vol. 14, 2004, pp. 1723-1737: ntrs.nasa.gov. ↩
  7. Directive (EU) 2023/1791, Article 12, and Commission Delegated Regulation (EU) 2024/1364 of 14 March 2024, reporting annually by 15 May for the previous calendar year, first report due 15 September 2024: eudca.org and dccomply.com. ↩
  8. European Commission, COM(2026) 500 of 21 September 2026 on the energy efficiency of data centres in the EU, published with the delegated regulation establishing a common Union rating scheme, C(2026) 3472: eur-lex.europa.eu. ↩
  9. Germany, Energy Efficiency Act: data centre defined from 300 kW of non-redundant rated connection power, reporting extended to sites between 200 and 500 kW from 1 July 2025, national deadline 31 March: dccomply.com. ↩
  10. Center for Software Engineering, University of Southern California, COCOMO II Model Definition Manual, version 2.1, 2000, sections 2.4.1 and 2.4.2, after Selby (1988) across nearly 3,000 modules at the NASA Software Engineering Laboratory, with the 47% maintenance figure from Parikh and Zvegintzov (1983) and the interface check formula from Gerlich and Denskat (1994): COCOMO II manual. ↩
  11. Poulin’s model, relative cost of reuse and relative cost of writing for reuse, summarised in « Integrated cost model for software reuse »: researchgate.net. ↩

About the author. Bernard Lecuivre leads industrial and infrastructure programmes from Brussels through Coeos Transformation. He currently heads a technology project office running a data centre relocation with physical server migration, including infrastructure design input for new buildings and continuity of operations throughout. Previously head of the project office for the handover of a federal department’s complete business application landscape, covering development, infrastructure and production with a team of 20, and programme manager for a €5M digital transformation of proton therapy product development against a 10% internal efficiency target. Certified Prince2 Practitioner, Agile Project Management, Professional Scrum Product Owner, ITIL and OCEB.

For an interface review or programme leadership on a facility build: coeostransformation.eu | LinkedIn