Networks, Automation, Connectivity & Data Centres · Chapter 13 of 18
Data centre engineering: what the numbers actually say
Tiers, PUE, rack density and cooling — with the published figures, and the four places the industry's own marketing contradicts its own survey data.
The previous chapter gave you the anatomy of a colocation deal. This one gives you the engineering, and it is worth reading carefully because there is an unusually large gap here between what is repeated in press releases and what the industry's own measurements show. Being the person who knows which is which is genuinely valuable in a facilities, infrastructure or capacity conversation.
Tier certification: the distinction almost nobody checks
Uptime Institute's tier system is the shared vocabulary, and there are two things about it that vendor material consistently gets wrong.
- Tier III means concurrently maintainable, not ‘99.982% uptime’. The Uptime standard does not assign availability percentages at all. The figures everyone quotes come from a 1990s white paper and are repeated by marketers rather than by the standard. What Tier III actually promises is that any single component or distribution path can be taken out of service for planned work without dropping the IT load. That is about maintenance, not about surviving faults — surviving an unplanned failure is Tier IV.
- There are three separate certifications, and only one of them means the building was inspected.
| Certification | What it actually covers | Validity |
|---|---|---|
| Tier Certification of Design Documents (TCDD) | A document package only. Uptime has not seen the built environment and states explicitly that it cannot speak to the standard the facility was built to | Expires two years after award for anything issued since 2014 |
| Tier Certification of Constructed Facility (TCCF) | Consultants visit the site and identify discrepancies between the drawings and the installed equipment | Lifetime |
| Tier Certification of Operational Sustainability (TCOS) | Staffing, processes and site management rather than the physical plant | Periodic renewal |
The question to ask a colocation provider, and why it works
“Tier III certified — which certification, and what date?”
The two-year expiry on design certification was introduced deliberately, because operators were marketing a design certificate indefinitely without ever building to it. An expired TCDD with no constructed-facility certificate is a marketing claim about a drawing.
This is a very cheap way to sound like someone who has actually procured data centre space, and it is a real diligence step rather than a trick question.
PUE, and the six years nothing improved
Power usage effectiveness is total facility power divided by IT power. The trajectory is worth knowing precisely because it contradicts the general impression that efficiency keeps improving: the surveyed industry average fell from 2.50 in 2007 to 1.65 by 2014, and has then sat between 1.54 and 1.59 for six consecutive years.
The explanation Uptime gives is entirely about economics rather than engineering. The cheap wins — hot and cold aisle containment, cooling control optimisation, electrical upgrades — had been taken by 2014. Beyond those, retrofitting an existing facility has a weak business case, and more than a quarter of surveyed facilities are over sixteen years old. Only about 15% of operators report 1.3 or better.
Two cautions when reading any published PUE figure.
The measurement boundary is not standardised. Google's reported figure explicitly includes substations, transformers, water treatment plants, cafeterias and office space. Most published figures do not. A like-for-like comparison would flatter Google further, and comparing two operators' self-reported PUE is comparing two different measurements.
A lightly loaded hall always reports a bad PUE, because fixed overhead is divided by less IT load. That is a measurement artefact of utilisation, not an efficiency problem, and treating it as one leads to bad decisions.
And climate is destiny. Microsoft reports 1.16 across the Americas and 1.28 across Asia-Pacific — the same company, the same engineering, different weather. That is precisely why Indian facilities run around 1.5 to 1.6, and it means an Indian data centre burns roughly 14% more facility power per IT-kilowatt than a Virginia one before you pay for a single electron. India's regulator has responded: government tenders for AI cloud infrastructure now require PUE below 1.35, which is an aggressive target in that climate.
Rack density: the number that corrects the AI narrative
Survey data across roughly 700 operators puts the modal rack density at about 9 kW in 2025, up from 8.3 the year before — and 7.5 kW if you exclude the facilities running above 30 kW. More than 80% of operators have no racks above 30 kW at all, essentially unchanged year on year. Roughly one in eight facilities has some racks in the 30 to 59 kW band.
Two implications that matter in an interview. First, the densification that is happening below 30 kW is mostly not AI — it is ordinary enterprise consolidation, because many mainstream servers now approach or exceed a kilowatt when richly configured, so a handful of servers plus storage and a top-of-rack switch crosses 10 kW without any accelerator involved. Only above about 60 kW per rack do AI and HPC workloads become the primary driver.
Second, if you are sizing or costing a facility, using the AI figure as your planning assumption will produce a design that is wrong for 96% of the racks. Both numbers are true; using the right one for the question is the skill.
Cooling: the thresholds, and what actually gets deployed
| Threshold | What changes | Note |
|---|---|---|
| ~15 kW/rack | Air begins to strain in hot climates | This is where Indian designs start reaching for liquid earlier than Western ones |
| >20 kW/rack | ASHRAE TC 9.9 recommends direct liquid cooling | A heat-flux threshold, not a preference |
| 25–40 kW/rack | The practical ceiling for air with good containment | Beyond this the airflow volume and fan power become economically absurd |
| ~41 kW/rack | Frequently cited hard physical limit for air | |
| 45–50 kW/rack | Liquid wins decisively on both thermals and total cost | |
| >50 kW/rack | Direct-to-chip is the default for GPU clusters | |
| ~200 kW/rack | Single-phase immersion territory | Reserved for uniform high-density fleets — it is not a general-purpose answer |
The sequencing point that is usually missed: direct-to-chip removes heat from the processor cold plate and leaves residual heat in the air. Real facilities therefore run hybrid — liquid to the chip, air for everything else. Immersion is the exception rather than the trajectory, because it demands a uniform fleet and a completely different service model. If you describe a high-density design, describing it as hybrid is the accurate answer.
Why the constraint moved to the grid
The genuinely structural change of the mid-2020s is not cooling — it is that siting is now decided by where you can get power connected. Greenfield developments face grid interconnection timelines measured in years, which is why announcements increasingly concern power purchase agreements rather than buildings.
The regulatory consequence is starting to bite in both markets. Virginia approved a new rate class for customers over 25 megawatts, including data centres, taking effect in 2027 — the first serious American attempt to make data centres carry their own grid costs rather than spreading them across ratepayers. In India the movement went the other way: Maharashtra's regulator reclassified data centres from the commercial tariff to the industrial one, which is worth around ₹1.78 per kilowatt-hour on the energy charge, and added a wheeling discount for facilities taking renewable power through open access.
That second point is more consequential than it sounds, and it is covered in the next chapter, because it is the difference between Indian data centre economics working and not working.
Check yourself