Chapter 3 · Layer 0 · Facility Efficiency
The Bill Before the First Token
For every unit of power that reaches the computers, the average data centre uses 1.52. Germany now caps new ones at 1.2 by law. Facility efficiency, heat reuse and the bill before the first token.
Sanket Kulkarni · · 6 min read
Part of The Seven-Layer Tokenomics Stack
A law with a number in it
Since 1 July 2026, anyone opening a new data centre in Germany has had to meet two numbers written into law. Under section 11 of Germany's Energy Efficiency Act, the building must reach a power usage effectiveness of 1.2 or lower. And at least 10% of the energy it uses must be reused, mostly as heat for someone else. The reuse share steps up to 15% for centres opening from July 2027 and 20% from July 2028. Older German data centres get a deadline too: 1.5 from July 2027, then 1.3 as a yearly average from July 2030.
Germany's Energy Efficiency Act, section 11
A law with a number in it
July 2026
NewPower usage effectiveness 1.2 or lowerNewAt least 10% of energy reusedJuly 2027
NewReuse rises to 15%ExistingPower usage effectiveness 1.5 or lowerJuly 2028
NewReuse rises to 20%July 2030
ExistingPower usage effectiveness 1.3 or lower
New: data centres opening from that date. Existing: data centres already running.
Power usage effectiveness (PUE) sounds like jargon. It is one simple question: for every unit of electricity that reaches the computers, how much does the whole building use? A score of 1.0 would mean nothing is lost to fans, cooling, lighting or power conversion. Anything above 1.0 is overhead.
The industry is a long way from 1.0. The Uptime Institute, which surveys data centre operators worldwide every year, put the average at 1.52 in its 2026 survey. Its July 2026 release called the gains "gradual", slowed by older buildings. Google, by contrast, reported 1.09 across its global fleet for 2025, measured continuously over the year.
Chapter 2 was about the price of a unit of power. This chapter is about how many units you pay for that never touch a chip.
The service charge on the bill
Think of the final bill at a restaurant. The menu says ₹100 for the meal. Then come the service charge, the packaging and the delivery fee. You pay for all of it, but only the ₹100 was food.
Power usage effectiveness is that final bill divided by the menu price. At the industry average, every ₹100 of electricity for the computers comes with ₹52 of extras. At Google's reported fleet figure, the extras are ₹9. Same chips, same tokens, different bill.
Power usage effectiveness (PUE)
The service charge on every ₹100 of computing
Perfect building 1.00
nothing lost
Google fleet, 2025 1.09
measured over the year
German law, new centres 1.20
from 1 July 2026
Industry average, 2026 1.52
Uptime Institute survey
Sources: Uptime Institute, 2026 survey; Google, 2025 fleet average; Germany's Energy Efficiency Act, section 11.
Where do the extras go? Mostly to cooling. Every unit of electricity a chip uses ends up as heat, and that heat has to be pumped out of the room by fans, chillers and pumps that need power of their own.
The Tokenomics Stack lists three things under Facility Efficiency: power usage effectiveness, cooling overhead, and heat reuse. The third is the one most people forget. In our kitchen, it is the difference between letting the tandoor's heat escape up the chimney and using it to warm the water for washing up.
Finland shows what that looks like at city scale. The City of Espoo says waste heat from Microsoft's planned data centres, captured by the energy company Fortum, will eventually cover about 40% of the city's district heating needs (hot water piped to homes and offices), cutting coal emissions by about 400,000 tonnes a year. Eurelectric, the European electricity industry association, reports that about 75% of the centres' waste heat would go into that network each year, because demand for heat is low in summer.
Heat reuse in Espoo, Finland
The tandoor's heat, piped to the washing-up water
Data centres
waste heat
District heating
homes and offices
~40%
of the city's district heating needs, eventually
~400,000 t
of coal emissions avoided a year
~75%
of the centres' waste heat used each year; summer demand is low
Sources: City of Espoo, 2024; Eurelectric. Microsoft's planned data centres, with the heat captured by the energy company Fortum.
A price lever that multiplies every other price
This component is a price lever. It does not change how many tokens anyone asks for. It changes how much electricity each token drags along with it, and it multiplies whatever you agreed to pay for power in Chapter 2.
My own rough arithmetic: take computers that draw 100 megawatts (MW). At an overhead of 1.52, the building draws 152 megawatts. At 1.09, it draws 109. The gap is 43 megawatts, which is 43,000 units on an Indian electricity bill every hour, or about 377 million units a year. At October 2026's average exchange price of ₹6.66 a unit, that is roughly ₹250 crore a year spent on extras that make no tokens at all. It is an illustration, not a real site.
Illustration: computers drawing 100 megawatts (MW)
What 43 megawatts of overhead costs in a year
Gap
43 MW
152 minus 109
Every year
377 million units
43,000 units an hour × 8,760 hours
At ₹6.66 a unit
~₹250 crore
a year, on power that makes no tokens
My rough arithmetic, not a real site. Price: the Indian Energy Exchange day-ahead average for October 2026.
India is about to make this decision many times over. According to a government reply in the Rajya Sabha, reported on 13 August 2026, India's installed data centre capacity grew from 375 megawatts in 2020 to about 1,575 megawatts. The Central Electricity Authority (CEA) estimates data centre demand of about 17 gigawatts (GW) up to 2031 and 2032. India's Bureau of Indian Standards (BIS) has published national standards for power usage effectiveness and related measures. Every one of those new buildings will lock in its overhead for fifteen or twenty years the day its cooling is designed.
India's data centres, in megawatts (MW)
Overhead gets locked in the day the cooling is designed
Source: Rajya Sabha reply reported by Indian Infrastructure, 13 August 2026; the 2031 and 2032 figure is the Central Electricity Authority (CEA) estimate of about 17 gigawatts.
The framework adds a warning I agree with. Power usage effectiveness and its water equivalent, it says, "say nothing about how productively that energy is used." A building with a superb score full of idle chips is still wasting money. The measure the framework actually wants is energy per token, and eventually energy per useful outcome. Facility efficiency is one factor in that sum, not the whole sum. The idle stove problem belongs to Layer 2.
The decisions this component holds:
- What to build or rent: a cooling design that keeps overhead low as chips run hotter.
- What to measure: a yearly average from meters, not a design-brochure figure on the coldest day.
- What to do with the heat: sell it, give it away, or pay to throw it into the sky.
And the access point matters. Per the framework, if you own the facility, efficiency is your lever. In a shared building you see it only "as the operator discloses it." In the cloud, you see "what the provider discloses."
What I would do on Monday morning
- Ask for the real number. Whoever hosts your AI work, ask for power usage effectiveness as a twelve-month average from meters, the way Google reports its fleet. A single best-day figure is marketing. The international standard for this measure was updated in January 2026, so ask which method they follow.
- Put it inside your unit cost. Multiply your electricity price by the overhead ratio before you compare sites or providers. A cheaper unit in an inefficient building can cost more per token than a dearer unit in an efficient one.
- Plan for heat as a by-product, not waste. A chemical plant that vents its steam is leaving money on the table. If you are building or contracting new capacity, especially in India's coming wave, ask whether the site can sell its heat or is designed to. Germany has already turned that question into law.
Next
Cooling does not only use power. Much of it uses water. In Chapter 4, we look at Water: how much a data centre drinks, where that water comes from, and how much is used for every unit of power.
This series builds on The Seven-Layer Tokenomics Stack by the Tokenomics Foundation, used under a Creative Commons Attribution 4.0 licence (CC BY 4.0). The commentary, and any mistakes, are mine. Views are my own and do not represent my employer.
Sources
Every figure was checked against its source on 6 November 2026.
- Tokenomics Foundation, The Seven-Layer Tokenomics Stack (full paper), October 2026, Creative Commons Attribution 4.0 licence (CC BY 4.0). Facility Efficiency component (power usage effectiveness, cooling overhead, heat reuse); energy per token and per outcome as the measures that matter; "say nothing about how productively that energy is used"; "as the operator discloses it" and "what the provider discloses".
- Federal Ministry of Justice (Germany), Energy Efficiency Act (Energieeffizienzgesetz), section 11, in force as amended. New data centres from 1 July 2026: power usage effectiveness of 1.2 or lower and at least 10% reused energy (15% from 1 July 2027, 20% from 1 July 2028); existing centres 1.5 from 1 July 2027 and 1.3 yearly average from 1 July 2030.
- Uptime Institute Intelligence, The growing PUE advantage of larger data centers (Update 518), 6 August 2026. Annual average power usage effectiveness (PUE) of 1.52 from the 2026 Global Data Center Survey.
- Uptime Institute, 16th Annual 2026 Global Data Center Survey (press release, Business Wire), 28 July 2026. "Power efficiency gains remain gradual"; progress slowed by legacy infrastructure.
- Uptime Institute Intelligence, Uptime Institute Global Data Center Survey 2026 (Keynote Report 209), July 2026. Gains remain gradual; older infrastructure and cooling constraints slow progress.
- Google, Data center efficiency, accessed 6 November 2026. 2025 fleet-wide annual average of 1.09, measured continuously across the fleet as a trailing twelve-month figure including all sources of overhead.
- International Organization for Standardization (ISO), ISO/IEC 30134-2:2026, Data centres key performance indicators, Part 2: Power usage effectiveness, edition 2, January 2026.
- City of Espoo, Using waste heat from a new data center to cover 40% of Espoo's district heating needs, 15 July 2024, updated 26 August 2024. About 40% of district heating needs; about 400,000 tonnes of coal emissions a year; Microsoft and Fortum.
- Eurelectric, Microsoft data centers provide 40% of Fortum's electricity-based district heating in Espoo, Finland, accessed 6 November 2026. About 75% of waste heat used each year; low summer demand.
- Indian Infrastructure, Government promotes sustainable growth of data centres through energy-efficient and clean energy measures, 13 August 2026, reporting a Rajya Sabha reply. Capacity from 375 megawatts (2020) to about 1,575 megawatts; Central Electricity Authority (CEA) estimate of about 17 gigawatts up to 2031 and 2032; Bureau of Indian Standards (BIS) standards on power usage effectiveness.
- Business Standard, Spot power prices jump 149% to ₹6.66 per unit in October on demand surge, 3 November 2026. ₹6.66 per unit used in the illustrative calculation.