Carbon Explorer: A Holistic Framework for Designing Carbon Aware Datacenters

A datacenter can count as fully renewable on a yearly balance sheet and still run on coal power at three in the morning, so Carbon Explorer checks every hour of a real year and prices each fix against the carbon released to build the hardware that fix requires.

Why one hour matters more than one year

Datacenters around the world consumed 205 TWh of electricity in 2018. TWh stands for terawatt-hour, a unit of energy; 205 TWh is more electricity than Ireland or Denmark each consume in a year. Information and communication technology may reach 7% to 20% of all world electricity demand by 2030.

The industry has responded with money. Amazon, Meta and Google together invested in 22 GW of renewable generation. GW stands for gigawatt, a unit of power capacity. Meta alone holds 5,754 MW across thirteen sites in the United States — 3,931 MW of wind and 1,823 MW of solar. MW stands for megawatt, one thousandth of a gigawatt.

The way that money is counted is the source of the problem.

A power purchase agreement is a contract to buy the output of a specific wind farm or solar farm. For every megawatt-hour that farm generates, the buyer receives a renewable energy credit. A megawatt-hour, written MWh, is one megawatt of power held for one hour. At the end of the year the company adds up its credits and adds up the electricity its datacenters consumed. If the credits are at least as large, the company declares Net Zero.

Now look at the units. Credits are summed over a year. Electricity is consumed in an hour. A credit earned at noon in July does not deliver a single watt to a server at three in the morning in January. In that hour the datacenter draws whatever the local grid happens to be burning.

That difference is expensive in carbon. The paper lists how much carbon dioxide each kind of power plant releases per kilowatt-hour of electricity, written as grams of CO2 equivalent per kWh:

An hour powered by natural gas therefore emits about 45 times as much carbon per kilowatt-hour as an hour powered by wind. An hour powered by coal emits about 75 times as much.

So the question is not whether the yearly books balance. The question is what fraction of hours are genuinely covered. The paper measured that fraction under Meta’s existing renewable investments. In the two regions it examined in detail — one served mainly by wind, one served only by solar — the hourly figure was 46% and 51%. Roughly half of the year’s energy arrived in hours when renewable power was not there.

Two kinds of carbon, and why the paper insists on both

This distinction is the heart of the paper. Everything after it depends on getting it right, so it is worth stating slowly.

Operational carbon is released while the datacenter runs. It comes from the power plants that generate the electricity the servers consume. It is measured in grams of CO2 per kilowatt-hour consumed, which is the table above. Turn the servers off and it stops.

Embodied carbon is released before the hardware ever runs. It comes from mining the materials, manufacturing the parts, transporting them and assembling them — and for batteries, also from recycling them at the end of life. It is paid once, up front. It does not fall when the machine sits idle, and it does not fall if you later power the machine with wind.

Here is why the pairing matters. Every method for removing operational carbon requires a new physical object: a wind turbine, a battery, a server. Each object carries embodied carbon. So the real question is never how do I reach zero operational carbon. It is:

Does this object remove more operational carbon over its life than the carbon released to build it?

The paper supplies the numbers needed to answer that.

• A wind farm costs 10 to 15 grams of CO2 for every kilowatt-hour it will generate across its life, which is about 20 years. Compare that with the 820 grams per kilowatt-hour of coal: building the turbine costs between one eightieth and one fiftieth of what burning coal costs, per unit of electricity delivered.
• A solar farm costs 40 to 70 grams of CO2 per kilowatt-hour generated, over a life of 25 to 30 years. That is higher than wind, and still between one twentieth and one twelfth of coal.
• A lithium-ion battery costs 74 to 134 kilograms of CO2 for every kilowatt-hour of storage capacity it holds.
• A server costs 744.5 kilograms of CO2 to build. So one kilowatt-hour of battery capacity costs between one tenth and one fifth of what one server costs to build.

What people did before, and where it broke

Prior work was reasonable, and each strand solved a real problem.

Renewable investment through power purchase agreements is the industry standard, and it works for what it was designed to do. It genuinely lowers the carbon intensity of a datacenter’s electricity across a year. Earlier academic work went further and put generators on the datacenter site itself, with microgrids that switch between local sources. That approach suits small edge sites. Hyperscale operators avoided it by buying generation on the grid instead, at a scale no on-site installation could reach.

Batteries in datacenters are also not new. Operators already install them throughout racks and clusters, so that services keep running while power fails and diesel generators start. Prior research studied battery ageing and using batteries to reduce demand for grid power.

Scheduling flexible work around cheap or clean electricity has been studied for years, with objectives that included electricity price, carbon price, and the carbon intensity of grid power.

Three things broke.

First, the grid got harder, not easier. As more wind and solar farms are built, the swings in supply grow larger. When supply exceeds demand, grid operators must switch renewable generators off. That deliberate shutdown is called curtailment. In California, curtailment reached 6% of all renewable electricity generated in 2021, and has risen steadily since 2015. California’s grid was 33% renewable in 2020, against a United States average of 20% — so the rest of the country is heading toward the same condition.

Second, the variation is on the supply side, and the datacenter cannot absorb it. Datacenter demand is close to flat. Processor utilization at Meta swings by about 20% over a day, and Google’s published Borg traces show a 15% swing on average. But at datacenter scale the difference between maximum and minimum energy demand is only about 4%. Renewable supply is far less steady. In the California grid, renewable generation varied by more than a factor of three inside a single week. In the Oregon region, the best ten days of the year deliver about 2.5 times the renewable energy of an average day, and the worst days deliver almost none.

Third, and this is the gap the paper names, nobody priced the hardware. Prior proposals for demand response treated extra servers as free in carbon terms. The paper states plainly that the trade between operational and embodied carbon for that extra capacity had not been considered by any existing work.

The measure the whole paper turns on

Before the three levers, one definition.

Hourly coverage is the share of the datacenter’s yearly electricity that is supplied by renewable power in the same hour it is consumed. If in some hour the datacenter needs 20 MW and renewables deliver 12 MW, then 8 MW comes from the grid and counts against coverage. Sum that shortfall across all 8,760 hours of the year, divide by the year’s total demand, and subtract from one.

Coverage of 100% is what the United Nations 24/7 Carbon-Free Energy Compact asks for: every kilowatt-hour met by carbon-free sources, every hour of every day. The whole paper is about what reaching that number actually costs.

Lever one — buy more renewable capacity

The first lever is the obvious one. Buy more wind farms and solar farms until supply covers demand in every hour.

Carbon Explorer models this by taking the measured hourly generation of the local grid and scaling it in proportion to the investment. It then matches that hourly supply against the hourly demand.

The result is a long tail. Coverage rises quickly at first and then almost stops. For Meta’s Oregon site, going from 95% to 99.9% coverage takes more than 5 times as much renewable investment as going from 0% to 95% did. Buying the first 95% is cheap. Buying the last 5% is not.

Two properties of the local grid decide how bad the tail is.

• Regions served only by solar cannot pass roughly 50%. No quantity of panels produces power at midnight. This is a hard limit, not a diminishing return.
• Regions with both wind and solar have shorter tails, because the two sources are strong at different times and partly compensate for each other.

Across all thirteen sites, the carbon-optimal design using renewables alone reaches between 37% and 97%

Lever two — install batteries

The second lever stores carbon-free energy instead of generating more of it. Batteries charge in hours when renewable supply exceeds demand, and discharge in hours when it does not.

The paper sizes batteries in hours of datacenter compute, which is easier to read than raw capacity. Two hours for a 20 MW datacenter means 40 MWh of battery.

• Utah, a region with both wind and solar and an average datacenter power of 19 MW, reaches full hourly coverage by adding about 5 hours of battery on top of Meta’s existing renewable investment there.
• North Carolina, a region with solar only, needs 14 hours — almost three times as much, because it must carry the datacenter through every night.

For scale, the largest utility-scale storage project at the time of the paper offered 300 MW of power and 1,200 MWh of capacity. These are not rack batteries.

Carbon Explorer models lithium iron phosphate cells, a chemistry common in large stationary storage, using a published battery model that accounts for efficiency loss and for charge and discharge rates.

The depth-of-discharge trade

A battery’s life is counted in charge-and-discharge cycles, and how many cycles it survives depends on how deeply it is drained each time. Depth of discharge is the fraction of the battery’s capacity you allow yourself to use.

The cells the paper models survive 3,000 cycles at 100% depth of discharge, and 4,500 cycles at 80% — a 50% longer life for a shallower drain.

That looks like a free improvement. It is not. If you may only use 80% of a battery, you must buy a bigger battery to store the same usable energy. In the carbon-optimal configuration, dropping to 80% depth of discharge raises embodied carbon by 43%.

So one number falls and another rises. Carbon Explorer’s job is to decide which wins, and it does: across the regions, 80% depth of discharge lowers total carbon by 5% on average compared with 100%. Tuning this single setting moves total carbon by 3% to 9% depending on the region.

Going shallower keeps helping until it does not. A 60% depth of discharge implies 10,000 daily cycles and a 27-year battery life — long enough that other ageing effects would end the battery first.

Lever three — move the work to the clean hours

The third lever changes the demand instead of the supply. Carbon-aware scheduling delays work that can wait, from hours when grid power is dirty to hours when renewable power is plentiful.

This only works if some work can actually wait. It can.

A service level objective is the deadline a job is promised. User-facing services need a response now. Other work does not: model training, data processing pipelines, offline video processing. At Meta, offline data processing is about 7.5% of the whole fleet’s workload, and about 87.4% of that processing carries a deadline of four hours or more, most of it a full 24 hours. Google reported that jobs with 24-hour deadlines make up about 40% of the jobs its Borg scheduler handles. The paper’s main analysis uses that 40% figure as the share of work that is flexible.

The scheduler itself is deliberately simple. It is greedy. It moves flexible work out of the hours with the highest carbon intensity and into the hours with the lowest. It stops for one of two reasons: all flexible work has moved, or every server in the receiving hour is already busy. It obeys two constraints — the datacenter’s maximum power capacity, and the share of each hour’s work that is allowed to move.

What scheduling costs

Moving work later does not reduce the work. It concentrates it. If a day’s deferred jobs all execute in the six hours when the wind is strong, the datacenter needs enough servers to run them in those six hours.

The paper measures this. If every workload were flexible, reaching full hourly coverage would need 19% to more than 100% additional servers — in the worst region, twice the fleet. In the carbon-optimal designs, where 40% of work is flexible, scheduling needs 6% to 76% additional servers.

Every one of those servers carries 744.5 kilograms of

series 1: datacenter power demand, 8,760 hourly values
series 2: local wind + solar generation, 8,760 hourly values
site:     Utah, mixed wind and solar, average demand 19 MW

Why this works

The causal chain is short, and each link carries the one after it.

Because renewable credits are summed over a year while electricity is consumed in an hour, a datacenter can be Net Zero on paper and still emit carbon in half of its hours. → Because hourly coverage is decided by the worst days rather than the average day, only hour-by-hour data reveals the true cost of covering those hours. → Because the supply valleys are deep, closing them with generation alone needs a very large amount of capacity, much of which produces surplus power the rest of the year. → Because that unused capacity still had to be manufactured, the renewables-only answer carries the highest embodied carbon of any strategy, in every region. → Because a battery moves energy across hours instead of generating more, it closes the same valleys with far less hardware — which is why adding batteries reduces the total carbon footprint by about a factor of ten in every region. → Because deferring flexible work moves demand toward the supply instead of moving supply toward the demand, it closes part of the remaining gap with servers rather than with batteries, and reduces the battery capacity needed. → Therefore the lowest total is reached by a mixture, and the mixture depends on the local grid.

The last link produces the paper’s most surprising finding. Because the tail of coverage is long, and every step along it must be manufactured, there is a point where the next wind farm releases more carbon in its construction than it prevents in operation. Past that point, 100% hourly coverage is not the carbon-optimal design.

What it costs

• The answer is local, so it does not transfer. The carbon-optimal design for Texas is not the carbon-optimal design for North Carolina. Every site needs its own year of grid data and its own search.
• The batteries are large and real. The carbon-optimal designs need 200 to 1,800 MWh of battery capacity per site — utility scale, not rack scale.
• The extra servers are idle most of the time. Scheduling needs 6% to 76% more servers, and their embodied carbon is paid whether or not they are used.
• Deep discharge shortens battery life. The optimizer maximizes battery usage to avoid grid power, so batteries are often either fully charged or fully discharged. That is good for operational carbon and bad for the battery.
• The analysis is offline. It replays a year that already happened. A real deployment would need to forecast supply rather than know it.

What the measurements showed

The study covers thirteen Meta datacenter sites in the United States. Hourly grid generation comes from the Hourly Grid Monitor of the United States Energy Information Administration, for 2020. Hourly demand comes from Meta’s own production measurements. Three sites are shown in detail: Oregon, served mainly by wind, with an average datacenter power of 73 MW; North Carolina, served only by solar, at 51 MW; and Utah, served by a mix, at 19 MW. The main analysis assumes 40% of the work is flexible, matching the reported share of jobs with 24-hour deadlines.

Two findings about where to build follow from the table. Nebraska, served mainly by wind, together with Utah and Texas, which have both wind and solar, are the best of the thirteen locations for minimizing total carbon. Their supply valleys — the days of the year with the least renewable generation — are the shallowest. The solar-only sites, North Carolina, Georgia, Tennessee and Alabama, do not reach full coverage and carry the highest total carbon footprints.

One cost figure changes how to read the battery result. At the battery price the paper cites, 350 dollars per kilowatt-hour, even an 1,800 MWh installation is a small fraction of a hyperscale datacenter’s overall cost, which reaches billions of dollars. The battery is expensive in carbon and cheap in money.