MTS

Wiki/Primary documents

LBNL Data Center Energy Reports — 2016 vs 2024, Annotated

A side-by-side reading of Lawrence Berkeley National Laboratory's two Congressionally-mandated U.S. data center energy reports — the 2016 report that famously predicted the sector's electricity demand would stay flat, and the 2024 update that put a tripling on the table within four years.

Why these two reports matter

Almost every grid planner, utility integrated-resource-plan (IRP), and trade-press article on AI's power footprint eventually cites a number from a Lawrence Berkeley National Laboratory (LBNL) data center energy report. There have effectively been only two such reports in the past decade, both produced under Congressional direction:

  • LBNL-1005775, United States Data Center Energy Usage Report, June 2016 (Shehabi, Smith, Sartor, Brown, Herrlin, Koomey, Masanet, Horner, Azevedo, Lintner). This was the response to the 2007 EPA report on data centers and a baseline for federal efficiency policy.
  • 2024 United States Data Center Energy Usage Report (Shehabi, Smith, Hubbard, Newkirk, Lei, Siddik, Holecek, Koomey, Masanet, Sartor), released December 20, 2024 by the U.S. Department of Energy. This is the Energy Act of 2020-mandated update.

The 2024 report is widely quoted ("data centers could be 12% of U.S. electricity by 2028"). The 2016 report is rarely cited because its central conclusion — that demand would stay essentially flat through 2020 — was rendered embarrassing by GPU-driven AI workloads almost as soon as the ink dried.

Reading them together is the most direct way to understand:

  1. How fast the forecasting consensus broke. A team of the same authors, using a refined version of the same bottom-up model, moved from a flat outlook to a 4–7× growth range within eight years.
  2. How much of the AI capex story is genuinely new versus recovery of growth that was always coming. The 2016 report explicitly described an "decoupling" of services from electricity — that decoupling has now broken.
  3. The honest uncertainty in any 2028 number. The 2024 high scenario is ~80% above its low scenario. That spread is itself a finding.

The 2016 report: the world before AI

Headline numbers (LBNL-1005775, June 2016)

Year Electricity use Share of U.S. total
2000 ~28 billion kWh —
2005 ~56 billion kWh —
2010 ~70 billion kWh ~1.8%
2014 (base year) ~70 billion kWh ~1.8%
2020 (forecast) ~73 billion kWh ~2%

Source: LBNL-1005775, executive summary, p. ES-1 (per the OSTI abstract and the eta.lbl.gov landing page).

The decade arc is the whole story: U.S. data centers nearly doubled their electricity demand between 2000 and 2005, grew 24% from 2005 to 2010, then grew only ~4% between 2010 and 2014. The 2016 team projected another ~4% over the next six years.

The verbatim conclusion

The most consequential — and now most archived — sentence in the 2016 report is the deceleration thesis:

"The combination of these efficiency trends has resulted in a relatively steady U.S. data center electricity demand over the past 5 years, with little growth expected for the remainder of this decade." — LBNL-1005775, Executive Summary

And on the upside of further policy:

"There are additional energy efficiency strategies and technologies that could significantly reduce data center electricity use below the approximately 73 billion kWh demand projected in 2020 … an annual saving in 2020 up to 33 billion kWh, representing a 45% reduction in electricity demand when compared to current efficiency trends." — LBNL-1005775, Executive Summary

The companion peer-reviewed paper (Shehabi, Smith, Masanet, Koomey, Environmental Research Letters, 2018, "Data center growth in the United States: decoupling the demand for services from electricity use") made the underlying mechanism explicit: workload demand was growing exponentially, but electricity was flat because of three reinforcing efficiency vectors.

The 2016 assumption set

What did the 2016 team believe would absorb all that compute growth without raising electricity demand?

  1. Hyperscale consolidation. The 2016 model assumed that by ~2020, more than half of U.S. server installed base would sit inside large cloud / hyperscale facilities, where utilization is dramatically higher than internal corporate server rooms. The ERL paper pins this at ~53% of servers in hyperscale by 2021.
  2. Virtualization-driven utilization gains. Service-provider facilities assumed ~25–45% utilization vs. ~10–15% in internal data centers — meaning the same workload needs far fewer physical machines.
  3. PUE compression. PUE (Power Usage Effectiveness, total facility power / IT power) was assumed to keep falling as workload migrated from small server closets (PUE often >2.0) into hyperscale builds (PUE ~1.1–1.2). The 2016 weighted-average PUE assumption was near 1.7 falling toward ~1.5 sector-wide.
  4. Server-side improvements. Performance per watt of x86 servers was assumed to roughly double per Moore's-law era cadence, while server shipments were assumed to grow only modestly.

What the 2016 report did not model:

  • Accelerated computing as a category. GPUs and AI accelerators appear nowhere as a meaningful contributor to server stock. The model was an x86-shaped world.
  • Training-class clusters. Multi-megawatt training clusters were not yet a planning input.
  • Generative-AI inference workloads at consumer scale. ChatGPT was six years in the future.

The "decoupling of services from electricity" was a real empirical observation through 2014 — and a forecast that depended on the absence of a workload category that did not yet matter.

How accurate did the 2020 forecast turn out?

This is more nuanced than the popular story. The 2024 report's own back-cast estimates U.S. data center electricity at:

  • 2014: 58 TWh
  • 2018: 76 TWh
  • 2020: (interpolated from 2024 report time series) roughly 90–100 TWh

So through ~2018, the 2016 forecast was reasonable (within ~5%). It was 2019 onward — as GPU-accelerated server shipments began ramping in earnest — when the curve started to diverge. By 2023, actual consumption was ~176 TWh, more than double the 2020 forecast. The 2016 report did not get 2020 catastrophically wrong; it got the shape of the curve after 2020 completely wrong.

Even the back-cast is contested. Note that the 2016 report's 2014 estimate was 70 billion kWh, while the 2024 report's 2014 estimate is 58 TWh — the 2024 team revised the base year down using better shipment data. Some of the "growth" the 2024 report measures is therefore a recalibration of the 2016 baseline rather than a pure increase.

The 2024 report: the inflection

Headline numbers (2024 LBNL report, released Dec 20, 2024)

Year Electricity use Share of U.S. total
2014 58 TWh ~1.5%
2018 76 TWh ~1.9%
2023 (base year) 176 TWh ~4.4%
2028 low scenario ~325 TWh ~6.7%
2028 high scenario ~580 TWh ~12.0%

Implied power-capacity range at 50% utilization: ~74–132 GW of continuous data-center load by 2028. Sources: DOE press release (Dec 20, 2024); Berkeley Lab newsroom (Jan 15, 2025); reseller summaries citing pp. ES-1 through ES-4 of the report.

The AI breakout

The single most consequential new finding in the 2024 report is the quantification of accelerated servers (GPUs / AI-class chips) as a discrete energy line item:

  • AI-related (accelerated) server electricity, 2017: ~2 TWh
  • AI-related (accelerated) server electricity, 2023: ~40 TWh

That is a ~20× rise in six years, and it accounts for the bulk of the post-2017 inflection in total data-center demand. From the report:

"In 2017, the overall server installed base started growing and Graphic Processing Unit (GPU)-accelerated servers for artificial intelligence (AI) became a significant enough portion of the data center server stock that total data center electricity use began to increase." — 2024 LBNL report

In growth-rate terms:

  • 2014–2018 CAGR: ~7%
  • 2018–2023 CAGR: ~18%
  • 2023–2028 projected CAGR: ~13–27%

The 18% CAGR through 2023 is the post-decoupling regime. Whatever the 2028 number turns out to be, that regime is now the data, not a forecast.

PUE assumptions and trajectory

The 2024 report explicitly moves away from a single sector-wide PUE number to a distribution by space type and cooling system:

  • Hyperscale, indirect evaporative cooled: PUE clusters around 1.10–1.20.
  • Service-provider colocation: typically 1.30–1.50.
  • Internal / enterprise rooms and closets: remain 1.8–2.5.
  • Aggregate 2023 U.S. weighted PUE: approximately 1.4–1.5 depending on space-type weighting, falling slowly toward ~1.3 by 2028 in the model.

The key methodological shift: the 2024 team models PUE and WUE jointly, recognizing that lower PUE can be bought with higher water use (evaporative cooling) or vice versa.

"Key updates include extensive quantification and characterization of accelerated servers used for AI applications, consideration of various cooling system types and outdoor temperatures when modeling PUE and WUE, calculation of carbon and water footprint of electricity consumed based on local grid mixes, and an estimate of electricity demand from cryptocurrency mining." — 2024 LBNL report, abstract / executive summary

Water use (new chapter)

The 2024 report is the first LBNL data center report to systematically quantify on-site water consumption:

  • 2023 direct on-site water use: ~17 billion gallons (~64 billion liters) for cooling.
  • 2028 projection: could double or quadruple depending on cooling-mix decisions.
  • WUE (Water Usage Effectiveness, liters per kWh of IT load) modeled as a function of cooling system × climate zone, then weighted by the geographic distribution of capacity.

The water chapter also separates direct (on-site evaporation and blowdown) from indirect (upstream water embedded in the electricity mix), and concludes that for most U.S. data centers indirect water use is the larger number.

Methodology

The 2024 report uses three modeling approaches in parallel:

  1. Bottom-up: server / storage / network shipments × installed-base survival curves × per-unit power draw, summed across space types. This is the spine of the model and the direct lineage from the 2016 report.
  2. Top-down: utility / EIA data and large-operator disclosures, used as a constraint and cross-check.
  3. Extrapolation: statistical extension of the bottom-up trajectory under different assumptions about shipment growth, utilization, and accelerator share.

Key 2024 modeling refinements not present in 2016:

  • GPU thermal modeling. Maximum operational power is modeled at ~80% of nameplate TDP rather than = TDP, after measurement work on H100-class servers.
  • Utilization uncertainty band. Server utilization is modeled with a central estimate of ~70% and an uncertainty range of 30–80%, which the team explicitly flags as the dominant driver of the 2028 spread.
  • Cooling-system mix. Liquid cooling (direct-to-chip and immersion) is modeled as a growing share for AI deployments.
  • Cryptocurrency mining is quantified separately so it does not contaminate the AI-attributable line.

The uncertainty bands

This is what makes the 2024 forecast unusual for a federal report: the high scenario is ~78% above the low scenario (580 vs. 325 TWh). On a base of 176 TWh in 2023, that is the difference between +85% growth and +230% growth in five years.

The Mobius Risk Group summary notes that the report's range corresponds to an "additional 149–404 TWh" of electricity by 2028 — i.e., the spread of new demand (~255 TWh) is itself larger than total U.S. data-center demand was in 2018.

The 2024 team is also candid that even the high scenario may not capture the right tail. From Shehabi's public remarks around the release:

"We don't yet know how much AI will increase electricity demand, but we do know it will increase quickly."

And from the report itself, on what is not captured:

The model cannot fully account for software-side efficiency innovations (model distillation, sparsity, mixture-of-experts inference), the enterprise shift toward smaller domain-specific models, or the possibility that frontier-training compute pauses if economics change.

Side-by-side numerical comparison

Metric 2016 report (LBNL-1005775) 2024 report
Authors (lead) Shehabi, Smith, Sartor et al. Shehabi, Smith, Hubbard, Newkirk, Lei, Siddik, Holecek, Koomey, Masanet, Sartor
Base year 2014 2023
Base-year electricity ~70 billion kWh 176 TWh
Base-year share of U.S. electricity ~1.8% ~4.4%
Forecast horizon 2020 (6 yrs out) 2028 (5 yrs out)
Central forecast ~73 billion kWh mid-point ~450 TWh
Forecast share of U.S. electricity ~2% 6.7–12.0%
Forecast scenarios Single central + efficiency upside Low / Mid / High band
2014 estimate (per-report) 70 billion kWh 58 TWh (revised down)
AI / accelerated server treatment Not modeled Discrete line: 2 TWh (2017) → 40 TWh (2023)
Aggregate PUE assumption ~1.7 trending to ~1.5 ~1.4–1.5, modeled by space type + climate
Water Not quantified 17B gallons direct in 2023; 2× to 4× by 2028
Cryptocurrency Not separated Quantified separately
Power capacity implication ~10 GW 74–132 GW (50% utilization)
Headline conclusion Growth "relatively steady … little growth expected" "Could double or triple by 2028"

The most striking row is the forecast share of U.S. electricity: from ~2% to as much as 12%. That single change in a federal modeling product reframes how every utility integrated-resource plan, every PJM / ERCOT / MISO load forecast, and every state siting policy must approach the rest of the decade.

What the comparison teaches about forecasting AI infrastructure

1. Bottom-up models are blind to category breaks

The 2016 model was not a bad piece of work. Its forecast tracked actuals through 2018 within a few percent. It failed because the 2014–2018 server-shipment data it extrapolated did not contain meaningful GPU-accelerated server volume. A model that does not have a SKU cannot project that SKU's demand. The lesson is not that LBNL was wrong but that any bottom-up shipment-based model has a fundamental category-introduction blind spot.

The 2024 report partially fixes this by carving accelerated servers into a distinct category — but it inherits the same risk for whatever category emerges next (custom ASIC training clusters, on-device inference, edge AI, post-transformer architectures).

2. The "decoupling" thesis was conditional, not permanent

The 2018 ERL paper said the decoupling depended on continued efficiency gains at hyperscale. The authors warned even then:

"efficiency measures of the past may not be enough" — Shehabi et al., Environmental Research Letters, 2018

That hedge was real, and it has played out. Hyperscale PUE is already near a thermodynamic floor (~1.08–1.10). Server-side performance/watt gains continue but cannot offset 10× more compute per workload.

3. The inflection was visible in shipments before it was visible in totals

The 2024 report dates the inflection to 2017 — six years before the 2024 release and exactly one year after the 2016 report was published. NVIDIA DGX-1 launched April 2016; ChatGPT was six years away. The LBNL team's back-cast suggests an attentive shipment-data reader could have caught the turn by 2018–2019. That nobody updated the federal forecast until 2024 is a policy lag, not a modeling failure.

4. The uncertainty is now the policy variable

A forecast range of 325–580 TWh is not a forecast in the classical sense — it is a statement that the next five years will be determined by adoption decisions that have not yet been made. Utilities and regulators reading the 2024 report are effectively being told that the answer depends on:

  • How fast hyperscalers actually build out announced capacity (vs. announce-and-defer);
  • What share of inference moves to smaller / on-device models;
  • Whether liquid cooling deployment hits scale before the next training generation;
  • Whether crypto demand stays bounded (modeled separately) or rebounds.

This is unusually candid for a federal report and worth quoting executives back to themselves when they cite "the LBNL number" as if it were a single point estimate.

Specific quotes

From the 2016 report

"In 2014, data centers in the U.S. consumed an estimated 70 billion kWh, representing about 1.8% of total U.S. electricity consumption." — LBNL-1005775, Executive Summary

"The combination of these efficiency trends has resulted in a relatively steady U.S. data center electricity demand over the past 5 years, with little growth expected for the remainder of this decade." — LBNL-1005775, Executive Summary

"U.S. data centers are projected to consume approximately 73 billion kWh in 2020." — LBNL-1005775, Executive Summary

"There are additional energy efficiency strategies and technologies that could significantly reduce data center electricity use below the approximately 73 billion kWh demand projected in 2020 … an annual saving in 2020 up to 33 billion kWh, representing a 45% reduction in electricity demand when compared to current efficiency trends." — LBNL-1005775, Executive Summary

From the 2018 ERL companion paper

"efficiency measures of the past may not be enough" — Shehabi, Smith, Masanet, Koomey, "Data center growth in the United States: decoupling the demand for services from electricity use," Environmental Research Letters, 2018

From the 2024 report

"In 2017, the overall server installed base started growing and Graphic Processing Unit (GPU)-accelerated servers for artificial intelligence (AI) became a significant enough portion of the data center server stock that total data center electricity use began to increase." — 2024 LBNL report

"Key updates include extensive quantification and characterization of accelerated servers used for AI applications, consideration of various cooling system types and outdoor temperatures when modeling PUE and WUE, calculation of carbon and water footprint of electricity consumed based on local grid mixes, and an estimate of electricity demand from cryptocurrency mining." — 2024 LBNL report, executive summary

"Total data center energy estimates range from roughly 325 to 580 TWh in 2028, corresponding to 6.7% to 12% of total U.S. electricity consumption." — 2024 LBNL report, executive summary (paraphrased in DOE press release Dec 20, 2024)

Author quotes around the 2024 release

"By showing what the energy use is and, more importantly, what's causing the growth in energy use, it helps us think about what opportunities there are for efficiencies." — Arman Shehabi, Berkeley Lab newsroom, Jan 15, 2025

"We don't yet know how much AI will increase electricity demand, but we do know it will increase quickly." — Arman Shehabi, public remarks around the release

Electricity shortfalls represent "more of a local problem rather than a national problem." — Sarah Smith, public remarks around the release

Open questions about the 2024 forecast

The 2024 report is the most carefully constructed federal data-center forecast in a decade — and it still leaves several first-order questions open. Anyone using its numbers in a model or policy memo should be tracking these:

  1. Is the ~80% high-vs-low spread itself an underestimate? Several private-sector forecasts (Goldman Sachs, McKinsey, Wood Mackenzie, the IEA's Energy and AI report) cluster around the LBNL mid-to-high range but with their own upside cases pushing toward 700–1,000 TWh by 2030. None of those models, including LBNL's, can credibly bound the tail set by frontier-training-cluster announcements.

  2. What share of "announced" hyperscale capacity actually gets built on schedule? The 2024 model assumes announced builds translate to deployed load. As of 2025–2026, the gap between announced gigawatts and energized gigawatts is widening — interconnection queues, transformer lead times, and gas-turbine backlogs are stretching project timelines by 24–48 months.

  3. What happens to the curve if frontier-model training plateaus? The 2024 model implicitly assumes training compute keeps scaling. If scaling laws yield diminishing returns at a known point in 2026–2028, the high scenario is unreachable. Conversely, if test-time-compute (inference-heavy reasoning) scales, the high scenario is conservative.

  4. PUE floor. The model lets aggregate PUE drift toward ~1.3. But for AI-heavy buildings, liquid cooling can push PUE below 1.1 — at the cost of higher water use or higher capital cost. The PUE assumption is doing a lot of work in the low scenario.

  5. The base-year mismatch. The 2024 report revised 2014 down from 70 to 58 TWh. The next LBNL report will likely revise 2023's 176 TWh further. The 4.4% share of U.S. electricity is itself a model output, not a measurement.

  6. Water is the underestimated story. A 2–4× increase in direct on-site water use (from 17 billion gallons in 2023) lands disproportionately on a few drought-exposed counties in Arizona, Texas, Virginia, and Oregon. The aggregate national number obscures siting-level conflict that is likely to constrain deployment more than the electricity number does.

  7. The Energy Act of 2020 cadence. Congress mandated this report; nothing mandates the next one. If the LBNL series remains the most credible federal baseline, the next update (2028 or 2030?) will be the document against which the 2024 high and low scenarios are graded. Worth bookmarking.

Sources