How close are we to capacity?
№ 201 · Throughput vs capacity
Definition
MetricActual throughput as a share of demonstrated capacity
Unit% of capacity utilised
Two lines (or area) showing actual throughput against capacity. The gap equals under-utilization. Pull from operations data. Persistent gap = overcapacity; throughput hitting ceiling = expand capacity. Example: 42K capacity, 32K throughput = 76% utilization (good headroom).
Benchmarks
| Bottom 30% | Median | Top 30% |
|---|---|---|
| below 70%, or persistently above 95% with no headroom | 75% to 80% | 85% to 92% sustained, with deliberate headroom held for peak |
US total industry and manufacturing capacity utilisation from the Federal Reserve G.17 release, 2025 to 2026 monthly readings against the 1972 to 2025 long-run average. Warehouse space utilisation from the WERC DC Measures survey. This metric is a band, not a maximisation target, so both tails are failure states. · Direction is shape rather than higher_better, which matters more here than on any other card in the section. Running above roughly 92% for sustained periods converts every demand spike straight into backlog and overtime, and the US manufacturing base has averaged 78.2% over five decades rather than anything close to full. The critical definitional issue is demonstrated versus nameplate capacity. Nameplate assumes no changeover, no micro-stops and no quality loss; typical discrete manufacturing overall equipment effectiveness sits at 55 to 65%, so a ratio computed against nameplate overstates real headroom by a third or more. Cloud and warehouse figures in the category table use the same ratio but are not comparable to each other in level.
| Category | Bottom 30% | Median | Top 30% |
|---|---|---|---|
| US total industry (Federal Reserve definition) | below 74% | 76.1% | above 80% |
| US manufacturing | below 73% | 75.6% to 75.7% | above 78.2% (the long-run average) |
| Warehouse and distribution centre space | below 75% | not separately published in the public summary | 92% or above |
| Kubernetes compute (CPU) | below 5% | 8% to 10% | above 40% after automated rightsizing |
| Plant equipment (OEE) | below 40% | 55% to 60% | 85%+ |
When it looks bad
Throughput pinned against the capacity line for weeks with no visible gap, so every demand spike converts directly into backlog and overtime. The mirror image is equally bad: a wide permanent gap meaning fixed cost is being carried for volume that never arrives.
Throughput at 97 to 99% of capacity for nine consecutive weeks, backlog up four days, overtime at 18% of hours worked and on-time delivery down six points.
What to do about it
- Set the operating band explicitly, for example 80 to 88%, and treat breaches at both ends as exceptions requiring an owner. US manufacturing has run in the mid 70s for most of the last decade against a 78.2% long-run average, so a line sitting at 95% is not efficient, it is fragile (Federal Reserve G.17).
- Recompute the denominator as demonstrated capacity before drawing any conclusion. Nameplate ignores changeover, micro-stops and quality loss, and typical discrete manufacturing OEE of 55 to 65% means nameplate overstates usable capacity by a third or more (Evocon 2024, Lean Production).
- Buy peak capacity variably rather than structurally. Cross-docking and third party overflow protect the ratio without adding permanent fixed cost, which is how best-in-class distribution centres sustain 92% average space utilisation without collapsing at peak (WERC DC Measures 2024).
- Where the constrained resource is cloud compute, rightsize automatically before adding nodes. Measured CPU utilisation across 2,100 plus organisations was 8% in 2025 with 69% CPU overprovisioning, and automated rightsizing halved provisioned CPUs while cutting out-of-memory kills to near zero, so the reliability argument for overprovisioning does not hold (Cast AI, 2026).
Sources
- Federal Reserve Board Total industry capacity utilisation 76.1% in June, 3.3 points below the 1972 to 2025 long-run average. Manufacturing 75.7%, 2.5 points below its long-run average. Mining 87.4%, utilities 69.5%. federalreserve.gov ↗
- Federal Reserve Board Manufacturing capacity utilisation rose 0.4 points in January to 75.6%, 2.6 points below its 1972 to 2025 long-run average, confirming the multi-year band in the mid 70s. federalreserve.gov ↗
- Federal Reserve Board Total industry capacity utilisation fell to 74.5% in 2020, jumped to 78.3%, and stood at an estimated 75.8% in August 2025 against a 1972 to 2024 long-run average of 79.5%. Utilisation revised down about 1.5 points on average across 2021 to 2024. federalreserve.gov ↗
- Warehousing Education and Research Council, reported by Hyster-Yale Average warehouse capacity used was the single most-tracked DC metric in 2024, with best-in-class at 92% or above and unchanged year over year. Peak warehouse capacity used ranked second. scg-lm.s3.amazonaws.com ↗
- Cast AI Average CPU utilisation 8% in 2025, down from 10%. Memory utilisation fell from 23% to 20%. CPU overprovisioning rose from 40% to 69% year over year, memory overprovisioning 79%. cast.ai ↗
- Cast AI Average CPU utilisation across Kubernetes clusters 10%, down from 13% the prior year; memory utilisation 23%. Clusters partially using Spot Instances achieved 59% average compute cost reduction, exclusively Spot 77%. cast.ai ↗
- Evocon Most manufacturing organisations run overall equipment effectiveness closer to 55 to 60%. Most set their high target at 56 to 60% rather than at the 85% world-class level, so the realistic capacity ceiling is far below nameplate. evocon.com ↗
- Vorne, Lean Production 85% OEE is world class for discrete manufacturers, 60% is fairly typical and 40% is common for organisations just starting to measure. Demonstrates the size of the gap between nameplate and demonstrated capacity. leanproduction.com ↗