Deep dive 3 of 3 · Defense

Defense: Constant Capability or Constant Force?

Spending Decomposition 2000→2025, a deep dive on one line: defense discretionary spending. Written 2026-09-03; sections 3, 6 to 9 revised and section 10 added the same evening after the defense benchmark study. Figures are percentage points of CBO potential GDP unless stated, the research baseline; the website's default basis (actual spending over actual GDP) reads about 0.01 to 0.03 lower on each level and the same stance to a hundredth. Revised 2026-09-03 after Jason's decisions on the defense benchmark study (war and emergency funding is a memo line) and again on 2026-09-06, when the capability wedge was set at 0.8 percent a year and sections 3, 6, 8 and 10 were rewritten to present that figure as a judgment rather than as a finding of CBO's. Dollar figures are as published by the source named. Years are federal fiscal years, October to September.

1. Why defense needs its own benchmark

For a civilian program "the same services" has a natural meaning: the same real services per person served, at what the inputs cost, less whatever productivity the government can be expected to capture. The exhibit builds every nondefense category that way, with a category-specific cost index and a workload scalar. Defense has no caseload, and it has a question the others do not: is "the same defense" the same machine, or the same position relative to the people the machine is for? A 2000 carrier group in 2025 is a weaker instrument than it was, because the adversaries it faces spent the intervening years building. Jason's phrasing of the benchmark, "comparable defense capabilities as foes," is the second reading; the constant-force benchmark is the first. Both are defensible. The gap between them is not noise to be split; it is a finding, and the exhibit reports it as one.

2. The constant-force benchmark

The Department of Defense publishes its own composite price indexes for each appropriation title, the "Green Book" deflators, and its spending by title. The constant-force counterfactual takes 2000 spending in each title, grows it at that title's own deflator, and sums. Military personnel, operation and maintenance (including the Defense Health Program) and procurement are priced by their own deflators; the remainder, research and development, construction, housing and revolving funds, by the total-DoD index. Retired pay is excluded from the shares because it moved to accrual accounting in 1985. Workload is one, and there is no productivity credit, because a constant force is defined as the same people and the same equipment.

DoD's own prices rose 93 percent over the window, against 77 percent for the economy-wide GDP price index, so the constant-force benchmark already allows defense inputs to get dearer faster than everything else. Potential GDP rose about 3.0 times in nominal terms. The 2000 force at 2000 prices, 2.97 percent of GDP, grown by 1.926 and divided by the growth of GDP, is 1.89 percent of GDP in 2025. Actual spending net of war and emergency funding was 2.84, so the force benchmark says the government bought 0.95 points of GDP more defense than in 2000.

Two things are inside the DoD deflators that a purist would take out. Military pay rose faster than the employment cost index from 2000 to 2010 by act of Congress, and the deflator carries that as a price; the exhibit treats it as the cost of the force as operated, which understates stance slightly. And the procurement deflator prices a stable design: RAND's work on ship costs finds that a ship of unchanging design escalates at roughly the consumer price index, and that most of the above-inflation growth in weapons costs is capability growth, more sensors, more stealth, more range. That is exactly the growth the constant-force benchmark excludes and the capability benchmark includes.

3. The constant-capability benchmark

The number this benchmark rests on is a judgment, and the page says so before it says anything else. There is no measurement of what it costs to hold defense capability constant, and there cannot be one: capability is not a quantity with a price index, the way a ton of asphalt or a hospital day is. The exhibit needs a figure, that figure moves the answer more than any other single parameter on the discretionary side, and so what follows is where it came from and how far it could reasonably be pushed either way — not a derivation.

The one usable anchor. CBO has analyzed the Department's Future Years Defense Program for two decades and finds the same thing each time: holding the planned force constant costs more in real terms every year. The reasons are structural. Military pay and benefits grow faster than civilian pay to keep recruiting competitive; military health care costs grow faster than the economy; operation and maintenance per service member rises as equipment ages and gets more complex; and each generation of weapons costs more than the one it replaces, because staying comparable to adversaries means buying the newer generation. In the 2025 edition the growth is 1.04 percent a year for military personnel, 1.16 for operation and maintenance, 0.98 for weapons and 2.3 for construction, against the GDP price index.

But note what CBO is costing: a plan. The projection holds force structure at the size the five-year program ends at, and it keeps that program's planned modernization — which is a policy choice about how fast to recapitalize, riding inside the growth rate. The consequence is visible straight down the series. Across the twelve editions reached for the study of 2026-09-03, the sequester-era plans of 2012 and 2016, which held force size and modernization down, cost 0.5 to 0.6 percent a year to carry forward; the five editions since 2019 run 0.87 to 1.2. Same method, same agency, different program. That spread is not measurement error — it is the ambition of the plan being costed, and it is about as large as the parameter itself. A benchmark is meant to be the fixed thing policy is measured against, so a benchmark that moves with policy is doing a little of the job it was hired to prevent.

Pulling the other way: part of CBO's figure is already counted below it. CBO measures against the GDP price index. The constant-force benchmark underneath is priced at the Department's own deflators, which rose 0.34 percent a year faster than the GDP price index over the window and already carry the military and civilian pay growth that is about 40 percent of what CBO projects. The military-personnel deflator rose ×2.035 over 2000 to 2025; the employment cost index rose ×2.036. The Department's pay deflator simply is the employment cost index, so stacking CBO's full rate on top of it counts pay twice. That much is mechanical and can be read off the two indexes. Netting out the pieces the deflators already price — military and civilian pay, construction — and keeping health, operations per member and acquisition gives 0.58 to 0.63 percent a year, the range turning on whether military health cost growth is a price or a capability. That arithmetic indicates a size; it does not settle a number, since it applies 2000 weights to a 2029-to-2039 component split and inherits the same plan-dependence as everything above it.

So the wedge is set at 0.8 percent a year, chosen rather than computed (Jason, 2026-09-06). Round on purpose: a figure like 0.63 would advertise a precision that does not exist. Below CBO's roughly 1.05, because the pay double-count is real and visible. Above the netting arithmetic's 0.6, because that arithmetic is soft and because the recent, more expensive plans are the ones describing the modernization the country is actually trying to buy. And on the high side of the plausible zone deliberately, because high is the conservative direction here: a bigger wedge means a bigger capability bucket and a smaller measured stance, so an error in this parameter costs the exhibit's headline finding rather than subsidizing it.

Which way each known consideration points, so a reader can move the number instead of taking it on trust:

Compounded over twenty-five years the 0.8 wedge is a factor of 1.22. Applied to the constant-force counterfactual it gives 2.31 percent of GDP, and stance of +0.53 rather than +0.95. The 0.42 between them is booked to a conditions bucket of its own, "defense capability," rather than folded into income and economy, so the exhibit never hides the parameter that drives it. (At CBO's gross 1.05 the counterfactual is 2.46 and stance +0.38; at 0.65, the netting arithmetic's figure and the exhibit's setting from 2026-09-03 to 09-06, it is 2.22 and +0.61.)

Income and economy is the same under both, −1.08: it is the drift of a real-constant force against a growing economy, the same term that makes a price-indexed benefit shrink as a share of GDP. The only thing the benchmark changes is whether the capability wedge is the government's doing or the world's.

4. The diagnostic that does not depend on either benchmark

Real budget authority per active-duty service member, at DoD's own prices in 2025 dollars, went from about $263,000 in 1980 to $376,000 in 2000 to $621,000 in 2025, a rise of 65 percent since 2000 on a force that shrank from 1.45 million to 1.37 million. Whatever one calls it, the United States spends two thirds more per person in uniform, in the Department's own prices, than it did in 2000. The capability benchmark says about half of that is the price of a constant position; the force benchmark says all of it is a choice. The number itself is neither.

5. What the discretionary deflator dial does to this line

The defense benchmark is a dial of its own, but it lives inside the exhibit's own constant-real-services deflator. The three simpler deflators override it and give defense the same treatment as every other category.

Every row is measured against the 2.84 spent net of war and emergency funding, so the nominal-GDP row now reads a small cut: the share as appropriated was flat, the share net of the wars fell. The current-services row is the one to notice. The statutory baseline that both budget agencies use grows defense by the economy-wide price indexes, which run below DoD's own, so it reads an even larger stance than the constant-force benchmark. That is the sense in which "current services" is not neutral for defense: it assumes the Department's inputs get cheaper relative to their own history. The inflation-and-population row is the Center on Budget's benchmark applied to defense, and it lands between the two DoD-priced benchmarks only because population growth happens to offset the gap between consumer prices and DoD prices.

6. Other windows

The two benchmarks bracket the answer differently depending on where the window starts. Rerun on 2026-09-06 at the 0.8 wedge with defense net of war in every year: from a 1980 base, the Cold War force, today's budget is above the constant-force cost by 1.05 and above the constant-capability cost by 0.27; from 1992, the post-Cold-War force, it is above the force cost (+0.49) and below the capability cost (−0.22); from 2011, the caps era, it is above both (+0.36 and +0.07), because the 2011 base is itself net of a full point of GDP of war funding. As appropriated, with the wars inside the line, the 2011 window had read as a cut against constant capability; that was the wind-down of the wars, not a retreat in the force. Because the wedge compounds, raising it from 0.65 to 0.8 costs more stance the longer the window: 0.17 from 1980, 0.08 from 2000, 0.06 from 2011.

The 1992 row moved for a second reason unrelated to the wedge: it read 4.64 / 2.37 / +0.47 before the Gulf War memo was added on 2026-09-04, and 4.61 / 2.35 / +0.49 after. Corrected here on 2026-09-06.

7. The sources, one by one, and how each was used

DoD, National Defense Budget Estimates for 2025 (the Green Book), April 2024. Table 5-4, deflators by public-law title, 2025 = 100; Table 6-1, total obligational authority by title in current dollars; Table 7-5, end strength. Parsed directly from the Comptroller's Excel tables, no numbers retyped. The old comptroller.defense.gov addresses refuse non-browser clients; the files are live at comptroller.war.gov. No 2026 edition had appeared when this was built, and none had by 2026-09-03 (the Comptroller's page lists one, but the file returns not-found), so 2024 and 2025 are the request and the annualized continuing resolution, not outturns, and the deflators for those two years embed OMB's projected inflation as of April 2024. Used for the constant-force index, the title weights and the per-member diagnostic; Tables 2-1, 5-11 and 5-12 were added on 2026-09-03 for the war split and the pay-raise check.

CBO, Long-Term Implications of the Future Years Defense Program, twelve editions from the 2011 program to the 2025 program. The finding that a constant planned force costs about 1 percent a year more in real terms, the decomposition into pay, health, operations and weapons, and the statement that the growth is measured against the GDP price index. The 2025 and 2016 reports and the 2011 testimony were read in full from mirror copies; the other editions through CBO's own page summaries, because cbo.gov refuses scripted requests. The 1.25 figure the design note first used was the Department's projection rather than CBO's, which is why the review re-centered it. Used as the anchor for the wedge — the anchor only: the exhibit's 0.8 is a judgment built on CBO's roughly 1 percent, not a figure CBO publishes (section 3).

SIPRI, Military Expenditure Database, 1949 to 2025. Rivals' spending in constant and current dollars, for the peer-relative reading in section 10. CRS Insights IN12274 and IN12425 for the 2024 and 2025 emergency appropriations, and IF10260 and RL33446 for the pay raises against the employment cost index.

RAND, cost escalation of ships and weapons. The finding that a stable design escalates at about the consumer price index and that most above-inflation growth is capability. Used to interpret what the procurement deflator does and does not contain.

OMB, Public Budget Database and Historical Tables. The defense category itself, outlays by subfunction, and the split of grants from direct spending that the nondefense categories use. Used for the line's actual values.

The war and emergency lumps. The exhibit's 2025 endpoint includes the Ukraine, Israel and Indo-Pacific supplemental of April 2024 and the war funding of earlier years, because emergency appropriations are appropriations. A separate series was built on 2026-09-03 from the Green Book's split of budget authority into base, war and supplemental funding and its spend-out rates: war and emergency outlays peaked at a full point of GDP in 2008 to 2011 and are 0.13 of GDP in 2025, so the endpoint reads 2.84 without them. Jason made them a memo line the same night, so the defense line is now net of them in every year and the memo carries them; at the wars' peak that moves a full point of GDP out of the line.

8. What a skeptic would say, and what the answer is

"The wedge is a parameter, and it is worth four tenths of a point of GDP." Yes, and it is worse than that: it is not a measured parameter but a chosen one. It is the single most consequential number on the discretionary side of the exhibit, which is why it is a dial, why its bucket has its own name, and why this document exists. It has moved three times — 1.25, then 1.05 when the provenance turned out to be the Department's rather than CBO's, then 0.65 on the netting arithmetic, now 0.8 as an explicit judgment. Anyone who wants a different figure can set one; section 3 gives the direction and rough size of every consideration we know of, and the number is deliberately on the high side so that the exhibit's headline finding is not the beneficiary of the uncertainty.

"Comparable to foes is not measured; China's spending quadrupled." True, and the capability wedge is a cost-of-a-constant-force concept, not a threat-relative one. A benchmark that scales U.S. spending with rivals' spending was built on 2026-09-03 and turns out not to be a benchmark: the United States spent nine times China and Russia together in 2000 and under twice in 2025, so holding the 2000 ratio would require 13 percent of GDP and read defense stance as −10. Even a constant share of world military spending reads −1.7. The wedge, the cost of keeping a planned force current, is the right instrument for "comparable to foes"; the ratio is a memo line (section 10).

"DoD's deflators embed policy: Congress chose the pay raises." Also true. The military pay raises above the employment cost index from 2000 to 2010 were votes, and treating them as price understates stance by a small amount: the raises ran 4.8 percent above the index across the board over 2001 to 2010 (7.8 percent counting targeted raises), the raises below the index of 2014 to 2016 reversed most of it, and the personnel deflator matches the index over the whole window, so the amount hidden in the force benchmark is 0.01 to 0.03 of GDP. The procurement deflator runs the other way: it prices a stable design below the consumer price index, where RAND finds a stable-design ship escalates slightly above it, so it carries none of the capability wedge and, if anything, understates the cost of the same machine by up to 0.09.

"The last two years are budget estimates." They are, and the 2026 Green Book will replace them with actuals and rebase the deflators; the queue has it. The effect on a twenty-five-year index is small.

"A flat share of GDP is the neutral benchmark; you have manufactured a policy expansion." A flat share is what the nominal-GDP setting assumes, and it is available. It asserts that the same defense costs a constant share of a growing economy, which is the claim that neither the Department's own prices nor CBO's force-cost analysis supports. The exhibit shows it so that a reader who holds it can see what it implies everywhere else.

9. What would sharpen the line

The 2026 Green Book, for actuals and rebased deflators, when it appears. A defensible way to price a stable-capability weapons program independently of DoD's plans, which is the one thing that would turn the wedge from a judgment into an estimate. The rest of the list this page carried on 2026-09-03 morning, a wedge by title, a peer-relative benchmark, a war series and the other-window runs, was done that evening and is reported above and below.

10. The harder look of 2026-09-03

Larry Summers said the defense line needed to be studied hard, and it was, in a note filed in the research folder (defense benchmark study_2026 09 03_claude.md). Seven questions, one answer that matters.

Files: engine Program-Accounting/Combined/combined.py (CAP_WEDGE_RATE, DODR, disc_cf, the --wedge= override) and Discretionary/engine/engine.py (dod_cost_ratio, DEF_CAPABILITY_WEDGE); evidence Data/defense/ (defense_deflators.csv, defense_toa_by_title.csv, defense_end_strength.csv, defense_sources.md) and Data/defense/study-20260903/ (the study's series and sources); the discretionary design note and v1 results of 2026-09-01; the review memo of 2026-09-02; the defense benchmark study of 2026-09-03.

← Back to the methodology page