Nuclear arsenals: the 83% decline is real, finished, and mostly a filing cabinet

Status: drafted 2026-08-13 Reachability swept first (per the standing rule from the maternal-mortality run): nuclear-warhead-stockpiles fetches clean; nuclear-weapons-tests 404s; nuclear-warhead-inventories exists but is a 2026 cross-section, not a time series — useful for diagnosis, unusable as a Datoid series. Vintage checked first: FAS 2026 release, OWID lastUpdated 2026-05-13, nextUpdate 2027-05-13, data through 2026. Current.

Segment: The nuclear decline is over, and most of it was a dismantlement queue

• Source theme(s): Batch 24 #1 (the 80% decline). Cross-refs: Batch 24 #2 (the nine arsenals diverging), Batch 24 #3 (arms-control coverage collapsing).
• Spine role: main story
• The framing people carry: Two incompatible ones, held simultaneously. The optimist's: the Cold War ended, the arsenals were slashed, nuclear war is a 20th-century problem. The pessimist's: we are in a new arms race and it has never been more dangerous. Neither camp usually knows the actual number.
• Hypothesis (from the bank): "Total warheads peaking ~1986 at ~70,300, falling steeply through the 1990s-2000s post-START, flattening ~2017 at ~12,000-13,000; the decline almost entirely US+Russia drawing down Cold War surpluses; the curve now bottoming out as modernisation replaces reduction."
• Verdict: CONFIRMED — every clause of it, including the flattening date. And the checking turned up a mechanism the bank did not state, which is the better story: the post-2000 half of the "decline" was overwhelmingly the dismantlement backlog draining, not weapons leaving service. The queue has now almost run out, and the usable stockpile bottomed in 2017 and has risen in eight of the nine years since.

The numbers (every row sourced — no exceptions)

Primary source: Federation of American Scientists, Estimated Global Nuclear Warhead Inventories (2026 release), via Our World in Data grapher nuclear-warhead-stockpiles, accessed 2026-08-13. Tier A — ingested by the existing OWID connector. Treaty facts: FAS, The Aftermath: The Expiration of New START — Tier B.

A measurement asymmetry that must not be charted away. OWID's own descriptionKey states it plainly: "Retired warheads are only included in the global total." The World row is total inventory (military stockpile + warheads retired and queued for dismantlement); the country rows are stockpiles only. So World ≠ the sum of its countries, by construction, and a panel implying otherwise would be wrong. This is why the dashboard puts World on its own panel. Independently confirmed against the 2026 cross-section grapher nuclear-warhead-inventories, whose retired column sums to 2,442 — exactly the World-minus-countries gap.

The texture (what narrative flattens)

• The lead finding: two true lines pointing opposite ways. Total inventory is down 82.7% from its 1986 peak. The usable military stockpile bottomed at 9,272 in 2017 and is 9,745 today. Both are the same dataset. The headline number everyone quotes is falling because a filing cabinet is emptying, not because weapons are leaving service.
• The mechanism, which is the actual story. The gap between the two lines is the dismantlement backlog: warheads retired from service but not yet taken apart. It peaked at 12,518 in 2007 — larger than the entire world stockpile today — and is now 2,442. Between 2007 and 2026, global inventory fell by about 9,000; roughly 10,000 of that movement was the queue draining. The "decline" of the last two decades has been, in large part, an accounting catch-up on decisions taken in the 1990s. It is nearly finished, so the famous falling line is about to stop falling for arithmetic reasons alone.
• Concentration, then dispersal, then re-concentration. In 1986 the US and USSR held 90.2% of everything. The share fell as they drew down. It is 66.5% today — but the direction of travel has reversed: the two largest arsenals are flat-to-rising while seven others grow.
• China is the only arsenal changing fast. 350 (2020) → 620 (2026), +77% in six years — the steepest expansion in the dataset, and the reason the aggregate has stopped falling. Everyone else's growth is in the tens.
• The legal scaffolding went first. The arsenals stopped shrinking in 2017; New START expired on 5 February 2026 with no successor. For the first time since 1972 there is no negotiated ceiling on the two largest arsenals — and no verification regime. The constraint disappeared after the decline had already stopped, not before.
• One country built them and gave them up. South Africa went 0 → 6 by 1989 and back to 0 by 1990. It is the only entity in the dataset that ever did. It is also invisible on any chart scaled to the others, which is a reason to give it its own panel rather than a reason to leave it out.
• Japan-style reputational inversion, for the pessimists: the "new arms race" framing is directionally right about stockpiles and badly wrong about magnitude. Today's entire world stockpile is smaller than the USSR's dismantlement queue was in the mid-1990s.

Datoid artifact

• No existing dashboard touches defence or security — the shelf is macro, climate, development, health and media. This is a genuinely new domain.
• Tier A with zero new connector work. Every series is nuclear-warhead-stockpiles@<entity> on the existing OWID connector, using the @entity addressing added on 2026-08-10 for the maternal-mortality births series. Eleven entities available (9 nuclear powers + South Africa + World); all annual, 1945–2026, one axis (count).
• Artifact: the nuclear-arsenals dashboard, 4 panels:
  1. World inventory 1945→2026 — the 82.7% fall, on its own panel because it is a different measure from the country rows.
  2. The two that did the reducing — US vs Russia, 1945→2026.
  3. The seven that are adding — China / India / Pakistan / North Korea / UK / France / Israel, 1990→2026.
  4. South Africa, 1975→1995 — built six, dismantled six.
• What is deliberately NOT charted: the global military-stockpile line and the retired-queue line. Both are sums/differences across country rows, i.e. computed series, and Datoid has no mechanism for those (the same call taken against deriving births from population × crude birth rate in the maternal-mortality run). Their figures are quoted in commentary with FAS cited, since FAS publishes both directly — but they do not become fake series.

Story justification (Stage 1 rule 5)

• Does the argument have a reveal that needs sequencing? NO → skip.
• Because: the punchline is a second line — global stockpile against global inventory — and that line is not on the platform and cannot honestly be put there (it is a computed aggregate). A scrollystory would therefore sequence prose over charts that do not themselves carry the turn, which is a transcript with pictures: precisely the failure mode the gate exists to prevent. The four dashboard panels plus commentary carry the composition argument in one view, and the dashboard is the landing surface the episode needs. Revisit if a mechanism for derived series ever lands — this theme is the strongest argument for one that the project has hit so far.

Open questions / where I'm guessing

These must NOT become spoken assertions.

1. FAS's own page and FAS's own dataset disagree slightly. The prose on Status of World Nuclear Forces says "an estimated 12,321 at the beginning of 2026"; the dataset behind the chart says 12,187. A ~1% discrepancy, probably a stale sentence or a different cut-off. Use the dataset figure and say "about 12,000" out loud — never present the precise number as authoritative to four digits.
2. The stockpile trough year (2017) is derived, not published. It is the sum of nine country rows, each of which is itself an estimate. FAS asserts the direction ("increasing again") independently, which is the claim worth making. Do not assert "2017 was the exact minimum" — say "bottomed out around 2017".
3. Every one of these numbers is an estimate of a state secret. Only the US, UK and France have ever published stockpile figures, and not recently. Israel has never acknowledged its arsenal at all; North Korea's is inferred from fissile-material production. The whole dataset is a well-sourced guess, and the episode must say so rather than implying a warhead census exists.
4. The retired-queue decomposition is my arithmetic, not FAS's series. The 2026 endpoint checks out exactly (2,442 both ways), which is strong validation, but the historical queue path is World-minus-sum year by year and inherits every country estimate's error. Say "most of it" and "roughly", never a precise share of the decline.
5. I have not resolved what happens to New START's limits in practice. Russia offered in September 2025 to observe the central limits for a year past expiry without verification; reporting on 5 February 2026 described this as unresolved. Do not state that the limits are being observed, or that they are not. State the treaty lapsed and no successor exists.
6. "Modernisation replaces reduction" is the bank's causal claim and remains unsourced here. Programme spending is a different dataset (SIPRI, ICAN) that I have not pulled. The quantity claims stand on their own; leave the causal one out or attribute it explicitly as an interpretation.
7. Warheads are not equivalent to each other. OWID's chart note says so directly: yields vary by orders of magnitude, and the modern stockpile skews very differently from the 1986 one. A count is not a measure of destructive capacity, and the episode should not let "83% fewer" imply "83% less dangerous".