The Gap Builds Itself

Not wrong in random directions, the way honest error scatters. Wrong in the same direction, every time — the direction that justified the procurement.

1955: the bomber gap. At the Soviet Aviation Day flypast, a formation of M-4 Bison bombers passed the reviewing stand, circled out of sight, and passed again. Western attachés counted each pass. Projections of Soviet long-range bomber strength were revised sharply upward, and the B-52 program was expanded to answer a fleet that had been, in significant part, the same aircraft doing laps.

1957 to 1961: the missile gap. Sputnik in October, the Gaither Report in November, and by 1960 a presidential campaign fought over an American deficit in intercontinental ballistic missiles. The gap was real in the estimates and inverted in reality. When the imagery finally came in, the Soviet Union had a handful of operational ICBMs. The United States had many times that. In February 1961 Robert McNamara told reporters on background that there was no missile gap and set off a political firestorm — not because he was wrong, but because the procurement had already been ordered against a number that no longer existed. The American stockpile went from roughly seven thousand warheads in 1958 to just over thirty-one thousand at its 1967 peak.

1976: Team B. The CIA convened an outside panel, chaired by Richard Pipes, to check whether the National Intelligence Estimates were underestimating Soviet capability and intent. It concluded that they were, across nearly every category, and it produced the single most instructive sentence in the history of threat assessment: that the Soviets likely possessed a non-acoustic anti-submarine warfare system, the evidence for which was that no such system had been detected. Absence of evidence, promoted to evidence. The panel's substantive estimates — Soviet defense spending, Backfire bomber range, missile accuracy — were later shown to be wrong in the same direction as everything before them.

1979 to 1983: the window of vulnerability. Soviet heavy ICBMs, the argument ran, could destroy the American Minuteman force in a first strike, so the United States needed the MX. The Scowcroft Commission was convened in 1983 to confirm the window and instead quietly dissolved it, observing that silo vulnerability is not arsenal vulnerability when bombers and ballistic-missile submarines exist. Fifty MX missiles were deployed anyway.

Four for four. And the argument has arrived again in its usual clothing, this time with China in the numerator: the adversary is expanding, therefore we must expand.

Which is where this piece has to refuse the obvious move, because the obvious move is the exact error it exists to describe. Four inflated estimates invite a pattern-match — they were wrong, this resembles them, therefore this is wrong — and that inference is no better when a dove makes it than when a hawk does. It is also unavailable on the facts. The China number does not share an evidentiary basis with its four predecessors. The bomber gap was a failure of observation: attachés counting passes with their eyes. The missile gap was a failure of observation: no overhead imagery until CORONA. Team B was a failure of inference from thin data, on purpose, by a panel selected to reach a conclusion. The several hundred new missile silos at Yumen, Hami and Ordos were counted off commercial satellite imagery by academics and NGOs with no procurement stake, published openly, and then confirmed rather than generated by the Pentagon — whose own reporting now puts China's operational stockpile past six hundred warheads and projects it beyond a thousand by 2030.

That is the most independently checkable threat estimate in the entire seventy-year sequence. So concede it. The count is probably right. It may well be low.

And it changes nothing, because the error was never in the count. The error is in the step from count to requirement, and that step fails for reasons that have nothing to do with whether the number is accurate.

i · the asymmetry in the ledger

Every expansion argument in the record shares a single accounting move. The adversary's arsenal is counted as capability — what it could do, at maximum, under the worst assumptions about intent. One's own arsenal is counted as a number — a quantity, insufficient, requiring augmentation.

The two sides of the ledger are denominated in different units, which means the trade always looks good at the moment it is booked. That is not anyone's dishonesty. It is a property of the instrument, and it produces the same output regardless of who is holding the pen.

The reason it matters is that deterrence is not a function of quantity. It is a function of survivability. A retaliatory force deters if the adversary believes enough of it will survive a first strike to impose unacceptable cost. Add a thousand warheads to vulnerable, fixed, MIRVed silos and you have not added deterrence — you have added targets, and you have concentrated more value into fewer aimpoints, which is the definition of a tempting first strike.

Thomas Schelling described the mechanism in 1958 and it has not been improved on: the reciprocal fear of surprise attack. If I believe you might strike first, my incentive is to strike first. If you know I believe that, your incentive rises. Neither of us wants a war. Both of us are being moved toward one by the structure of the thing we built for safety.

Accuracy makes it worse than volume does. A larger arsenal that is also more precise reads, from the far side, as a counterforce posture — an arsenal shaped for disarming the other guy rather than punishing him. The rational response to that is not to be deterred. The rational response is to shorten your decision timelines, move toward launch-on-warning, predelegate release authority, and disperse. Each of those steps degrades crisis stability, and none of them is visible in a warhead count.

So the apparent paradox — that building more could leave you deterring less — is not a paradox. Expansion degrades the architecture that gives the weapons their deterrent value. The excess is the problem, and it is the problem specifically because the adversary can see it and must respond to it.

Which disposes of the China argument without disputing a single silo. The two-peer framing does the real work in Washington now: if you must deter Beijing and Moscow simultaneously, the argument goes, you need to match the sum. You do not. Deterrence is not additive, because it is not transacted against a coalition balance sheet — it is transacted separately with each capital, and each capital has to believe only one thing, which is that retaliation would cost more than the prize. Sums are how procurement thinks. Survivability is how deterrence works. They have never been the same quantity, and the confusion between them is the most expensive category error in the history of the state.

ii · the one time the alarm was right

An argument that explains everything explains nothing, so here is the case that should break this one.

The Soviet buildup of the 1970s was real. It was not an artifact of estimation. Between the mid-1960s and 1986 the Soviet stockpile grew from a few thousand warheads to roughly forty thousand. The SS-18 was real, it was heavy, it was MIRVed, and it was aimed where the alarmists said it was aimed. The hawks of that decade got the specifics wrong in almost every particular, but they were not hallucinating a direction.

Concede it fully. Then look at what caused it.

The Soviet buildup was the delayed echo of the American buildup of 1958 to 1967 — the one ordered against a missile gap that ran in the opposite direction from the one advertised. Moscow watched the United States quintuple its arsenal in nine years while holding a handful of ICBMs, drew the obvious conclusion about its own survivability, and spent the next twenty years correcting it. Khrushchev's successors were not going to be caught in that position twice.

Which means the single case in the record where the threat was genuine is the case the previous cycle manufactured. That is not a counterexample. That is the loop closing.

The estimate does not merely fail to describe reality. Acted on, it produces the reality that retroactively vindicates it. Give it fifteen years and the file will show that the hawks of 1957 were right about 1975, and nobody will file the variance report explaining that they were right because they were listened to.

iii · what killed the missile gap

Here is the part that should worry anyone who finds the above merely amusing.

The missile gap died. It did not die because someone made a better argument. It died because on August 18, 1960, a CORONA reconnaissance satellite returned a film capsule containing imagery covering more Soviet territory than every U-2 flight combined, and over the following year the photographs made the estimate untenable. By the autumn of 1961 the official count of Soviet ICBMs had collapsed from hundreds to a few dozen. The correction was possible because verification existed.

That is the mechanism that has broken.

On-site inspections under New START stopped in March 2020 for the pandemic and never resumed. Russia suspended its participation in the treaty's verification regime in February 2023. And on February 5, 2026, New START expired — the last binding limit on the strategic arsenals of the two largest nuclear powers, gone for the first time since 1972.

There was an offer on the table and it was never answered. In September 2025 Putin proposed that both sides go on observing the treaty's central limits informally for a year after expiry, conditional on American reciprocity; Moscow's foreign ministry noted on February 4 that no formal response had come. The American position on the day of the lapse was that it would pursue a new and modernized treaty instead. Six days later Russia's foreign minister told the Duma that Moscow would hold to the old ceilings only for as long as Washington did. That is not a floor. It is two governments each watching to see what the other does, with nothing in writing and nobody authorized to look.

The treaty capped each side at 1,550 deployed strategic warheads. Both sides hold far more than that in reserve — the total American stockpile is estimated around 5,200 warheads, the Russian around 5,500. The gap between deployed and stockpiled is not a rounding error; it is an upload capacity. Neither country needs to manufacture a single new warhead to roughly double its deployed strategic force. It needs a decision and some months of wrench time. The 2023 Strategic Posture Commission has already recommended preparing to do exactly that.

So the setup is complete, and it is worse than 1957 in the one respect that used to save us. In 1961 an estimate could be falsified by a photograph. In 2026 there are no inspectors, no data exchanges, and no treaty.

And "less checkable" is the modest version of the claim. CORONA worked because a silo is concrete. It exists in the world, it takes years to build, it cannot be moved, and no amount of denial survives the film. Upload capacity is not an object at all. It is a decision plus some months of wrench time, applied to warheads that already exist, in bunkers already under observation, which look identical on the day before and the day after. Restore the inspection regime tomorrow — inspectors back at the gate, data exchanges resumed — and you would be verifying an intention rather than a structure. The 1961 rescue is not merely unavailable. It is unavailable in kind.

One further thing about the loop, which is that it has two participants and roughly a hundred and ninety bystanders. The non-nuclear signatories of the Non-Proliferation Treaty gave up the option in exchange for a promise of eventual disarmament, and have now watched the last binding limit on the two largest arsenals lapse with no successor and no account owed to them. Reciprocal fear is a coherent logic for the two states inside it and an imposition on everyone it endangers — which is a fair part of why nothing inside the loop has ever been able to break it.

iv · the prediction

Here is the sequence, and it is not a forecast so much as a replay with the dates left blank.

An estimate will appear showing that Russia or China has uploaded past the old New START ceiling. It will be sourced to means that cannot be independently checked, because the checking apparatus was allowed to lapse. It will be met by an argument that the United States must match, and the upload will be authorized — quickly, because the hedge warheads are already built and the delivery systems are already there.

The adversary will observe the American upload and respond in kind, because from where it sits the American move is a counterforce posture and the response is survivability, not aggression.

And in roughly a decade the published numbers will confirm the original estimate. The file will close. It will read as vindication.

Nobody will note that the estimate was a cause. Nobody ever has. That is the only genuinely reliable finding in seventy years of this: the loop has never once been broken by being correctly described, only by someone building an instrument that made the claim checkable.

The instrument did not simply expire. Instruments can be rebuilt, and a treaty can be signed again by two governments that decide they would rather. What happened is smaller and considerably worse: the quantity that now needs measuring stopped being the kind of thing that can be photographed, and the treaty's lapse only removed the last reason to pretend otherwise.

There is no CORONA for an intention. Someone would have to invent one, and the first requirement for inventing it is admitting that the count was never the problem.