In 1914, every general staff in Europe planned for a short war. In 1928, Giulio Douhet proved on paper that bomber fleets would break societies within days. In 2022, assessments in Washington reportedly gave Kyiv seventy-two hours. Defence forecasting fails so reliably that failure itself is the field’s best-documented finding.
The usual lesson is humility: nobody can know the future of war. It is half right. Prediction fails where it depends on contingency — who fights whom, which invention arrives first. But some statements about the next decade follow from physics that will not change by 2036, from economics no procurement office can repeal, from institutional incentives that survive every election. They can be deduced rather than predicted — and, unlike predictions, deductions show their work.
One rule governs everything below: every claim states what it rests on, how strongly it holds, and what would prove it wrong. Anyone can check the steps; that is the point.
Seven statements
1. Whatever emits will be found, and whatever is found can be killed. Sensors ride civilian cost curves — smartphone cameras, software-defined radios, small satellites — and get cheaper every year; signature reduction is bespoke military engineering and does not. Precision falls in price on the same curves. So the gap between detected and destroyed closes toward minutes, and everything that radiates, moves, or gives off heat lives under permanent risk of engagement. Signature management becomes the first condition of survival. This is not a trend reading but a consequence of cost curves already locked in. What would break it: deception and jamming scaling faster than sensing — a thousand decoys making ten thousand sensors worthless. The battlefield would then turn noisy rather than transparent; the fog of war returns as data garbage, not darkness.
2. Mass returns as industrial mass — and any long war becomes a contest of production rates, not stockpiles. Precision was expensive, so militaries bought few exquisite platforms. Once precision rides civilian components, the calculus flips: many cheap, losable systems beat few costly ones whenever the defender pays more per kill than the attacker pays per unit. Attrition migrates from people to machines, but it stays — and expendable systems empty a peacetime depot in weeks. The measure of military power shifts from what sits in the depot to what leaves the factory each month. The logic holds as long as its premises do. It would fail if cheap defences — guns, jammers, directed energy — flip the cost curve back; then cheap mass was a window, not an era. Or if future wars were reliably short — which nobody can guarantee.
3. Physics keeps the cheap systems tactical; the war gets layered. Battery energy density sits an order of magnitude below liquid fuel, and small airframes trade payload, range, and speed brutally against one another. Cheap small systems therefore dominate the last ten to fifty kilometres; operational and strategic depth remains the province of larger, faster, costlier effectors. There will be no “drone war” but a layered war: saturation at the edge, expensive precision in the depth, and a contested transport problem in between. This one rests on physics, and no energy-density breakthrough is due within the decade. The failure mode is not technical but intellectual: inferring strategic dominance from tactical dominance — the Douhet trap in miniature.
4. The electromagnetic spectrum becomes the precondition domain — the new air superiority. Every uncrewed system depends on a radio link, navigation signals, or both — and electrons are cheaper than airframes: jamming costs structurally less than the thing it jams. No force will operate without first, or simultaneously, winning its patch of spectrum. It is air superiority’s old role — invisible, local, re-contested in minutes. The only thing that could repeal this is autonomy severing the dependence entirely — but autonomous systems still must perceive and navigate, and perception can be blinded. The spectrum never loses its value; it relocates it.
5. Autonomy in the final seconds becomes standard — as a forced countermeasure, not a doctrinal choice. If jamming makes remote control unreliable in the terminal phase, two options remain: abort, or let the system decide its last seconds. Terminal target recognition is a narrow, bounded perception problem — precisely the class machine learning solves today. Selection pressure will settle the ethics debate: whoever refuses terminal autonomy loses effect against anyone who jams. This follows necessarily — but for the final seconds only. The failure mode is extrapolating it to “autonomous warfare”: between steering an endgame and running an operation lie unsolved classes of problems — context, intent, law under ambiguity.
6. Economics forces layered defence — and the honest metric is cost per prevented damage. A six-figure interceptor against a four-figure weapon is a losing strategy even when every intercept succeeds: each cheap shot drains an expensive magazine. Defence must layer — jamming and deception first, at near-zero marginal cost; guns and cheap interceptors next; the exquisite missile reserved for the exquisite threat. The economic logic is inescapable; whether institutions follow it in time is a separate and much weaker claim — peacetime procurement can ignore economics for years, wartime consumption cannot. You can watch this one fail in practice wherever success is still measured as intercept rate: the metric that vendors of expensive systems prefer.
7. The decisive bottleneck is organizational, not technical. Hardware iterates in years, software in days. Once military effect is software-defined — frequencies, seekers, jamming profiles — every enemy adaptation devalues stocks overnight and is countered within days — faster than any procurement or training cycle. Whoever holds update authority at the front adapts; whoever holds it in ministries loses the duel regardless of technical quality. Peacetime militaries are optimized against this: without the market test of combat, careers and budgets run on proxy metrics — auditability, error avoidance, platform counts — that reward the opposite of delegation and fast failure. That the bottleneck sits here is deducible. Who solves it first is not. The test is not demos or strategy papers but one question: who at the front may change what, without asking permission?
What nobody can know
Four questions cannot be settled from first principles, and anyone claiming a confident answer is usually selling something — a budget, a product, a doctrine, or a book.
Does sensing or deception win the long duel? Both ride the same technology curves; the equilibrium depends on the order of future inventions. Does cheap autonomous mass work against a first-class opponent whose defences are intact? Every data point so far comes from wars between degraded or asymmetric forces. How does escalation behave when autonomous systems interact under the nuclear shadow? Zero data points exist, and the one experiment that would produce them must never run. And how fast do large organizations adapt once their existence is at stake? Peacetime observation measures the wrong variable: bureaucracies ridiculous in peace have transformed in months under existential pressure; agile ones have frozen.
These questions are not a footnote of humility; they are the planning agenda for the decade — not to answer them, which nobody can, but to build forces that survive either answer. The seven statements tell a planner what to stand on; the open questions, what to hedge. Confusing the two lists is how militaries end up, once a century, proving Douhet right on paper and wrong in the field.