The week sterling stopped an army
In late October 1956 British and French troops landed in Egypt to retake the Suez Canal. Militarily the operation worked. Paratroopers took their objectives. The canal zone was within reach. Two of the most experienced armies on earth were executing a plan they had rehearsed for months.
Then the pound started to fall. Speculators attacked sterling and the Bank of England burned through foreign exchange reserves at a rate London could not sustain. Chancellor Harold Macmillan went to the International Monetary Fund for an emergency loan. Washington blocked it. The United States also refused Export-Import Bank credit and threatened to sell its own sterling holdings, a move that would have devalued the pound outright. The message from Eisenhower was simple. No ceasefire, no money.
Britain accepted a United Nations ceasefire on November 6, days into a successful operation. France followed. Prime Minister Anthony Eden resigned within three months. The National Army Museum's own summary of the lesson is blunt. From that point on, any major British operation would need American approval.
Nobody defeated the British army. The army depended on an economy, the economy depended on a currency, and the currency depended on a foreign government's goodwill. The dependence sat three layers below the soldiers and none of them could see it. Capability on loan is capability with conditions. Britain learned the terms in the worst week possible.
Every decisive capability creates the same question
The pattern is older than Suez. In the age of steam, navies ran on coal and coal ran on refueling stations. A fleet without coaling rights was a fleet with a radius. In the age of oil, whole national strategies bent around access to fields and sea lanes. The capability that decides wars always generates a second question underneath it. Who can turn it off?
GPS made the question precise. For years the United States intentionally degraded the civilian GPS signal through a feature called Selective Availability, which pushed accuracy out to roughly 100 meters. Every nation and company building on GPS accepted a switch it did not control. The US turned Selective Availability off in May 2000, and accuracy improved five-fold overnight for every receiver on earth. Nothing about those receivers changed. A policy changed.
Washington later committed never to use the feature again and built new satellites without it. But the deeper lesson had already landed in Brussels. Europe spent billions building Galileo, its own satellite navigation constellation, and the European Space Agency states the rationale plainly. Europe wanted a guaranteed positioning service under its own civilian control. Galileo exists because a close ally of the United States looked at a switch in someone else's hands and refused to build its future on it.
That refusal came from friends. The US never used Selective Availability against Europe. It did not have to. The existence of the switch was enough.
The current version runs on silicon
In October 2022 the United States showed the world what a modern capability cutoff looks like. The Bureau of Industry and Security issued export controls that blocked China's access to high-end AI chips, to the American software used to design advanced chips, and to the equipment needed to fabricate them. CSIS analyst Gregory Allen called it an unprecedented degree of US government intervention, a policy built on chokepoints that made evasion extraordinarily difficult rather than merely inconvenient.
Set aside whether the policy was wise. Look at what it proved. A single government demonstrated that it could reach into the supply chain of a rival superpower and remove the substrate of machine intelligence. If that lever works on a nation with China's resources, it works on everyone. Every defense ministry on earth watched that announcement and did the same math. Whatever we build on top of someone else's compute, someone else can take away.
The controls have tightened and shifted repeatedly since 2022, and each revision made a second point. The terms of technological dependence are not stable. They move with elections, with trade disputes, with escalation cycles a dependent nation does not control. A capability whose terms can change without your consent is not a capability you have. It is a capability you are currently allowed.
Intelligence with three off switches
Now put AI into this history. Defense establishments everywhere are wiring machine intelligence into targeting, logistics, intelligence analysis, and command decisions. Most of that intelligence arrives the same way. It runs in a distant data center, reached over a network, under a commercial license, paid from an annual budget line. That architecture contains three separate off switches. Not one of them belongs to the force that depends on it.
The first switch is policy. A cloud model can be revoked by its vendor, restricted by the vendor's government, or altered by a terms-of-service update nobody in uniform ever reads. This is Selective Availability rebuilt in software, except the degradation can be selective by customer, by country, and by use case.
The second switch is the spectrum, and an adversary holds it. Ukraine is the proof. Russian electronic warfare degraded US-made precision weapons so badly that Excalibur guided shells fell from over 50 percent effectiveness to under 10 percent, and Washington stopped supplying them. These were weapons built by the most sophisticated defense industry on earth, defeated by jammers. Any AI capability that needs a live connection inherits the exact same failure mode.
China is building the same lever at theater scale. CSIS documented a rapidly expanded electronic warfare complex at Mumian on Hainan Island, erected in just over a year, with antenna arrays positioned to collect signals and conduct electronic attack across the South China Sea. In any Pacific contingency, the assumption that a forward unit can reach a data center is not conservative planning. It is a hope.
The third switch is money. Metered intelligence gets rationed. When inference is billed per query, someone above the tactical level decides how much thinking a unit can afford this quarter. No army in history has treated ammunition that way and won.
The sovereignty stack
Governments have started to respond. The United Kingdom's AI Opportunities Action Plan declares that Britain must be an AI maker and not just an AI taker, and stands up a sovereign AI unit to back national champions. South Korea, Taiwan, and the EU have announced sovereign AI programs of their own. NATO's revised AI strategy pushes adoption and interoperability across the alliance. The instinct is right. The architecture is usually incomplete.
Sovereignty in the AI era is a stack with four tiers, and a nation needs all four.
| Tier | What it means | The switch it removes | |---|---|---| | Sovereign compute | Inference runs on hardware the force owns, down to the individual operator's device | A supplier or blockade cannot cut the substrate | | Sovereign models | Open-weight models held on national or unit infrastructure | No vendor or foreign government can revoke or degrade the intelligence | | Sovereign data | Mission records, corrections, and outcomes owned by the unit that generated them | No platform holds the force's institutional memory hostage | | Sovereign operations | Full capability with zero connectivity, as the design baseline | Adversary jamming becomes an inconvenience instead of a decapitation |
Most national strategies address the first tier and gesture at the second. They fund data centers and negotiate compute access. Almost none address the fourth, because the fourth is not a procurement line. It is an architectural decision that must be made at the start and cannot be retrofitted. A system designed around connectivity can add offline modes forever and still fail when the spectrum goes dark, because every workflow silently assumes the link.
The tiers compound. Open-weight models are worthless in the field without hardware to run them. Owned hardware is worthless without models licensed to survive on it. Both are worthless if the data generated in use flows back to a platform the nation does not control. And all three are worthless if the whole assembly needs a satellite link to function. Two tiers out of four is not half sovereign. It is not sovereign.
The layer that makes the stack usable
Here is where EdgeLance sits, stated openly because a strategy paper that hides its author's interest is worth less than one that declares it. The four tiers of the stack are necessary and not sufficient. A warehouse of owned hardware and a drive full of open weights do not answer a question at a checkpoint at 2:00 a.m. Something has to turn raw sovereign capability into a working mission tool in an operator's hands. That something is the mission layer, and it is the tier we build.
The mission layer runs any capable open-weight model on whatever hardware a unit already fields, from a command post workstation to a phone in a pocket. It is model-agnostic on purpose. Models improve every few months, and a force locked to one model inherits that model's ceiling and that vendor's terms. When the weights live on the device and the application layer treats them as interchangeable, the force upgrades intelligence the way it upgrades ammunition. It swaps in the better round.
It treats disconnection as the baseline, not the exception. Every function works with zero connectivity, and connectivity is treated as a temporary windfall for syncing rather than a requirement for thinking. This is the fourth tier made real. A jammer that cuts the link cuts nothing that matters.
And it closes the data tier. Every mission generates records, corrections, and outcomes that stay with the unit that produced them. That archive compounds into something no vendor can sell back to the force, because the force already owns it. When a mission ends, the unit can burn the record completely and prove the deletion. Ownership includes the right to destroy.
For allies, this architecture means daily operations carry no American off switch. A partner nation running open weights on its own hardware through a mission layer it controls cannot be degraded by a foreign policy shift, because there is no remote dependency to squeeze. For the United States, the same architecture is insurance against its own supply chain politics. Export rules change, licenses lapse, and vendors merge. Capability that lives entirely inside the force survives all of it.
What allies buy when they buy sovereignty
This architecture changes the alliance math in three practical ways.
Foreign military sales get simpler. The hardest part of exporting AI-enabled capability is what travels with it. A cloud-tethered system exports a permanent dependency, and every partner knows it, which is why sophisticated buyers increasingly discount such offers. A sovereign-stack system exports a capability that the buyer fully possesses on delivery. That is a stronger product and an honest one, and it competes directly with the no-strings pitch that Chinese vendors make to the middle powers both blocs are courting.
Coalition operations get cleaner. Releasability is the quiet tax on every combined operation. When each nation's mission data lives on its own infrastructure under its own control, sharing becomes a deliberate act with defined scope instead of a standing leak. NATO's push for interoperability works better between sovereign systems than between tenants of the same foreign cloud, because sovereign systems can interoperate at the level of shared formats without merging custody.
Partner trust gets provable. Coalition partners today must simply believe that shared data will be handled as promised. Provable deletion converts that belief into evidence. A partner who can watch the mission archive burn at the end of an exercise will share more at the start of the next one. Trust in coalitions has always been built from verified small promises. This makes one more promise verifiable.
In 1956 Britain discovered that its army answered to someone else's treasury. The discovery cost a prime minister his job and an empire its confidence, and it arrived mid-operation, when nothing could be done about it. Every nation now wiring rented intelligence into its forces is scheduling the same discovery. The sovereignty stack is how a nation cancels the appointment. The force that owns its compute, its models, its data, and its disconnected operations can be defeated in battle. It cannot be switched off from a desk in someone else's capital.