Wall Street noticed the edge
In The Wall Street Journal on October 11, Andy Kessler argued that the AI data-center boom will end the way the mainframe boom did. Brute force gives way to smaller, cheaper machines. He borrows venture capitalist Steve Jurvetson's rule for when a technology shift happens: the new approach is 10 times better, 10 times cheaper, or both. Jurvetson calls it the factor of 10.
Kessler's evidence is already on the table. Open-weight models are selling tokens for about 90 percent less than the premium American labs. A compressed 27-billion-parameter model now fits in under 6 gigabytes, small enough for a high-end phone. Stanford researchers found that local models can answer 88.7 percent of single-turn chat and reasoning queries. His conclusion is a line Silicon Valley has repeated for decades: intelligence moves out to the edge of the network.
Defense needed it before the price fell
For a consumer, the edge argument is about cost and privacy. No token fees, and you own your data. For a squad, the argument started somewhere else. The network isn't there. An adversary jams it, distance stretches it, or emissions discipline turns it off. A model that lives in a data center does nothing for a team that can't reach the data center.
We have been writing about this for a year. In October 2025 we pointed out that the most capable computer on the battlefield is already in every operator's pocket. A month later we argued that a rented AI capability is one someone else can switch off. In February we warned that metered API costs would hit defense budgets and showed why local AI keeps working when links get contested. Last month we noted that the 1.7 million people on GenAI.mil still can't take it past the wire.
Kessler reaches the same place by way of the balance sheet. Price is the commercial reason to move AI to the edge. A contested spectrum is the military one, and it doesn't wait for a price curve.
Build for the curve, not the model
A factor of 10 every year or so has an uncomfortable implication. Whatever model a program fields today is a placeholder. Something smaller and better will ship before the accreditation package is finished. Software tied to one model or one chip ages as fast as that model does.
EdgeLance treats the model as cargo. Every device is scored into a capacity tier when it enrolls, so the system knows what it can carry. Models travel in signed mission packs that check fit before the team steps off. Every call runs through one inference gateway that applies capacity, classification, and thermal policy, then signs a provenance record onto the answer. A unit's own lessons ride on top as adapters that work with any open-weight base model.
So when a model that needed a laptop last year fits on a phone this year, the phone moves up a tier. The mission software, the operator's workflow, and the record stay the same. Each factor of 10 shows up as a free upgrade to hardware the unit already owns.
Cheaper weights raise the provenance question
One detail in Kessler's column matters more to defense than to anyone else. The models selling at a 90 percent discount come from Chinese labs. By his count, American labs fell from 70 percent of that market to 30 percent in a year, mostly on price.
A commercial developer can pick the cheapest model that works. A defense program can't. It has to know who trained the weights, under what license, and whether anyone changed them between download and the device. As the price of a model falls toward zero, trust is the only question left.
EdgeLance's default loadout is open-weight models under permissive licenses, which a program can own outright. Each model's license and source are recorded in the mission pack and the software bill of materials, and the gateway stamps every answer with the model that produced it. Falling prices mean more models, swapped more often, on more devices. The record gets more valuable every time the price drops.
The hard 11 percent
The Stanford figure covers single-turn chat and reasoning. Mission work is rarely one turn. An intelligence update draws on a dozen documents, three sensor feeds, and the last four hours of reporting. A planning question needs maps, tools, and the unit's own SOPs.
That remainder is where architecture matters. EdgeLance answers from the unit's own documents and cites the source for every claim. The gateway sends the hardest calls to a base server or an approved cloud endpoint when policy and the link allow it. The local model carries most of the load, and the routing policy covers the rest.
Cheaper inference means more agents
Kessler doubts the Jevons argument for data centers. If prices drop 90 percent, demand has to grow tenfold just to keep revenue flat. At the edge the math runs the other way. When inference is close to free on hardware the unit already owns, the unit doesn't spend less. It runs more: more agents per operator, more cameras logging continuously, more sensor data summarized on the device instead of streamed back.
Every one of those agents makes calls somebody will have to explain later. That is why the audit ledger is the moat. Cheap intelligence at scale needs a record of who approved what, which model said it, and what was destroyed when the mission ended.
Gradually, then suddenly
Kessler closes on the old line about how change happens: gradually, then suddenly. For defense, the gradual part is done. The phones, laptops, and heads-up displays are already issued, and the models now fit on them. Whether a unit gets the sudden part depends on software built to take any model on any device and keep a record of all of it.
That is what EdgeLance was built to do. See how the platform routes, records, and burns mission AI on hardware you already own, or explore the AI capabilities.