The Depreciating Asset

Models Are the Depreciating Asset. The Interface Is the Moat.

In 2023, the thing you needed a frontier cloud model to do ran in a data center. In 2026, a lot of it runs on the laptop in front of you. The capability curve bent toward local, and it has not stopped bending. Most people building AI products are investing in the part of the stack that depreciates fastest. The model is the wasting asset. The interface, the coordination layer, and the trust properties of where inference runs are the part that compounds.

That is the whole thesis. The rest of this essay is me trying to convince you it changes how you should build.

The depreciation schedule nobody puts on the balance sheet

Every model advantage is a lease, and the lease keeps getting shorter. The gap between the best cloud model and the best model you can run on hardware you own used to be measured in years. Then it was quarters. Watch any capability that mattered in 2024: structured extraction, code generation, summarization at length, tool calling, decent multilingual reasoning. Each of them crossed the “good enough on local hardware” line, and none of them crossed back.

The mechanism is not mysterious. Open-weight labs are compressing frontier capability into smaller footprints on a relentless cadence, consumer silicon now ships with neural compute that would have embarrassed a 2022 inference cluster, and the quantization and distillation toolchain gets better every month. Frontier cloud models keep improving too, and the top end will stay ahead for a long time. But “the frontier is ahead” is the wrong measurement. The measurement that matters for a product is: can the model available in this environment do this task? For an expanding majority of the tasks people actually pay for, the answer on local hardware went from no, to barely, to yes, to easily.

Here is the uncomfortable accounting. If your product’s value is “we wrap the best model,” you own an asset with a known depreciation schedule and no salvage value. You are renting your differentiation from a lab that also rents it to your competitors, and the open-weight release six months from now marks it toward zero. A moat that rests on model quality is a moat on a countdown, and you do not control the clock.

What actually compounds

So what should you own instead? The parts of the stack that get better, not worse, when a new model ships.

The native interface. The thing the user actually touches is not the model. It is the app: how conversations are organized, how context gets assembled, how results become documents and actions, how the whole thing feels on a phone versus a desktop. Interface quality is earned the old way, through thousands of small decisions and years of refinement, which is exactly why it does not evaporate when the weights underneath it change. Nobody’s UI got worse the day a better model came out.

The harness. Raw model capability and useful work are separated by an enormous amount of unglamorous machinery: tool orchestration, memory, retrieval, permissions, error recovery, the scaffolding that turns “the model can code” into “the work got done.” I spent years as a CTO learning that the model was never the hard part of shipping AI into production. The harness was. That machinery is accumulated engineering judgment, and it transfers intact to every new model you drop in.

The router. Once you accept that no single model wins every task, model choice becomes a scheduling problem: the small fast model for classification, the big one for synthesis, the specialized one for code. A client built to route across models gets stronger with every release from every lab, because each release expands its option set. A client welded to one provider gets stronger only when that one provider ships, and weaker relative to the field the rest of the time.

The trust properties. Where inference runs is an architectural decision with permanent consequences. Data that never leaves hardware you own is a guarantee. Data sent to a third party under a policy is a promise. Guarantees and promises are different products.

Notice the shape of all four: none of them get worse when a new model ships. They get better. Every open-weight release is a free upgrade to a client built to be model-agnostic. The labs are collectively spending billions of dollars a year improving a product they do not own. You just have to be positioned to receive it.

The one advantage that never erodes

Everything above is about relative speed of decay. Privacy is different in kind, because it does not decay at all.

The cloud AI apps have a real quality edge today, and honest people should say so. But look at what each side’s core advantage is made of. The cloud’s edge is model quality, which is the depreciating asset this essay is about. The local side’s edge is that your data, your conversations, your documents, your company’s internals never leave hardware you control. That property is structural. No model release erodes it, no pricing change touches it, and no cloud provider can match it without ceasing to be a cloud provider.

So the two curves only cross in one direction. The quality gap narrows every quarter, because that is what the capability curve does. The privacy gap never narrows, because it is not a capability, it is an architecture. When a temporary advantage competes with a permanent one, you should bet on the permanent one, and you should make that bet before the crossing point rather than after, because after is when everyone else makes it too.

This is already visible in who runs local models today. It started as hobbyists, and it is now doctors who cannot send patient notes to an API, lawyers with privileged documents, founders with cap tables and board decks, engineers with proprietary code, and ordinary people who have read enough terms of service to be done with them. That population grows with every data breach, every training-data controversy, and every enterprise security review. Nothing on the roadmap of any cloud lab reverses it.

The obvious objection

“The frontier will always be ahead, and people always want the best.” The first half is true and the second half is false, and the false half is where products live. People do not want the best model. They want their task done, at a quality bar, with acceptable cost and acceptable risk. Nobody drafts an email with the reasoning mode that solves competition math, and no CFO chooses where inference runs without asking where the data goes. Once the local model clears the bar for the task in front of it, the remaining frontier advantage is surplus the user cannot feel, and the things they can feel take over: speed, cost, interface, and trust. Capability thresholds, not capability leads, decide markets. The thresholds keep migrating onto hardware people already own.

What this means if you are building

Run the test on your own product: if the next major model release shipped tomorrow, open weights, better than what you use today, does your product get stronger or weaker? If the answer is weaker, you have built on the depreciating asset, and the countdown is running whether you acknowledge it or not. If the answer is stronger, you have built something the labs are now improving for you, for free, forever.

Build the thing the next model release cannot obsolete. The interface, the harness, the router, and the trust architecture are that thing. The model is the commodity flowing through it.

I am betting on this with a company. More on that soon.