There's a quiet consensus forming across AI investors and operators, and it's worth saying plainly: raw model intelligence is commoditizing. Frontier labs ship "good enough" models on a rolling basis, open-weight alternatives close the gap within months, and inference prices keep collapsing. In that world, the model is not a moat. The durable advantage has moved up the stack — to the proprietary feedback loop that captures how a system is actually used, and feeds that back into making it better.

This is the "data flywheel," and it's not a new idea. It's why Google search compounds, why Tesla's driver-assist improves with every mile driven that competitors can't replicate, why Netflix's recommendations get eerily good. Each turn of the loop — usage generates data, data improves the product, a better product drives more usage — strengthens a moat that, after twelve to twenty-four months of compounding, often becomes uncrossable.

Agentic coding is where this is playing out fastest right now. The pattern is clean: strong benchmark performance attracts developers; developer adoption generates real-world usage data; that data improves the next model; the better model attracts more developers. Industry coverage traces the current AI coding race directly to this loop — pointing at the moment a frontier lab posted credible SWE-bench scores as the moment the flywheel "started spinning in earnest," forcing every other major lab into reactive mode.

Now put a number on the scale of what's being collected. By early 2026, one leading coding agent was reportedly authoring on the order of 4% of all public GitHub commits — roughly 135,000 a day, with a single-day peak north of 300,000. Set aside whether the exact figure holds; the order of magnitude is the point. That is an extraordinary, continuous stream of real engineering trajectories — problems posed, approaches taken, tests run, failures corrected — flowing through the labs' interfaces and infrastructure. It is, without exaggeration, one of the richest training corpora in software history, and it is being generated by developers, for the benefit of the platform.

Here's the part that should reframe how you think about your own work.

Your interaction data is not generic

The public internet has been largely scraped; high-quality text for pre-training is nearly tapped out, which is part of why recent model jumps feel more linear than exponential. The scarce resource now is specific, high-signal, real-world interaction data — and the most specific interaction data in the world is yours. Your codebase conventions. Your domain's edge cases. The particular way your team decomposes problems, the fixes you accept and the ones you reject. No public dataset contains it. It is, functionally, your fingerprint — and increasingly, your primary intellectual property.

And the mistakes matter most

This is counterintuitive but central. Modern alignment techniques — RLHF and its successors like DPO — run on preference and correction data. A wrong trajectory, labeled wrong, teaches the system something a thousand generic successes cannot. The moat, as one analysis put it, is not the technique; it is ownership of the feedback dataset and the discipline to keep refreshing it. Your errors are not waste. They are the curriculum for the next generation of your agents.

So what does it actually mean to own this?

It means controlling two layers that most teams currently rent:

• The interface — where the interaction happens. This is not cosmetic. Consider a concrete contrast: some cloud coding agents work by uploading a clone of your repository to the vendor's infrastructure, running the task asynchronously in an isolated sandbox, and handing back a pull request. Others run locally, on your machine, against your filesystem. In the first model, the richest version of the interaction trace naturally accumulates on the vendor's side. In the second, it can accumulate on yours. The interface determines who is positioned to capture the fingerprint in the first place.

• The infrastructure — where the memory lives, and under whose control it can be queried, retained, exported, and reused. Ownership that stops at "we get a read-only export on request" is not ownership. Real ownership means the accumulated memory is a first-class asset you hold, port, and build on — not a byproduct sitting in someone else's system, governed by someone else's terms.

None of this is an argument against using great tools. The frontier agents are genuinely excellent and you should use them. It's an argument about where the compounding value settles. If you adopt the tools but let the interaction data flow entirely into infrastructure you don't control, you've captured the short-term productivity and forfeited the long-term asset — the one thing that was actually unique to you.

The honest caveat, so you can make this argument without overclaiming: the labs improved their models through many mechanisms — scaling, synthetic data, algorithmic advances — not interaction data alone. The flywheel is a real and documented dynamic, not a sole cause. But you don't need the strong causal claim for the conclusion to hold. You only need this: interaction data is now among the scarcest and most valuable inputs in AI, the most specific version of it in your world is yours, and value compounds for whoever holds it.

Own your agentic memory. It's the one moat you're generating for free — right up until you give it away.