October 2026
Six days ago my job-screening pipeline ran on a hosted model through an API. This week it runs on a 2-billion-parameter AWS model on my own GPU, and getting there was a one-word change to a command-line flag. Three engines, three completely different origins — a paid API, an independent open-source lab, a hyperscaler shipping a training recipe — all behind the identical endpoint. The 29-job test pool went from minutes over the network to 7.8 seconds local, in 4.4 GB of VRAM.
That's not a price collapse. That's the interface collapsing until the choice of engine becomes a config value.
IBM shipped a compelling WebSphere capability, and Oracle answered in WebLogic, or JBoss pushed out a matching function, or Tomcat caught up — each vendor trying to be the best, and the rest of the market following within roughly a quarter. Nobody held a feature lead for long. What survived was never the feature — it was switching costs and accumulated data nobody could reproduce.
That's what makes this phase interesting, and it's worth getting right: it wasn't open source against the giants. It was everyone against everyone — IBM, Oracle, JBoss, Tomcat, proprietary and open alike — and capability converged anyway. And the clock has collapsed too. That race played out in quarters. My version of it took six days — the swap I described up top happened last week. A hyperscaler publishing the training recipe and the training data under Apache 2.0 is that race running again at fast-forward, except this time the recipe is the accelerant. Anyone can fork it, fine-tune it, and self-host it. The moat against a new entrant is zero by construction.
Here's the part that stung a little. I spent real evenings hand-tuning that screening stack — the score thresholds, the model tie-breakers, the local-vs-hosted choice. All of it is now free and better everywhere. Hand-tuning it was a means, not a moat.
What survived is the stuff I'd been treating as plumbing. Years of job descriptions archived from employer ATS pages, because the job board truncated them at 5,000 characters and cut the requirements section off entirely.
And a label set. That word needs defining, because the rest of this piece turns on it. A label is a recorded human verdict: this role was a real fit, that one wasn't, and this is where the screener went wrong. One verdict is trivia. A history of them, accumulated across every application, is something no competitor can download.
Plus a CV audit that proved seven employers received a stub instead of a CV. All of it derived from my applications and my outcomes. Nobody can clone it, because nobody else lived it.
That's the same lesson the benchmark taught, one level down: when the thing you're competing on goes free, the scarce resource moves to the labels — the verdicts that distinguish the cases everyone gets wrong.
There is a risk in this phase that nobody prices in. When the engine is free and swappable in an afternoon, testing gets skipped — the evals, the guardrails, the hard-case pools. It works on the demo, so it ships. I watched my own benchmark prove how easy that is: 90% agreement on a pool containing zero hard cases measured only the easy half of the problem. A guardrail tested against nothing hard is a guardrail that leaks.
And the debt compounds in two places. The tech debt is the unevaluated edge — every engine swap inherits behaviors nobody re-measured, and the bill arrives in production. The data debt is quieter: when the model is free, the only thing you are really betting on is your own labels, and most teams have never written theirs down, let alone audited them. I audited mine and found seven employers had received a stub instead of a CV. That was not a model failure. It was plumbing nobody had tested — exactly the kind of thing this phase encourages you to skip.
The training set isn't buildable yet. Fifteen outcomes exist and they're the wrong verdicts — "got an interview" is a noisy proxy for "structurally a fit." Of those, only one application still has recoverable text, because the pipeline overwrote its input daily. So this is an accumulation problem, not a modelling problem. Accumulation problems are what moats are made of.
So: a hyperscaler will ship your differentiator for free, and it will ship it faster than you can evaluate it. What it can't ship is your labels, and it can't tell you which of your measurements were measuring nothing. I've now watched this happen three times — twice from the inside, and once to myself.
The capability that wins the quarter is the capability that gets commoditized within the quarter. Bet accordingly, then check what your metrics were actually pointed at.
What's the label set in your world? Your judgement, your labels — the accumulated ground truth nobody else can reproduce.
Also published on LinkedIn.
#AIEngineering #AgenticAI #PlatformEngineering
Built on a home lab, powered by local models, and owned by Andrew Katana.