Models stopped being the bottleneck
July 2026
Somewhere around 2024, the question that matters in applied AI quietly changed. It used to be can a model do this? For most business tasks the answer is now yes, or close enough that waiting for the next release is not a strategy. The question that decides outcomes today is different: can this institution actually run it?
The numbers say mostly not. Multiple 2026 surveys converge on the same shape: the large majority of enterprise AI pilots, most estimates put it between 80 and 90 percent, never reach production, and the cited causes cluster around data readiness, governance, integration, and ownership rather than model quality. Roughly nine out of ten enterprise AI pilots never make it to production, and the post-mortems almost never blame the model. They blame data that was clean in the demo and filthy in the warehouse. Evaluation criteria that were never defined. Integrations nobody scoped. A pilot that had a champion but a production system that had no owner. The technology industry keeps solving the part of the problem it enjoys, and the part it enjoys stopped being the constraint.
I recognize this pattern because I have lived through it once already. I spent most of a decade in digital assets, first on stablecoins and confidential computing inside a bank consortium, then allocating capital at a fund backed by two large institutional players. The lesson of that decade was never that the cryptography was insufficient. The technology worked years before the institutions did. What separated the projects that mattered from the ones that evaporated was never a whitepaper. It was whether anyone had done the unglamorous work at the point of contact: custody, compliance, integration, an answer to the question who is accountable when this misbehaves?
AI in 2026 is at exactly that point of contact, at much larger scale and speed. And the uncomfortable implication is that the scarce skill is no longer building models. It is a kind of bilingual competence: enough technical depth to know what a system is really doing, and enough institutional fluency to know what an organization will really tolerate.
Why I run everything at home
For the past year I have been running a private version of this experiment. Frontier-quality open models on hardware I own, quantized and accelerated until they are genuinely usable. The most instructive project so far: grafting speculative-decoding heads across mixture-of-experts post-trains, where donor selection turned out to matter far more than the method itself. A writeup is coming. An agent that manages my calendar, reminders, and email, and has to survive contact with my actual life rather than a benchmark. Experiments on where trained model behavior helps and where it breaks in sensitive settings.
I do this partly for privacy conviction: I think capable AI should not require sending your data to someone else’s computer, and for a lot of corporate work it eventually will not be allowed to. But mostly I do it because it keeps my judgment honest. It is very easy to have opinions about why AI deployments fail. It is harder to hold those opinions while being the person who has to fix the malformed chat template, watch the acceptance rate collapse, and discover which failure modes only appear on day forty of daily use. Every deployment problem I have seen in an enterprise post-mortem, I have met a miniature of on my own machines.
The small conclusion from all of this is that the adoption gap is not a temporary condition that better models will close. Better models widen it, because they raise what is possible faster than institutions raise what is operable. The gap is the job. It was the job in digital assets, it is the job in AI, and I suspect it will be the job in whatever comes after.