A couple of years ago, the consensus in the AI space was crystal clear. We all thought foundational models were destined to become pure commodities. The logic was sound. The interface was always going to be "English in, English out." That meant swapping one model for another would be trivial if a better or cheaper option came along. The natural result would be a hyper-competitive race to the bottom, leaving no sustainable moat for any of the major players.
The Shift in the Landscape
Back then, with OpenAI and Google releasing their own open-weight offerings a few years ago, it seemed like the whole industry was marching inevitably toward a universally open baseline.
Fast forward to today in 2026, and the reality looks wildly different. The open-weight space is now thoroughly dominated by Chinese models. Architectures from labs like DeepSeek and Qwen are pushing the boundaries, while the frontier American open-weight models we expected to see have largely vanished from the top tier.
The Enterprise Reality Check
Despite this abundance of open-weight power, the corporate market hasn't fully embraced it. There is a deep, structural distrust of Chinese models within American enterprises. Thanks to intense compliance, security, and legal concerns, these open models simply are not getting the enterprise adoption we once anticipated.
Right now, buyers are ruthlessly pragmatic. They care about raw capabilities and not cost. They care more about raw intelligence than customization and local deployment. This has allowed proprietary models to dominate the conversation. The top closed models are moving in lockstep, constantly leapfrogging each other and shattering records on SWE-bench and other software engineering evaluations. As long as these proprietary models keep delivering massive, measurable gains in usability, the enterprise world seems to care a lot less about the open-weight philosophy.
The Inevitable Future
But if you think the dream of commoditization is dead, you are looking too closely at the present.
Commoditization is still inevitable. Keeping model weights closed as a long-term business strategy is simply not going to work. The current breakneck progression of proprietary models will eventually slow down.
More importantly, model distillation has fundamentally changed the game. Because of distillation techniques, there is really no such thing as a truly closed-source model anymore. The expertise, reasoning patterns, and capabilities required to build a frontier model constantly diffuse into the broader ecosystem. It is only a matter of time before that knowledge gets baked directly into the open-weight community.
We are still heading toward a future where the underlying model is a cheap, hyper-competitive, low-margin commodity, mostly dominated by open weights. The real moat, and the true value creation, will always be built on top of that in the application layer.