Essay
Open vs. closed is the wrong question
The Kimi K3 launch and the open letter that followed put real weight behind open models. But for enterprises, picking a side misses the point. The strategic question is which layers of the stack you commit to pay for, and where scarcity works in your favor.
It has been interesting to follow the debate the Kimi K3 launch set off. Last week, an open letter asked Washington not to restrict open-weight models. It launched with 25 signatures, with Nvidia, Microsoft, Dell, IBM, Palantir, and Hugging Face among them. Within a day it had more than 50, including OpenAI and Google.
That is a remarkable pin in the map. When the labs with the most to lose from open weights sign a letter defending them, something has shifted. For enterprises, the practical implication is simple: we now have the option of getting near-frontier intelligence from the open-source world, rather than depending solely on closed frontier labs.
But I don’t think the right response is to pick a side. “Open or closed” is potentially the wrong question.
The wrong question
As a general trend, models will keep getting more capable on both sides of the line. The open world is closing the gap; the closed world keeps moving the frontier. Betting your architecture on which side “wins” is betting on a race whose leaderboard changes every quarter.
In enterprise technology, the question that actually matters is about your stack: which layers are you structurally committing to pay for, and how do you obtain intelligence with maximum flexibility of setup?
That reframing gives you three moves.
Keep the model layer swappable
The first move is architectural: treat the model as a component, not a foundation. If your workflows, evaluations, and integrations can move between models without a rewrite, your intelligence cannot easily be repriced. Every abstraction you build between your business logic and a specific model API is negotiating leverage you bank for later.
Use the inference market as leverage
The second move is commercial. Today’s inference market is genuinely competitive: hyperscalers, OpenRouter, Together, and the neoclouds are all hosting the same open weights. When identical intelligence is available from a dozen suppliers, price discovery works in your favor. That only holds if you preserved the first move; a swappable model layer is what turns a competitive market into an actual negotiation.
Invest where scarcity is yours
The third move is the one that compounds. Put your investment in the layer nobody can commoditize: your data and your workflows. Models are converging; everyone’s agents will reason at roughly the same level. What they reason about is your domain, your processes, and your accumulated operational knowledge: the only part of the stack where scarcity works in your favor.
Where this lands
In two years, I suspect “open vs. closed” will sound as dated as “Mac vs. PC,” because the capital will have finished moving to the layer underneath. For enterprises, that shift will be concrete: the model stops being a strategic decision and becomes a procurement decision, while the real negotiations move to compute contracts and data leverage.
The open letter matters not because open weights are going to win, but because their existence disciplines the whole market. You don’t have to run them to benefit from them. You just have to be architected so that you could.