Open Models, Technology Diffusion & the Shift from Renting to Owning AI
Kimi K3 shows open models can compete with the best, and might provide a better foundation for enterprises as they shift from renting to owning their AI infrastructure
The launch of the Chinese AI model Kimi K3 challenges the winner-take-all strategy of the frontier model providers OpenAI and Anthropic, but it also has wider implications for enterprises currently building out their AI stacks.
Kimi 3 is a 2.8T parameter Mixture of Experts model, and developers Moonshot say they will release its weights next week. But it is 24x more expensive than the original breakthrough open model Deepseek, whilst still being cheaper than leading frontier models. In the Frontend Code Arena, it is rated as the best model, ahead even of Claude Fable 5.
As Azeem Azhar notes, this will not necessarily lead to companies moving away from frontier models en masse, but it could be a nail in the coffin of their current business models:
For the AI economy as a whole, for companies around the world, for governments that aren’t rich, this is probably a net positive. The inference margins that OpenAI and Anthropic enjoy are significant, and they can maintain them because they have the very best models. But it’s pressure, not displacement. Enterprises don’t buy on price alone. They value security, support and possibly the fancy professional services on offer. And the harnesses OpenAI and Anthropic have built remain a differentiator.
AI economics and geo-politics
The existence of a frontier-level open weights model that is open to adaptation for any purpose - good or bad - poses all kinds of questions regarding safety, regulation and geopolitics. Nathan Lambert digs into some of these in his analysis of Kimi’s impact on the open model sector in the United States.
As Ben Thompson notes in his detailed analysis of Kimi’s impact, Chinese AI strategy is partly about avoiding the US achieving an asymmetric advantage through the dominance of closed frontier models, whilst commoditising their complements to protect the core value proposition, following the Microsoft, Google and Meta playbooks.
As we wrote nearly a year ago, this strategy creates more value for the ecosystem overall, and will accelerate innovation in the application and harness layers, which is where we need to advance more quickly to turn synthetic intelligence into business capabilities:
The diffusion of economic benefits from general purpose technologies is a key determinant of their overall impact. Value capture by a few dominant firms is rarely the optimal outcome.
Electrification and the development of the internet were quite diffuse innovations, with both acting as a force multiplier for all kinds of industries and activities, both old and new. In contrast, the transistor and telephony both led to value capture by a handful of large firms, and only later became easily accessible general purpose technologies.
OpenAI’s reaction to Kimi suggests they see open models as an existential threat to their business model. But perhaps more worryingly, it also points to the possibility of more US government market interference and the bizarre possibility of frontier models being restricted as a strategic national resource.
Europe finds itself stuck in the middle, lamenting its own lack of AI model innovation and fearing AI becoming a geopolitical weapon, as the FT reports today. But with such gain of function in open models, there is a huge opportunity to slipstream China’s smarter AI strategy and use it to build out Europe’s application layer, especially in advanced industrial sectors, whilst reducing its dependence on the United States.
But we have value chains at home
Regardless of geopolitical considerations, the debate in enterprise AI is already shifting away from reliance on expensive, closed frontier models and towards owning more of the emerging value chain around AI. In a piece titled Enterprise AI Is Entering Its ‘Own vs. Rent’ Phase, Shubham Goswami recently likened the frontier model debate to cloud repatriation, where many companies realised in the past few years that they could run a lot of workloads cheaper and more effectively on their own systems:
The cloud repatriation analogy breaks down in one important way: when you move workloads back from AWS, the knowledge of how to run them is largely portable. If your learning loop — your prompts, evals, fine-tuning data, institutional workflows, get built inside OpenAI’s or Anthropic’s infrastructure, is that knowledge actually portable when you decide to own it? Or have you built the thing that’s supposed to be your competitive advantage in someone else’s house?
Two recent pieces shared by Constellation Research point to areas where this is already happening. Salesforce has cut inference bills by ‘rightsizing’ towards open models, which suggests other big token consumers are also probably feeling the pinch of inference costs. And they also carried a short update on the banking sector that observed banks are increasingly looking at building and owning their own AI systems as sources of competitive advantage.
The model is not the product. The harness around it is more valuable, and the real prize is how we combine these into an organisational operating system that combines synthetic and human intelligence in the most optimal ways.
Learning loops and change flywheels
Those firms that focus on adaptation, not just adoption of new tools will create the most sustainable ‘thick value’ using AI.
That means doing the patient, detailed work of re-imagining processes and workflows, turning support functions into scalable services, and replacing bureaucratic controls with codified automated systems that can move at the speed of technology, not committees.
And those that bake in organisational learning loops and continual improvement will accelerate and pull away from those who run a single one-off transformation and then run an updated, but unchanging new operating model on top.
Michelle Parke recently wrote about the power of recursive organisational learning as an accelerant, referencing Niklas Luhmann’s work on social systems:
Organizations do not merely accumulate information; they continually reproduce themselves through communication. Niklas Luhmann described social systems as maintaining continuity not by preserving static records but by recursively generating new communications that become the conditions for subsequent communications. Organizational intelligence, therefore, emerges less from isolated computational capability than from the recursive reproduction of meaningful interactions that continually shape future decisions (Luhmann, 1986).
This is not obscure social theory, but rather a practical use case. AI needs to explicitly learn from an organisation’s knowledge context in order to be effective. But AI also generates an audit trail of decisions, calculations and actions that could supercharge organisational learning if captured and used in service of continual improvement.
This has the potential to become a virtuous circle of improvement: better context → better agentic performance → better learning to feed back into the system.
Enterprise AI leaders would do well to think about this as a key element of their agentic AI architecture, especially if they want to create a self-improving agentic ecosystem as we wrote a couple of weeks ago.