The Real Battleground for Enterprise AI Is Control Over the Learning
As companies deploy large models, the key advantage shifts from access to algorithms to ownership of the feedback and knowledge loops that drive ongoing improvement.
This article originally appeared on Weijin Research's WeChat Official Account on July 14, 2026. Original Chinese title: 「企业AI的真正战场，是学习过程的控制权」. It has been translated and adapted for an English-speaking audience.
There's a famous saying from the internet era: if you're using a product for free, you are the product.
In the age of AI, that line gets an upgrade: even if you're paying for AI, you might still be its training material.
Companies pay for tokens to buy a model's capabilities. But in real work, they keep feeding the AI system business context, workflows, human corrections, preference signals, and edge cases. These look like ordinary usage traces, but they're actually the most valuable learning signals of the AI era.
The old software business model was companies buying tools. Today's large-model business model could become companies buying tools while also becoming part of how those tools evolve.
This is the knot the token economy still hasn't untied, and it's the question that everyone from Satya Nadella to Alex Karp hasn't fully figured out: are enterprises consuming intelligence, or are they also producing intelligence for model providers?
The agent wave keeps pushing AI assets higher. The market believes that as AI truly enters enterprise production workflows, a new business model seems to have taken hold: companies keep buying tokens, models keep creating value, and intelligence eventually becomes infrastructure, just like cloud computing.
But as the token economy actually unfolds, things aren't that simple. Several big names in American tech have started warning bluntly: today's large-model business model rests on an increasingly strained value exchange.
The problem isn't just that tokens are expensive. It's who controls the learning process generated by token consumption.
The learning process here isn't just model pre-training, and it isn't just employees learning. It's the process of AI completing tasks in real business, receiving feedback, correcting errors, accumulating context, embedding workflows, and forming judgment standards. It includes prompts, tool calls, human corrections, task trajectories, evaluation results, exception handling, permission rules, and business semantics.
Together, these things form a kind of "particular intelligence." It's not the general knowledge found on the internet. It's the capability that a specific company, a specific industry, a specific set of processes has built up through long practice.
Real competitive advantage often comes from this kind of continuous learning. For both companies and models, what builds a moat isn't one-time knowledge. It's the process of constantly accumulating, correcting, and feeding back through real work.
The problem is that today, this learning process is increasingly concentrating in foundation models and AI platforms. When companies buy AI, they aren't just consuming intelligence. They're also continuously creating it. The knowledge, feedback, workflows, and experience they generate can, in turn, become fuel for the models, platforms, and agent systems to keep evolving.
Microsoft is one of the biggest beneficiaries of this AI wave, having invested heavily in OpenAI and Anthropic. But Satya Nadella recently published a series of posts pointing directly at this telling issue. He finds it ironic that model providers reserve the right to learn from customer usage and interaction data while simultaneously imposing restrictive terms on "knowledge distillation."
This is the second month in a row he's voiced concern. Last time, he said society would never allow an AI future that hollows out an entire industry. A platform that can truly endure shouldn't just keep absorbing value; it should help more companies create new value.
The contradiction is this: model companies want to expand their own learning flywheel, while enterprises want to preserve their own closed learning loop.
Palantir CEO Alex Karp says that technical customers want control over their own compute, models, data stack, and core competitive advantage. They want to make sure they own the means of production, rather than handing them over to someone else.
That line gets to the deep anxiety of enterprise AI. Companies aren't just worried about data leaks, and they aren't simply trying to lower token costs. What they're really afraid of is that by outsourcing AI capabilities over the long term, they'll also outsource their organizational learning capacity, industry semantics, process knowledge, and ability to innovate.
Learning by doing is one of the key sources of innovation. When manufacturing companies outsourced production, they didn't just lose factories and workers. They lost the craftsmanship, experience, and organizational capabilities that accumulate continuously on the production line. In the AI era, a similar problem is entering the realm of knowledge work.
If companies just keep burning tokens on external models without being able to turn that experience into their own assets, then the productivity gains from AI may show up more as ongoing payments to model providers than as compounding returns on the company's own capabilities.
This also explains why open-source models are becoming an alternative competitive path. Open-source models are no longer just a tool for lowering token costs. They're an institutional choice for companies to retain a closed learning loop.
As long as open-source models are strong enough, companies can handle inference, fine-tuning, evaluation, routing, and continuous optimization inside their own infrastructure. The feedback, anomalies, corrections, and workflow experience generated in real business can be retained as much as possible within the company's own system, rather than being continuously outsourced to model providers.
In other words, what open-source models offer isn't just a cost advantage. It's a kind of "learning sovereignty."
Of course, open source isn't a universal answer. Companies still need to evaluate capability, security governance, toolchains, model serving, and the ability to iterate continuously. Many companies also lack the ability to maintain a full AI infrastructure on their own. The future market won't have just one shape. It will take multiple forms: closed-source hosted, private deployment, open-source self-built, multi-model hybrid, industry platforms, and more.
But the direction is already clear: the more a company possesses core data, core processes, and strong regulatory requirements, the more it will value control over the learning loop.
This is what makes the distillation debate so important. On the surface, the distillation controversy is about copying between models. At its core, it's a fight over the right to learn: who has the right to learn from whose outputs, feedback, and work processes?
Anthropic argues that without distillation, the gap between open-source models and frontier closed-source models would widen. But similarly, without the tacit knowledge and practical feedback from users' real tasks, the scaling of closed-source models may not always go smoothly.
As several leading American frontier closed-source model companies intensify lobbying in Washington to crack down on open-source models, Nathan Lambert, a former researcher at the Allen Institute, argues that at this stage, "knowledge distillation" has to some degree evolved into a regulatory capture campaign, because many of the proposed solutions clearly favor the frontier model companies pushing for distillation restrictions.
That doesn't mean there's no reason to restrict distillation. Model companies invest enormous capital to train frontier models and naturally worry about their capabilities being cheaply copied, or about safety measures being circumvented. But if restrictions are designed too broadly, they could also increase the scarcity of frontier models and weaken the ability of the open-source ecosystem and the research community to catch up.
So this debate can't be reduced to a simple closed-source versus open-source fight. It's a collision between returns on innovation, safety governance, and the diffusion of knowledge.
At the same time, model companies are extending into the application layer, workflows, and hardware entry points. The reason is simple: whoever controls the workflow controls the feedback; whoever controls the feedback owns the fuel for the next round of model evolution.
OpenAI keeps expanding into applications and hardware entry points. SpaceX acquires Anysphere, the company behind Cursor. Model companies send forward-deployed engineers into client organizations. At root, none of this is just about increasing revenue. It's about gaining entry points into the real world where learning signals are continuously generated.
Code, finance, healthcare, manufacturing, office productivity, customer service, design: these are all high-value learning streams. Once a model enters these workflows, it's no longer just answering questions. It's observing how tasks are proposed, how they're broken down, how they're corrected, how they're evaluated. That's the most precious fuel of the agent era.
The same logic applies to price competition. Recently, companies like OpenAI, Meta, and xAI have been lowering API costs. On the surface, they're competing for customer budgets. At root, they're competing for learning signals. Whoever can enter more real workflows at a lower price gets more task trajectories, evaluation feedback, and application scenarios.
But one question remains. If enterprises simply migrate from one expensive closed-source model to a cheaper closed-source model, while the learning flow remains controlled by an external platform, then the knot they are trying to untie today is merely being retied around another company.
The real problem in the token economy, therefore, is not whether enterprises should use closed-source models, nor whether open-source models can fully replace frontier models. It is how enterprises can avoid permanently outsourcing their own learning process.
The answers differ. Jensen Huang's solution is to build a broader alliance of open-source and ecosystem partners. Satya Nadella advocates building a "frontier ecosystem," not just a "frontier model." Richard Sutton has started a new venture, searching for a learning paradigm different from today's large models — one that moves away from datasets and learns instead from its own experience.
Which approach will ultimately prevail remains unanswered.
But one thing is becoming increasingly clear: what is truly scarce in the AI era is not just model capability, but the ability to learn continuously in real work. Whoever owns this learning loop owns the compounding returns of the token economy.