By Mohamed Ali - Technology & Digital Economy Commentary
July 10 , 2026 - Every company that pays for AI tokens right now is running the same experiment, and almost nobody is tracking the second-order effects.
The first-order effect is obvious: you get an answer. A contract gets summarized. Code gets written. A customer list gets analyzed. The ROI looks clean on a spreadsheet.
But the second-order effect is structural. You are not buying intelligence. You are renting it. And the rent you pay is not just money. It is data. Specifically, the data that describes how your business actually works , your processes, your edge cases, your proprietary logic, the accumulated knowledge that makes your company worth more than the next one.
The model provider keeps all of it. Forever. This is not a hidden clause or a conspiracy. It is the business model. The more they know about how companies operate, the better their models become, the more rent they can charge, the harder it becomes to leave.
That is the mechanism. Here is what it means.
The Data Extraction Engine
In 2025, enterprises shipped more than 18,000 terabytes of internal data into AI tools. A 93% jump from the year before. ChatGPT alone triggered over 410 million data-loss-prevention violations. Most of that data was not random noise. It was source code, technical documentation, contract language, customer segmentation logic , the exact material that defines competitive advantage.
The companies sending this data are not stupid. They have security teams. They have policies. But the infrastructure of frontier AI is designed to make extraction frictionless. A single prompt box, no technical barrier, instant utility. The easier it is to use, the easier it is to leak. That is not a bug. That is the product.
And here is the part that gets missed in boardroom conversations: every prompt trains the model. Not in some abstract sense. Your proprietary logic becomes part of the weight matrix that the provider sells back to everyone, including your competitors. You are paying to build the system that eventually makes your specific expertise replicable at scale.
The Outsourced Learning Loop
When your data flows to someone else's servers, you lose something that is hard to see until it is gone: the improvement cycle.
If you run a model internally, every interaction makes the model better at your specific problems. The knowledge compounds inside your walls. Over time, the gap between your AI and a generic one widens.
If you use a frontier API, that compounding happens inside a Silicon Valley lab. Your competitive advantage is being externalized, averaged across thousands of companies, and sold back to you as a commodity. The model gets smarter, but not specifically for you. It gets smarter for everyone, which means, in relative terms, you stay flat.
That is the real cost. Not the per-token pricing. The fact that you no longer own the flywheel.
More from Kenya
The Control Problem
Then there is the infrastructure layer. Policies change. Access gets cut. Pricing shifts overnight. An Austrian fintech was fined €450,000 in March 2026 for running credit scoring through a US-based API because regulators classified it as an unlawful data transfer. Shadow AI breaches now cost an average of $4.63 million, $670,000 more than breaches without it.
These are not edge cases. They are the predictable consequences of building critical operations on infrastructure you do not control. When your vendor's roadmap, legal exposure, or pricing model changes, your options are limited. You can migrate, but migration assumes you have not already built your workflows around their specific interface, their specific model behavior, their specific failure modes.
Most companies do not realize how deep that dependency runs until they try to leave.
The Math Nobody Runs
Self-hosted open-weight inference costs roughly $0.001 to $0.04 per million tokens in electricity. Frontier APIs charge $2.50 to $15.00 for the same volume. At enterprise scale, the hardware pays for itself in months.
The objection is usually talent. "We do not have the engineers to run models internally." But that is changing fast. The tooling around open-weight deployment has matured to the point where the operational gap between API and self-hosted is narrowing every quarter. The real gap is not technical. It is attention. Companies are so focused on the immediate utility of frontier models that they are not modeling the five-year cost structure of dependency.
86% of enterprises expect their AI infrastructure budgets to triple in the next three years. 70% plan to scale on-premise or edge AI by 2028. Less than 5% run sovereign workloads today. That number is projected to hit 75% by 2030.
The shift is already happening. Just quietly. The companies making the move are not the ones with the flashiest demos. They are the ones that looked at the incentive structure and realized that renting intelligence from the same entity that sells it to your competitors is a long-term losing position.
What This Actually Looks Like
Running open-weight models in-house means your data stays inside your environment. Your learning loop compounds for you, not for a provider. Your hardware cannot be remotely throttled or repriced. Your regulatory exposure is bounded by your own infrastructure, not someone else's legal jurisdiction.
This is not anti-innovation. It is simply recognizing that the current default , ship everything to a third-party API, absorb the convenience, ignore the structural cost , is a choice that most companies made without fully understanding what they were choosing.
The companies that survive the next phase of AI adoption will not necessarily be the ones with the best models. They will be the ones that still own the knowledge that made them competitive in the first place.