The enterprise AI playbook of the last three years has relied on the assumption that renting the world’s most powerful foundation models, building your workflows around them, and moving fast is the key to success. It was an effective strategy for quick experimentation, but it has now resulted in a specific kind of structural dependency I like to call the “intelligence tax”.

The concept of a technology tax is nothing new, but this tax isn’t just measured in monthly cloud invoices. The intelligence tax, simply put, is the cost of depending on a capability you do not own, one that can be revoked, repriced, retained, or restricted at any time by a provider or a government entity. This tax is becoming more and more real as AI spend increases, and enterprises are paying it in sudden regulatory exposure, vendor lock-in, and the growing risk of relying on third-party infrastructure to run mission-critical operations.

The choice IT leaders now have to make is whether they pay the intelligence tax forever, or invest in owning the capabilities that provide a long-term advantage.

The Frontier Model Tipping Point

The main issue with relying exclusively on third-party large language models (LLMs) is that when you rent your intelligence, your operational resilience sits entirely in someone else’s hands. Regulatory mandates, export controls, and judicial rulings can instantly rewrite the rules of API availability. Recent history has made this vulnerability painfully clear.

For example, the US government recently issued a directive ordering Anthropic to suspend all access to its most powerful AI models for any foreign national, whether inside or outside the United States, including Anthropic’s own employees. The order arrived at 5:21 p.m. ET, and by the end of the day, their models were disabled for every customer with no migration window, no grace period, and no negotiation. These types of black outs are why relying exclusively on third-party LLMs puts your competitive intelligence at risk. When your intelligence is rented, your access, your continuity, and your proprietary data belong to someone else.

The True Cost of Token Economics

Many assume that because the tokenmaxxing phenomena has started to correct itself and token prices have been falling that AI bills will decrease. The reality is that the volume of AI tokens consumed per agent interaction is rising faster than prices are dropping. Even if tokens are cheaper per unit, consuming vastly more tokens means the total bill grows.

At enterprise production scale, that math compounds. Let’s say a company is running 100,000 agent interactions per day. At standard commercial API rates for top-tier frontier models, annual costs for running these workloads can ring you anywhere from $1.3 million to nearly $4 million in raw token usage alone.

Now, let’s look at the alternative. A self-hosted, fine-tuned 7B small language model (SLM) on two H100 GPUs costs roughly $94,000 per year, which includes hardware amortized over three years, power, cooling, maintenance, and a quarter of an MLOps engineer. Even if you double the staffing needed to a full-time employee of MLOps support ($200,000 loaded), the self-hosted cost rises to $244,000. This is still much cheaper than a rented frontier model. For high-volume enterprise workloads, the breakeven point between self-hosting and renting API access occurs at a modest 2,500 to 6,000 daily interactions. Any enterprise running agents at production volume is well past these thresholds.

Most Enterprises Don’t Need Massive Models

While these costs continue to build on themselves, IT leaders need to keep in mind that architecture is rarely all-or-nothing. The practical move is to identify the high-volume workloads where a fine-tuned model matches the frontier and migrate those first. The intelligence tax targets the workloads you should own but instead rent, and these high-volume workloads consume the vast majority of enterprise tokens. That is why low-volume, exploratory, and cross-domain tasks are best kept on the API.

IT teams do not need a model that can do everything, they need a model that does their thing well. They need to perform compliance reviews, claims processing, customer routing, and contract analysis better, cheaper, and without sending proprietary data to a third-party API. When the people closest to the models can’t tell the difference between the rented product and the open-weight alternative, the intelligence tax replaces intelligence for convenience.

The Difference Between Renting and Owning

The intelligence tax isn’t a one-time premium, it’s a compounding disadvantage. A recent Gallup study found a minimal measurable impact of AI on employee productivity. But those seeing no impact are not failing at AI, they are failing at ownership. They rent a model, bolt it onto existing processes, cut headcount to show the board a temporary number, and wonder why nothing is delivering the results they are after.

The difference between renting and owning AI is the difference between a tool and a system. A rented model answers today’s question at today’s price, and that is reflected on the final API bill. An owned SLM that is fine-tuned on your data, deployed in your environment, and improved with every production interaction learns your business and keeps costs proportional to your outcomes.

You can’t rent your way into domain expertise. Every quarter, fine-tuned domain-specific models get measurably better on the tasks that matter to that specific customer, while the frontier API charges the same per-token rate it charged on day one. A fine-tuned model that has absorbed and learned your specific data is not something a competitor can replicate by subscribing to the same model. The ability to learn from enterprise-owned data and contextual knowledge continuously and autonomously is what transforms your competitive moat into a flywheel of intelligence.

Own The Layer That Matters

The cycle of enterprise software always follows a familiar pattern. Early adopters race to rent capability for quick wins, but market leaders typically achieve sustainable success by owning the layers that matter.

In the agentic AI era, that layer is institutional intelligence.

Relying on rented models means paying an ongoing tax on your own operational efficiency while handing your core competitive advantage to platform vendors. The goal for enterprise IT is not to completely abandon frontier models, which still serve a purpose for low-volume, open-ended tasks. Leaders should aim to reclaim the high-volume, domain-specific workloads that drive the most business value.

The intelligence tax is an opt-in cost, one that IT leaders can choose to escape. Organizations that embrace SLMs where relevant will be able to avoid the looming intelligence tax and take ownership of their proprietary advantage.