The defining technology question is no longer whether a company uses AI. It is whether that company controls anything important underneath it. AI adoption is already broad. Stanford’s 2026 AI Index reports that 88% of surveyed organizations use AI in at least one business function, while 70% use generative AI. At the same time, industry produced more than 90% of notable frontier models in 2025. Access is spreading much faster than control.
That creates a different technology divide. Some companies will own the data, workflows and customer relationships that determine how technology creates value. Others will assemble businesses from rented compute, rented intelligence and rented distribution. Both approaches can produce successful companies, but they carry very different economics, dependencies and bargaining power. The important question is not how much of the stack a company owns. It is which parts it cannot afford to lose control over.
The Stack Has Unequal Layers
Technology stacks are often drawn as neat diagrams: infrastructure at the bottom, applications at the top, customers above everything. Strategically, those layers are not equal.
A payroll company does not gain much competitive advantage from building its own server hardware. A fraud-detection company may gain enormous advantage from controlling years of transaction data. A startup can reasonably rent a foundation model while retaining tight control over the evaluation system that decides whether its outputs are trustworthy.
This distinction matters because “build versus buy” is too crude. The more useful separation is between commodity layers and control points.
A commodity layer performs necessary work but contributes little unique advantage. Authentication, file storage and generic compute often fall into this category. A control point affects something harder to replace, such as margins, proprietary knowledge, customer access or product behavior.
| Technology decision | Commodity question | Control question |
| Cloud infrastructure | Can another provider run the workload? | Would migration materially disrupt the business? |
| AI model | Can another model perform the task? | Does the product depend on provider-specific behavior? |
| Business data | Can the information be recreated? | Does it improve the product over time? |
| Distribution | Are there alternative acquisition channels? | Can one platform sharply reduce demand? |
| Customer relationship | Can customers be contacted directly? | Does an intermediary control access to them? |
A well-designed stack does not eliminate rented technology. It makes sure rented components remain replaceable where replacement matters.
The New Scarcity Is Control
Cloud computing trained companies to think of infrastructure as something that could be abstracted away. AI complicates that assumption because the most powerful systems require large concentrations of compute, energy and capital.
The scale is visible in current spending. Microsoft said capital expenditure reached $41 billion in its fiscal 2026 fourth quarter, with roughly two-thirds directed toward short-lived assets such as CPUs and GPUs. Alphabet raised its 2026 capital expenditure outlook in April to $180 billion to $190 billion as it responded to demand for AI compute.
Those numbers do not mean ordinary companies should start building data centers. They show why control over the lowest layer of the AI stack belongs to a relatively small group of companies capable of financing enormous infrastructure programs.
Most businesses will rent that layer, and renting it is usually sensible. The strategic mistake is assuming that rented infrastructure has no consequences simply because the company does not manage the hardware itself.
A product with modest inference costs can switch providers without threatening its business model. A company whose margins depend heavily on inference pricing is in a different position. If every customer interaction creates a significant external compute bill, infrastructure economics eventually become product economics.
The same principle applies higher in the stack: dependence becomes strategic when somebody else’s pricing or policy can alter your product faster than you can respond.
Renting Intelligence Changes Competition
Foundation-model APIs have created an unusual market. Companies can now purchase reasoning, language generation, image creation, coding and multimodal capabilities without training the underlying models.
This removes an enormous barrier to entry. It also means competitors can often rent similar capabilities. The old software advantage was frequently based on what a company could build. AI shifts some of that advantage toward what a company can do around a model that everyone else can access.
Consider two products using the same foundation model.
The first sends customer prompts to the API, receives an answer and displays it in a polished interface. The second combines the same model with proprietary records, task-specific evaluation, workflow history and human corrections collected across thousands of completed jobs.
Technically, both rent intelligence. Strategically, only one is building a compounding asset. This produces a useful test:
If a competitor received access to your exact model, what would they still be missing?
Strong answers include accumulated customer context, specialized data, difficult integrations, proprietary scoring systems and operational knowledge encoded into workflows. Weak answers usually amount to interface design or prompt wording that could be reproduced quickly.
The model itself can be rented. The system around it determines whether the company is also renting its differentiation.
Data Is Where Rented AI Becomes Specific
A general-purpose model knows broad patterns. Businesses create useful AI systems by supplying the context the general model does not possess.
That context can include which customers regularly churn after a certain support issue, which supplier is unreliable for a specific component, how an internal policy is interpreted, which marketing claims have previously failed compliance review, or what an individual account has purchased over several years.
None of those advantages requires owning a frontier model. They require owning the information architecture around it.
This is why data strategy in AI should extend beyond collecting large datasets. The more valuable question is whether a business captures the feedback created when work is performed.
For example, an AI sales assistant generates a recommendation. A salesperson edits it. The prospect responds. The deal progresses or fails. If those events disappear into separate systems, the company gains little institutional learning. If they become structured feedback, the application acquires knowledge that the rented model did not originally provide.
The strongest AI products may therefore look technically ordinary at their foundation while becoming difficult to copy at the context layer.
Dependency Becomes Dangerous at the Exit
Vendor dependence is usually discussed in terms of how much a company spends with a supplier. A better measure is the cost of leaving.
A third-party API that can be replaced in three days is not strategically equivalent to an API around which five years of application logic has been built. Both are vendors. Only one may have real bargaining power.
Exit costs accumulate through architecture. Prompts can become tuned to one model’s behavior. Database schemas can become specific to one platform. Security reviews may have to be repeated after migration. Internal teams may develop operational processes around proprietary dashboards. Customers may depend on features that cannot be reproduced exactly elsewhere.
This suggests that companies should track switching difficulty, not merely vendor count. A practical dependency review should ask:
- How long would replacement realistically take? Include engineering changes, testing, procurement and compliance work rather than counting only the time required to sign a new contract.
- What data can leave with us? A service is more difficult to replace when historical records, metadata or learned configuration cannot be exported cleanly.
- Which product behavior is provider-specific? If customers depend on outputs or capabilities unique to one vendor, switching becomes a product problem rather than an infrastructure task.
- What happens to unit economics if prices change? A dependency deserves more scrutiny when supplier pricing directly determines gross margin.
- Is there a credible second option today? A theoretical alternative has little strategic value if migration has never been tested.
This is where technical architecture becomes negotiating power. The easier it is to leave, the less threatening dependence becomes.
Distribution Is Also Rented Infrastructure
Technology companies tend to think carefully about cloud lock-in while paying less attention to distribution lock-in. Yet a business can own its software, brand and website while remaining dependent on another company for access to demand.
Search engines, app stores, social platforms and marketplaces all sit between businesses and potential customers. They are not part of the software stack in a conventional architecture diagram, but economically they perform a similar role: they provide infrastructure the company does not control.
Professional services make the issue easy to see. A law firm can own its website and its content while still depending heavily on externally governed search results for discovery. Work around SEO for Lawyers therefore fits into a larger technology question about distribution control. The firm owns the destination, but visibility is partly determined by a system whose ranking logic and interface can change independently.
The same pattern affects SaaS businesses dependent on software marketplaces, merchants dependent on marketplace search and publishers dependent on recommendation feeds. Owning the product does not automatically mean owning the route through which customers discover it.
AI Adds Another Gatekeeper
The distribution problem becomes more interesting as AI assistants begin mediating discovery. Traditional search often sends users from a result page to a website. An AI system can potentially compress several steps. It can gather information, compare alternatives and recommend a shortlist before the user ever reaches the businesses being evaluated.
That changes where influence sits. A company may rank well in conventional search yet be poorly represented in AI-generated comparisons because its information is difficult to interpret, insufficiently differentiated or absent from sources the system considers useful. The customer relationship can therefore begin one layer further away from the company.
This does not mean websites or search disappear. It means businesses may face another intermediary capable of shaping demand.
The strategic response is not to chase every new distribution mechanism. It is to reduce dependence on any single one. Direct customer accounts, useful first-party data, repeat usage, branded demand and permission-based communication all become more valuable when discovery channels are volatile.
The strongest position is not maximum audience size. It is the ability to reach customers after a platform stops helping.
Three Companies Can Use the Same AI
Imagine three companies building similar AI workflow software.
The first company rents almost everything. It uses one model provider, one cloud environment, a hosted database and paid acquisition for most customer growth. Its advantage is speed. A small team can ship quickly because specialists operate most of the underlying infrastructure. Its weakness is that several outside decisions can affect its margins and demand simultaneously.
The second company rents selectively. It uses external models and cloud infrastructure but owns its workflow engine, evaluation data, integrations and customer history. It can test alternative models because the application’s most valuable knowledge sits above the model layer. This company spends less on unnecessary infrastructure while retaining control over what makes the product distinctive.
The third company builds deeply. It owns specialized infrastructure or models because latency, privacy, cost or accuracy at that layer directly determines product performance. This strategy requires more capital and engineering work, but the extra control can be justified when infrastructure itself creates competitive advantage.
There is no universal winner. The mistake would be assuming the third company is automatically more defensible because it owns more technology. Building a capability that the market already supplies cheaply can create cost without creating advantage. Selective control usually matters more than maximum control.
A Better Way to Audit the Stack
The most useful technology audit may no longer be an inventory of software licenses. It should identify places where external control intersects with business importance.
One simple model is to place every major dependency into four categories.
| Dependency type | Strategic response |
| Easy to replace, low business impact | Rent freely |
| Easy to replace, high business impact | Maintain tested alternatives |
| Hard to replace, low differentiation | Reduce unnecessary lock-in |
| Hard to replace, high business impact | Own, govern or actively diversify |
This exercise often reveals that the most expensive technology is not necessarily the most dangerous dependency.
A small identity provider can become mission-critical because users cannot log in without it. A single distribution platform can matter more than a large cloud contract if it generates most new customers. A proprietary dataset may appear as an operational asset until the company realizes competitors cannot easily recreate it.
Executives therefore need a different set of technology questions. Instead of asking only how much a system costs, they should ask who can change the terms, what the company can take with it and how long replacement would take. Those questions expose the actual balance of power inside the stack.
The Divide Will Be Selective
AI tools will make sophisticated technology easier to rent. That is good for competition because small companies can access capabilities that previously required huge engineering budgets. It also makes technology stacks more dependent on a small number of underlying providers.
The companies that handle this well will not respond by building everything themselves. They will identify the few places where ownership compounds.
They may rent compute while owning customer context. They may rent models while controlling evaluation and workflow logic. They may depend on search for discovery while building direct relationships that survive ranking changes.
This is the next technology divide: not builders on one side and renters on the other, but companies that know what should remain rented and what must remain theirs. The most dangerous dependency is not the technology you pay someone else to provide. It is the dependency you believed was replaceable until the day you needed to replace it.
