Analysis
The cost of buying a given level of artificial intelligence has been falling rapidly. The cost of putting it to work is another matter.
In a paper revised in March 2026, researchers at MIT estimated that the price of achieving a fixed level of performance on several AI benchmarks had been falling roughly five- to tenfold a year. Their study covered prices from April 2024 to November 2025. It also found that running the most capable models was becoming more expensive as those systems used more computation. Cheap capability and a costly frontier can exist at the same time. The Price of Progress.
For a business deciding where to spend its next technology budget, that distinction matters. Access to an impressive model is increasingly something a competitor can buy too. What is harder to acquire is the information, operating machinery and experience needed to make that model useful every day.
This is the strongest version of the argument that infrastructure is becoming AI’s competitive moat: as widely available capabilities become cheaper, more of the advantage can move into the systems around them.
For the companies supplying AI, that can mean chips, electricity and data centres. For a retailer, manufacturer or professional-services firm, it is more likely to mean reliable records, connections between business systems, clear permissions and a process that turns a suggestion into completed work.
The research supports that shift in emphasis. It does not establish that intelligence has become a uniform commodity, that every model is equivalent, or that buying more infrastructure guarantees a successful business.
Is artificial intelligence becoming a commodity?
There is a striking historical measure of how quickly access has changed. Stanford’s 2025 AI Index reported that the price of querying a model matching GPT-3.5’s performance on the MMLU knowledge benchmark fell from $20 per million tokens in November 2022 to $0.07 by October 2024—a reduction of more than 280 times. Tokens are the small units of text models process. Stanford AI Index.
That comparison holds a particular benchmark threshold constant. It does not mean a whole business process became 280 times cheaper, or that every capability of the two systems was identical. It is also a historical comparison, not a quotation for today’s services.
The newer MIT paper offers an important qualification. Its authors account for the tokens used to complete benchmark tasks, including the extra computation associated with reasoning. Their estimates are more modest than some earlier token-price comparisons. The study is a preprint, and its software-engineering estimates are less precise because that sample is smaller. Gundlach and colleagues.
The useful conclusion is that yesterday’s expensive level of capability can become substantially more affordable. That gives businesses more room to experiment and makes access alone a less convincing claim to exclusivity.
But a commodity is normally something buyers can substitute with little concern about who supplied it. AI has not reached that condition across the board. A system good at drafting routine correspondence is not necessarily dependable at interpreting an unusual technical document or carrying out a sequence of actions.
Stanford’s 2026 AI Index illustrates this unevenness. It reports major improvements in agents operating computers, while noting that they still failed roughly one in three attempts on a structured benchmark. Such results describe particular tests, not a universal failure rate for business AI. They nevertheless caution against equating impressive headline capability with dependable performance on every task. 2026 AI Index.
“Intelligence is everywhere” is therefore a useful shorthand with limits. What is spreading is access to increasingly capable tools. Access remains uneven, and a tool’s value depends on the job, the information supplied and the conditions under which it is used.
Why infrastructure can matter more as models spread
Economists have a name for the resources needed to turn an invention into a business: complementary assets.
The connection to AI is explicit in Old Moats for New Models, by Pierre Azoulay, Joshua Krieger and Abhishek Nagaraj. Drawing on innovation economics, the authors argue that control of specialised infrastructure and capabilities could be a more durable source of advantage than protecting the underlying ideas alone. Their analysis includes computing capacity and access to data. Author-hosted paper.
This is an argument about how competition may develop, not an experiment proving that infrastructure owners will win every market. Its importance is the distinction it draws: knowing how a technology works is different from possessing everything needed to deploy it competitively.
A model can be broadly accessible while some of the things needed to exploit it remain scarce. That helps explain why cheaper AI does not necessarily produce an equally distributed AI economy.
It also means two quite different infrastructure questions must be kept separate. Who can supply intelligence at scale? And who can apply it effectively to a particular business problem?
The physical limits behind an apparently weightless service
A chatbot arrives on a screen. Behind it is a system that needs power, cooling, networking, memory and computing equipment.
In its April 2026 report, the International Energy Agency projected that global data-centre electricity consumption would rise from about 485 terawatt-hours in 2025 to 950 in 2030. That is its central projection, not an observed outcome. The total covers data centres generally; electricity use at AI-focused facilities is projected to triple over the period. IEA: Key Questions on Energy and AI.
The agency identifies constraints in grid connections, energy equipment and advanced chip supply chains. It also describes a tension between improving efficiency per AI task and growing demand for more intensive applications.
For infrastructure suppliers, the implication is practical. An announced facility, an order for equipment and capacity that customers can actually use are different stages. Securing a dependable supply of power and bringing a functioning site online can matter as much as announcing its size.
That does not make every data centre a good investment. A facility still needs customers, sufficient use of its equipment and revenue that covers its costs. Scarcity can create an opportunity without protecting the owner from changing demand or better technology.
Nor does this imply that an ordinary business should buy its own servers. Renting computing capacity can be entirely consistent with building a durable AI advantage. The strategic question is which parts of the system need to be distinctive, and which can sensibly be bought from someone else.
The information a general-purpose model does not have
Consider a hypothetical customer asking a retailer to replace a damaged item. A general-purpose model may be able to write a courteous reply. To resolve the case, the business also needs the order record, the applicable returns policy, current stock and an accurate delivery address.
If one of those records is missing or wrong, better prose will not fix the problem.
This example captures why company-specific information can become more valuable as general capabilities spread. The advantage lies less in accumulating documents than in being able to supply the right information, with permission, at the moment it is needed.
A 2020 paper by Patrick Lewis and colleagues helped establish an influential approach to this problem: retrieval-augmented generation. It combines a language model with information retrieved from an external collection. In the researchers’ experiments, this improved performance on knowledge-intensive tasks and produced more factual language than the comparison model. NeurIPS paper.
The basic idea is accessible: let the system consult relevant material before answering. The paper does not show that connecting a model to any document store makes its answers trustworthy.
For the hypothetical retailer, useful infrastructure would distinguish today’s returns policy from an expired version. It would connect the correct customer to the correct order, record where an answer came from and prevent one customer’s information from appearing in another’s conversation.
There is also a difference between giving a model a lot of information and ensuring that it uses the important part. In Lost in the Middle, published in 2024, researchers found that the position of relevant information within long inputs could materially affect performance. Those findings concern the models and tasks tested then; they should not be treated as a verdict on every current model. Liu and colleagues.
The enduring practical lesson is to test with the information arrangements a business actually uses. A demonstration with one clean document says little about a collection containing conflicting versions, incomplete records and exceptions.
None of this requires describing an untidy shared drive as a proprietary asset. Data becomes strategically useful when it can support a better decision or a better service. Otherwise, it is simply more material to maintain.
An answer is not a completed job
The next difficulty begins when the system has enough information to act.
In the retailer example, approving a replacement might require reserving stock, creating a shipment, updating the customer record and sending a confirmation. If the connection fails after the shipment is created, repeating the whole process could send a second item.
That is a hypothetical failure, but it makes the distinction clear. The business needs a record of what happened, a way to resume unfinished work and a person responsible for exceptions. These are features of a working service, not qualities that can be inferred from the model’s writing ability.
The problem predates generative AI. In Hidden Technical Debt in Machine Learning Systems, published in 2015, D. Sculley and colleagues described how data dependencies, configuration, feedback loops and changing conditions create maintenance burdens around machine-learning systems. NeurIPS paper.
For today’s applications, this suggests a straightforward purchasing test: ask what happens when something goes wrong. Can the system explain which information it used? Can an employee correct a record? Can a failed action be recovered without duplicating it? Can the business detect when quality deteriorates?
Clear authority matters too. Drafting a response, recommending a refund and authorising a payment are different responsibilities. A business can automate one without handing over all three.
NIST’s voluntary AI Risk Management Framework and its companion playbook organise this work around governance, understanding context, measurement and management. They are guidance, not a certificate that a particular product is safe. NIST AI RMF Playbook.
The commercial point is simple: customers buy a resolved problem. A company that can deliver one consistently has something more valuable than a convincing demonstration.
What the workplace studies actually show
There is credible evidence that AI can improve work. There is also evidence that giving people access does not automatically reorganise a business.
In Generative AI at Work, published in the Quarterly Journal of Economics in 2025, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied 5,172 customer-support agents. Access to an AI assistant increased issues resolved per hour by 15% on average, with larger benefits for less experienced and lower-skilled workers. The assistant suggested responses inside the support process; people remained responsible for the conversation. Published paper.
That is evidence from a particular deployment at one company. It is not a forecast of the gain every employer will achieve, and the study does not isolate the commercial value of each piece of supporting infrastructure.
A separate field experiment, described in the November 2025 revision of Shifting Work Patterns with Generative AI, covered 7,137 knowledge workers across 66 firms. Among treated workers who used the integrated tool, time spent on email fell by about two hours a week in the second half of the experiment. The researchers did not detect broader shifts in the quantity or composition of tasks from providing AI at the individual level. Dillon and colleagues, working paper.
The two studies should not be forced into a single productivity figure. They examine different workplaces, interventions and outcomes. Together, they show why measuring a useful local improvement is different from establishing that an organisation has transformed.
An older economic framework helps explain the gap. In The Productivity J-Curve, published in 2021, Brynjolfsson, Daniel Rock and Chad Syverson describe how general-purpose technologies require complementary investments, including intangible assets. The effort spent developing new capabilities can precede the benefits and complicate productivity measurement. American Economic Journal: Macroeconomics.
Applied to AI, the implication is that training, process redesign and the work of making information usable belong in the investment calculation. The paper is not proof that every difficult rollout will eventually pay off. Some projects will remain poor uses of money.
Managers therefore need a more demanding question than how many employees have logged in. They need to know what work is being completed, how much checking it requires and whether the result is better. RohitAI has examined this distinction in its analysis of AI adoption and productivity measures.
When infrastructure becomes a moat—and when it does not
Infrastructure is not automatically a competitive advantage. A moat is an advantage that rivals cannot quickly or economically reproduce.
If every competitor can rent the same database, connect the same model and install the same software, ownership of that combination says little about defensibility. It may be necessary to compete without being enough to win.
The more promising advantages are specific: records that competitors cannot legitimately obtain; knowledge of difficult customer cases; a distribution channel people already use; or an operating process that has improved through careful correction of real errors.
These are possibilities, not a claim that every business already possesses such assets. A large archive may be obsolete. Customer relationships may not translate into permission to reuse information. A technically elaborate system may cost more to maintain than the benefit it creates.
There is a useful counterfactual. If two competitors received the same improved model tomorrow, would one still deliver a better service? If the answer rests on cleaner information, faster resolution and fewer errors, there may be an advantage beyond model access. If the only answer is the model’s brand, it is worth asking how long that difference will last.
The opposite counterfactual matters as well. Could a much better model remove the need for a costly part of the system? Infrastructure designed solely to compensate for a temporary limitation may lose value as the technology improves.
That is an argument for maintaining the ability to change suppliers and simplify a system when possible. It is not an argument for building an elaborate multi-provider platform before a business has identified a useful task.
What smaller businesses should build first
The practical response need not begin with a large AI programme. A narrower exercise can reveal whether the business has an opportunity worth pursuing.
Choose one recurring piece of work with a recognisable end point: a customer query resolved, an invoice reconciled or a quotation checked. Establish what an acceptable result looks like before introducing the new tool.
Then identify the information the task requires, who owns it and which versions are authoritative. Map the decisions that need a person, especially where an error is costly or difficult to reverse. Compare the existing process with an AI-assisted version on similar work, including the time spent preparing inputs and checking outputs.
The appropriate measure is the cost of an acceptable outcome, with quality reported alongside it. Model usage charges are only part of that calculation.
An illustrative calculation shows why. Suppose a task costs £0.02 in model usage and needs four minutes of review at an assumed labour cost of £30 an hour. The review costs £2, bringing the combined cost to £2.02 before other expenses.
A 90% reduction in the model charge saves £0.018 per task—less than 1% of that combined cost. Reducing review time from four minutes to two, while maintaining quality, saves £1. These are hypothetical inputs, not market prices or measured results. They demonstrate how a cheap model can sit inside an expensive process.
At much higher volumes, or with far less human review, computing costs may dominate instead. The answer has to come from the actual workflow. Track rejected attempts, repeat work, integration costs and maintenance rather than counting only the successful examples.
Only then does it become sensible to decide whether to buy a specialised application, connect existing tools or commission a custom system. Each choice should be justified by the problem it solves. The aim is not to own the most technology, but to retain control of the information and decisions that matter to the business.
The advantage is in making intelligence useful
Falling AI costs are a meaningful opening for smaller firms, workers and creators. They make capabilities available without requiring everyone to finance a research laboratory. They also raise the standard for what counts as differentiation.
The research does not support declaring the model race over. Nor does it support a simple future in which infrastructure owners capture all the value. Better models, product design, expertise and customer relationships will continue to affect the outcome.
It does support a change in the question. Once a capable model is available to several competitors, the next decision is where to invest so that it produces a result people will trust and pay for.
For many businesses, that work will be less conspicuous than the latest model launch: fixing records, connecting systems, defining responsibility and learning from failures. It is slower to demonstrate and harder to copy than access to a chatbot.
The scarce resource is not necessarily another answer. It is the ability to carry a useful answer all the way through to a dependable result.