An American software team can pay less for an AI response than it did a year ago and still submit a larger technology budget. It may be using the service more often, sending longer documents, adding human review and buying new security controls. Meanwhile, its cloud provider is spending heavily on the physical infrastructure needed to answer those requests. Cheap intelligence at the screen can coexist with expensive machinery behind it.
That is the economic tension running through the AI boom. For customers, AI is a potential cost-saving tool. For chipmakers and infrastructure suppliers, it is a source of demand. For investors, it is a claim on future earnings that may take years to arrive. Those are different perspectives on the same transaction, and they do not all benefit at the same time.
This is an explanation of the mechanisms, not a live market report or a recommendation to buy securities. The dated figures below come from identified research and financial statements, principally published in 2025. They are historical evidence, not estimates of today's stock prices, GPU rental rates or the latest quarter's spending. Forecasts are labeled separately.
First, separate the four bills people call ‘AI costs’
The first bill is model development: research, training runs, datasets and the engineering needed to make a model usable. The second is serving the model, often called inference: the computing required when a user submits a question or an application asks for an answer. A provider can spend heavily on training while reducing the price it charges for inference. Neither number, by itself, tells you whether the business is profitable.
The third bill is infrastructure: accelerators, memory, networking, buildings, cooling and electrical capacity. These expenses often arrive before revenue. The fourth is deployment inside an organization: connecting the model to records, controlling access, testing outputs, monitoring failures and changing how people work. Much of that bill is paid to employees and service providers rather than to the model company.
Keeping these categories separate prevents a common mistake. A cheaper model API is evidence of a lower price for one input. It is not evidence that a hospital, bank or software company can run an entire AI-enabled workflow more cheaply. The meaningful comparison is total cost per acceptable result, including the work required when the system gets something wrong.
Why useful AI can become cheaper even during a spending boom
Stanford's 2025 AI Index reports that the inference cost of a system performing at GPT-3.5's level fell more than 280-fold between November 2022 and October 2024. That finding concerns a defined capability level and historical period. It does not mean every model became 280 times cheaper, every enterprise bill fell by that amount, or the same rate of decline will continue. [1]
The direction nevertheless matters. Smaller models, more efficient serving, better hardware utilization and competition can make yesterday's premium capability inexpensive. Open-weight models also give buyers another option, although running a model yourself replaces a vendor invoice with your own computing, maintenance and staffing obligations. ‘Open’ is a licensing characteristic, not a guarantee of zero operating cost.
There is a second distinction between cost and price. Technical improvements can lower a provider's production cost, while commercial competition can lower its selling price. A low introductory price may also reflect a business decision to attract customers. Without provider-level financial evidence, it is unsafe to assume that a falling API price proves a sustainable margin—or proves a loss.
For a US business, the practical opportunity is to avoid paying for more capability than the task requires. A bounded classification task may not need the same model as a difficult research question. But cheaper only counts as better economics if the smaller system still meets the required accuracy, privacy and reliability standards.
A smaller unit price does not guarantee a smaller software budget
Consider a deliberately hypothetical example, not a market-price estimate. A company runs one million AI tasks per month at a model charge of one cent per task, making its model bill $10,000. If the charge halves but usage grows to three million tasks, the model bill becomes $15,000. Each task costs less; the monthly bill is 50 percent larger. That calculation excludes storage, software licenses, monitoring and labor.
More elaborate applications can magnify this effect. A single customer request may trigger document retrieval, several model calls, a tool action and a second pass checking the output. Reasoning-heavy requests and long inputs can consume more resources than a short chat exchange. The unit a vendor advertises may therefore be different from the unit the business cares about: an accurately resolved case, a reviewed contract or a completed transaction.
Software packaging adds another layer. AI may be included in an existing subscription, sold as a per-user add-on or charged by consumption. An included feature is not necessarily costless; the vendor may recover its expense through the broader contract. Buyers should examine renewal terms, usage limits, overage charges and whether a pilot's favorable pricing survives full deployment.
The defensible business case starts with an existing workflow and measures the full change. How much human time is genuinely saved? How often must outputs be corrected? Does faster drafting create a review backlog elsewhere? Task-level improvements are useful, but they do not automatically become lower payroll expense or higher company-wide earnings.
The physical bottlenecks: chips are only the beginning
An AI accelerator works inside a system. It needs high-bandwidth memory, advanced packaging, fast connections to other devices, and enough power and cooling to operate reliably. Increasing the supply of the processor alone does not guarantee that all the complementary components become readily available. A bottleneck can move from silicon production to packaging, memory, networking or installation.
This creates uneven economics. A supplier controlling a scarce component may have pricing power while another part of the stack faces intense competition. A cloud provider can buy expensive new equipment and still face pressure to reduce customer prices. A purchaser's capital expenditure becomes a supplier's revenue, but that accounting relationship does not prove the purchaser will earn a satisfactory return.
Rapid hardware improvement cuts both ways. Better performance can reduce the cost of serving a workload. It can also make an older system less competitive before its owner expected. Financial statements spread equipment costs over an estimated useful life, while the economic value of that equipment depends on demand, utilization and competing alternatives. A change in depreciation assumptions is not the same thing as a change in physical efficiency.
This is why a claim that ‘AI makes technology more expensive’ is too broad. It can raise demand for particular memory products or electrical equipment without raising the price of every laptop or every cloud service. Consumer hardware pricing also depends on inventories, product cycles, exchange rates and other demand. An AI explanation needs evidence for the specific product, not just a plausible story about scarce chips.
Electricity turns a technology story into a local infrastructure story
The International Energy Agency's April 2025 Energy and AI report estimated that data centers consumed about 415 terawatt-hours of electricity globally in 2024. Its base case projected demand of roughly 945 terawatt-hours in 2030. Those figures cover data centers overall, not AI alone, and the 2030 number is a forecast rather than observed consumption. [2]
The local constraint can be more important than the global total. A proposed data center needs power in a particular place, at a particular date, with dependable connections. New generation elsewhere does not automatically solve a crowded local substation or a lengthy grid connection process. Delays can leave valuable computing equipment waiting for the infrastructure required to use it.
For US communities, the consequential financial question is who pays for the additions. Rate structures, customer contracts and regulatory decisions determine how much is funded by the large electricity user and how much enters the utility's broader cost base. A signed power agreement does not, by itself, establish the effect on residential bills. That requires the local tariff and regulatory record.
Efficiency improvements do not settle the demand forecast. If each request uses less energy but companies deploy AI in many more activities, aggregate consumption can rise. The opposite is also possible if efficiency improves faster than adoption. Forecasts need both assumptions. The IEA's report is useful for framing scenarios; it should not be treated as a precise schedule for every US utility's earnings.
What a company filing reveals that an AI headline does not
Microsoft's fiscal 2025 annual report provides a concrete historical example. Cash additions to property and equipment were $64.551 billion in the year ended June 30, 2025, compared with $44.477 billion a year earlier. Cash from operations was $136.162 billion. These are company-wide reported amounts, not a clean AI-only spending total. [3]
The distinction between cash and earnings is essential. Buying equipment consumes cash upfront; much of the cost enters the income statement gradually through depreciation. A company can report rising operating income while directing a larger portion of its operating cash toward infrastructure. Operating cash less cash additions to property and equipment is a useful starting comparison, but it is not a complete valuation model.
Leases and commitments also matter. The same report discloses construction commitments, lease obligations and purchase commitments, including datacenter-related obligations. These are not interchangeable with the current year's cash spending. Comparing one company's cash capital expenditure with another company's broader measure including finance leases can produce a misleading ranking. Investors need consistent definitions and reporting periods.
Nor should all cloud growth be labeled AI revenue. Infrastructure can support conventional computing and AI together. The question is how much incremental, profitable customer demand the investment produces, after electricity, equipment depreciation, support and other expenses. A capacity announcement measures intended supply; it is not proof of eventual utilization.
Why stocks can react differently from the underlying technology
A useful technology is not automatically a good investment at any price. A stock reflects expectations about future cash flows, discounted for time and risk. If investors already expect extraordinary growth, an excellent business result can still disappoint. Conversely, a company can benefit from improving AI economics without being the most visible model developer.
The spending cycle has several beneficiaries with different exposures. Chip suppliers sell equipment. Cloud operators rent capacity and sell services. Software firms try to capture workflow value. Electrical equipment makers and construction companies provide physical capacity. Each depends on a different combination of order timing, competition, margins and customer willingness to pay. Treating them as a single ‘AI trade’ hides these differences.
There is also a portfolio concentration mechanism. A capitalization-weighted index gives greater weight to companies with larger market values. Where a small set of large companies shares exposure to an AI spending cycle, an investor who owns a broad index can still have meaningful indirect exposure to that theme. The exact weight changes over time and must be checked against current index data; no current concentration percentage is asserted here.
Lower computing costs can be good for adoption and uncomfortable for some suppliers at the same time. Cheaper capability might create far more demand, or it might reduce what buyers need to spend to achieve the same outcome. The market response depends on which effect investors expect to dominate. A technology breakthrough can therefore prompt a reassessment of earnings across the supply chain rather than a uniformly positive reaction.
Credit risk: the building, the tenant and the servers age differently
Infrastructure financing introduces a second financial-market channel. A company building a data center may borrow, sign a long lease or use a financing partner. Its ability to meet those obligations depends on contracted payments, actual occupancy, operating expense and the creditworthiness of its customers. Strong demand for AI in general does not guarantee cash flow for a particular facility.
A building and electrical connection can serve successive generations of equipment. The servers inside may face a much faster competitive cycle. Lenders and investors therefore need to understand what secures their exposure: real estate, computing hardware, a tenant contract or some combination. The value of one does not automatically protect the others.
A sensible stress test asks what happens if the opening is delayed, electricity costs rise, a major tenant reduces its needs, or market rental prices fall. It also asks who absorbs those changes under the contract. A long-term commitment can stabilize revenue, but only if the counterparty can and must pay under the relevant circumstances.
These are risk mechanisms, not evidence that a broad AI credit crisis is underway. Establishing that claim would require data on leverage, maturities, counterparties and losses. It is equally unsafe to assume that every project is resilient simply because the largest technology companies generate substantial cash. The risk belongs to a specific balance sheet and contract.
AI is also changing how financial markets operate
The investment boom is only one connection between AI and finance. Financial firms can also use AI to process news, summarize disclosures, monitor risk and support trading decisions. In an October 2024 analysis, the IMF described the possibility of faster, more efficient markets alongside higher trading volumes and greater volatility during stress. It presented those as potential effects and policy concerns, not a measured conclusion that generative AI already controls the market. [4]
The concern is correlated behavior. If many participants rely on similar models, data feeds or providers, they may respond to the same signal in similar ways. Faster reaction can improve price discovery in normal conditions while amplifying movements when liquidity is thin. Operational outages, manipulated information and confidently incorrect summaries are additional reasons to preserve independent checks.
None of this makes a chatbot a dependable investment adviser. A fluent explanation can omit a material liability, misread a filing or use an old number as if it were current. Decisions affecting money need traceable source documents, time-stamped data and accountability. The fact that information becomes cheaper to summarize does not make the judgment built on it risk-free.
The productivity case is real—but it needs a narrower question
Research on AI-assisted work offers evidence that the tools can improve performance in particular settings. Stanford's 2025 AI Index summarizes a growing body of productivity studies. But ‘AI helps on a task’ and ‘AI raises economy-wide productivity by a known amount’ are different claims. The latter also depends on adoption, workflow redesign, investment and how saved time is used. [1]
For a business, the useful test is concrete: compare the same kind of work, at the same quality threshold, with and without the tool. Include review time, rework and exceptions. An employee who drafts twice as fast may still wait for an unchanged approval process. A customer-service system that closes tickets quickly can be economically worse if it creates more repeat contacts.
This is also the bridge from technology costs to financial valuations. Sustainable productivity gains can support earnings through higher output, lower costs or better products. Competition may pass some of those gains to customers instead. The amount retained by a supplier is an economic question, not an automatic consequence of better model performance.
What to watch next—and what the evidence cannot yet settle
For technology buyers, watch cost per successfully completed workflow, contract terms and whether spending increases because of genuine demand or unnecessary usage. For infrastructure investors, watch utilization, cash spending, depreciation, power availability and the quality of customer commitments. For software investors, watch whether AI creates paid demand and durable margins rather than merely increasing the cost of an existing subscription.
The optimistic case is that lower prices broaden adoption, better tools raise output, and infrastructure earns attractive returns over time. The cautious case is that capacity arrives ahead of profitable demand, software prices become more competitive, and some hardware or contracts lose value faster than anticipated. Both are plausible mechanisms. Neither can be established by one benchmark, one quarter's capital expenditure or a rising share price.
The limits of this article matter. It does not estimate October 2026 aggregate AI spending, current stock valuations, local electricity-rate changes or the probability of a market correction. The historical sources establish specific developments and help explain the economics; they do not resolve those current measurements. Company disclosures also do not consistently isolate AI from the rest of their operations.
The bottom line is not that AI is making everything cheaper or everything more expensive. It is shifting costs and bargaining power between layers of the technology system. Users may gain affordable capability while providers face enormous bills. Financial markets are trying to determine who keeps the value—and whether the cash earned will justify the capital already committed.