AI Didn’t Save the Economy — It Rented It

The economy did not become intelligent overnight

Rent is a wonderfully clarifying word.

When a company rents GPU capacity, it is not buying “AI magic.” It is paying for access to a physical production system: chips, servers, cooling, electricity, networking, land, software orchestration, and a pricing model that turns machine time into invoices. That invoice may look like a cloud bill. In national accounts, however, it becomes something more prosaic and more useful: consumption, investment, exports, government expenditure, or an intermediate input.

That is where the current AI economy becomes interesting. Not in the press release. Not in the benchmark chart. Not in the boardroom sentence where someone says “AI is transforming everything” and everyone nods because nobody wants to be the person who asks for a cash-flow bridge.

The paper by Luisa Carpinelli, Filippo Natoli, and Marco Taboga, Artificial Intelligence and the US Economy: An Accounting Perspective on Investment and Production, is useful precisely because it refuses to start with intelligence. It starts with accounting.¹ The authors ask a narrower question: how does the current AI wave show up in measured US GDP?

That sounds dry. It is not. It is the difference between saying “AI saved the US economy” and saying “AI-related investment strongly supported demand, but much of the hardware was imported, while the more durable domestic contribution depends on operating data centers and selling computational services.”

Less glamorous. Much harder to misprice.

The central correction is simple: AI capital expenditure is not the same thing as GDP growth. A dollar spent on a data center can support domestic construction, import foreign hardware, create future cloud-service revenue, raise exports, or become an intermediate input for another firm. Those are different accounting channels. They have different implications for growth, margins, trade balances, public policy, and investor returns.

So the right business question is not “Is AI boosting GDP?” The better question is: which part of the AI production chain is being counted, where is the value added booked, and how durable is the resulting income?

The first accounting trap: capex is demand, not necessarily domestic output

The common story is familiar by now. Hyperscalers and large technology firms are spending enormous sums on data centers. Servers, GPUs, cooling systems, power infrastructure, and software investment have surged. Therefore, AI must be carrying US growth.

The paper does not deny the first half of that story. In fact, the investment numbers are striking.

In the first three quarters of 2025, US investment in information-processing equipment and software grew at annualized quarterly rates of 36.1%, 19.7%, and 5.4%. Within that, computers and peripheral equipment grew by 103.7%, 61.7%, and 46.1%. Large technology firms spent about $245 billion in capital expenditure in 2024 and more than $275 billion in the first three quarters of 2025, with most forecasts placing 2025 AI-driven capex by large tech companies around $300–350 billion.

If one only reads those numbers, the temptation is obvious: AI is not just a technology cycle; it is a macro engine.

But GDP is not a capex leaderboard.

On the expenditure side, the basic national-accounting identity is:

Investment raises measured demand. Imports subtract from domestic GDP. That subtraction is not a moral judgment; it is accounting. If a US firm buys a server assembled abroad, the spending supports investment demand in the United States, but part of the production value belongs to the country where the hardware was manufactured or assembled.

This is where the AI-growth narrative becomes smaller and more realistic. The paper shows that the surge in technology investment was mirrored by a surge in imports of computers, peripherals, and parts. Imports of the relevant categories rose sharply in the first three quarters of 2025, and more granular trade data suggest that server imports nearly doubled compared with the same nine months of 2024. Mexico and Vietnam appear as major sources.

Once the authors net out technology-related imports, the contribution of technology investment to US GDP growth falls materially: from above 1 percentage point to around 0.3 percentage points in Q1, and to about 0.6 and 0.2 percentage points in Q2 and Q3. On average, the import-adjusted contribution is around 20% of overall GDP growth in the first nine months of 2025.

That is still large. It is not “nothing.” But it is not the same as saying AI machines saved the economy. The resilient US consumer, according to the paper, contributed 0.4, 1.7, and 2.4 percentage points to GDP growth in the first three quarters. The consumer did not leave the stage. AI merely walked on with very expensive luggage.

This is the first mechanism the article must keep in view: \ast\astAI spending can be huge while its domestic GDP contribution is smaller than the headline number.\ast\ast

The second accounting trap: the data center is both a construction project and a factory

A data center enters the economy twice.

First, it enters as a construction and equipment project. This is the capex story: land, buildings, electrical systems, liquid cooling, networking, GPUs, server racks, and software. This phase boosts investment demand. It also creates import leakage if key equipment is manufactured abroad.

Second, once operational, the same data center becomes a factory for computational services. It sells GPU-hours, storage, networking, model-serving capacity, training runs, API infrastructure, and cloud services. These flows can appear in GDP through consumption, investment, government expenditure, or exports, depending on who buys them and for what purpose.

That double role is the paper’s most important mechanism.

For cloud infrastructure providers, the data center is the product. For AI labs, it is one of the largest input costs. For chip and server manufacturers, it is the main deployment location for their hardware. The data center is therefore not merely a building full of machines. It is the economic translator between physical AI infrastructure and measured digital output.

This matters because different uses of computational services enter GDP differently.

If a consumer buys a ChatGPT subscription, that is personal consumption. If an AI lab rents cloud capacity to train a new model and records that spending as R&D, it can appear as investment in intellectual property. If a government agency rents cloud infrastructure, it may appear in government expenditure. If a foreign firm uses a US-based cloud service, it can appear as exports of computer services.

But if a domestic manufacturer uses AI services as an intermediate input, the story changes. The cloud bill itself is not counted as final demand. It may still matter enormously by changing production costs, labor needs, product quality, or margins, but its direct GDP accounting route is different.

This is the point where hype usually becomes unhelpful. “AI adoption” is not one accounting category. It is a bundle of final demand, intermediate input use, capital formation, imports, exports, and sectoral value added.

The authors openly state that data are not yet granular enough to identify the precise distribution between final and intermediate use. That limitation is not decorative caution; it affects interpretation. If a large share of AI services is sold to final users, the revenue channel appears in GDP quickly. If a rising share becomes intermediate input for domestic firms, the direct final-demand effect is less visible, while the effect may instead emerge later through productivity, margins, prices, or new products.

For business readers, this is the right replacement for the lazy phrase “AI adoption.” Ask instead:

1. Is the AI service sold to a final user or used as an intermediate input?
2. Is the service produced domestically or abroad?
3. Is the buyer domestic or foreign?
4. Is the spending recorded as consumption, R&D investment, government expenditure, exports, or a production input?
5. Does the transaction create durable value added, or merely shift cost from labor to cloud vendors?

The boring questions are doing the work. Naturally, they get fewer conference panels.

The paper’s evidence is not one result; it is an accounting stack

The paper does not run a single causal experiment. It builds an accounting argument from several layers of evidence. That distinction matters. The evidence is not trying to prove that AI raises long-run productivity. It is trying to locate where AI-related activity currently appears in national accounts.

This matters because a weak reading of the paper would say: “AI contributes X to GDP.” A stronger reading says: “Different accounting views capture different slices of the AI economy.”

The expenditure approach captures the investment surge and the import offset. The value-added approach captures sectoral production, including the output of technology-intensive service industries. The trade data capture foreign supply-chain dependence and service export channels. The appendix connects physical infrastructure economics to possible revenue flows.

None of these layers is perfect. Together, they create a map.

Why the import adjustment changes the story without killing it

The import adjustment is easy to misunderstand.

It does not say AI investment is unimportant. It says the domestic GDP impact of investment depends on domestic value added. That is a sharper and more useful point.

Suppose a hyperscaler spends heavily on a US data center. Some of that money pays US construction firms, engineers, utility providers, landowners, and software teams. Some pays foreign hardware manufacturers or assemblers. Some supports profits of US-headquartered firms whose production network is global. Some later returns as cloud revenue, subscriptions, API usage, exports, or model-training services.

A simple capex headline cannot tell us how much of that becomes US GDP.

The paper’s mechanism helps separate four layers:

For investors, this separation is not academic. A company can report high capex, high revenue growth, and weak free cash flow at the same time. A country can host headquarters and still import much of the hardware. A cloud provider can enjoy scarcity pricing today and face margin compression tomorrow if capacity overshoots demand.

For policymakers, the lesson is equally direct. If the objective is domestic value capture, the location of high-value segments matters: chip design, data-center operation, cloud services, model development, power infrastructure, and service exports. Industrial policy cannot simply count server purchases and declare victory. That would be spreadsheet patriotism, which is less impressive than it sounds.

The service channel is where AI starts to look like production

The investment story is already visible. The service story is more important for the next phase.

The authors argue that as new AI data centers become operational and sell computational services, the resulting revenue stream could contribute to GDP on a scale comparable to the capex that built the facilities. That claim depends on two conditions: high utilization and current GPU-rental pricing.

The appendix illustrates the mechanism using NVIDIA’s GB200 NVL72 architecture. A modern rack with 72 GPUs and 36 CPUs is estimated to cost about $4.5 million to deploy once land, building, electrical systems, liquid cooling, and other infrastructure are included. The rack itself accounts for about $3 million, or roughly two-thirds of the total. Quoted annual rental prices for an equivalent rack range from about $6.6 million to $10.1 million, while estimated yearly operating costs are around $190,000.

At full utilization and current on-demand prices, the implied payback period is close to one year.

This is not a universal law of AI infrastructure. It is an illustrative calculation. The authors are careful about the unknowns: idle time, hardware failures, negotiated enterprise discounts, overhead costs, cloud-management software, network and storage add-ons, and the mix of long-term contracts versus on-demand pricing. The calculation is best read as a sensitivity anchor, not a prophecy.

Still, the business implication is powerful. If capacity remains scarce and utilization remains high, AI data centers are not merely a capex drain. They are machines for converting imported and domestic capital goods into high-frequency service revenue. That is why the paper’s title includes both investment and production. The production side is not a footnote. It is the next accounting act.

Other evidence points in the same direction. The paper reports that value added in technology-intensive sectors made a 0.48 percentage-point contribution to GDP growth in Q1 2025 and 0.69 percentage points in Q2. In Q2, computer systems design and related services alone contributed 0.38 percentage points. The three largest public clouds generated around $350 billion in revenue over the previous four quarters and continued growing above 20%. Exports of computer services reached an annualized $82 billion in Q3 2025, up 11% from the same quarter a year earlier.

Again, none of this is pure AI. The categories are broader than AI workloads. That is a measurement boundary, not a fatal flaw. The point is that the operating side of the digital infrastructure economy is already large enough to matter, and AI likely pushes it further.

The real business question is not growth; it is cash conversion

For business strategy, the paper’s accounting map becomes a capital-allocation map.

The current AI build-out can support growth through at least three channels:

1. Demand support: data-center construction, equipment purchases, software investment, and R&D spending.
2. Production income: cloud services, AI services, data processing, exports of computer services, and sectoral value added.
3. Potential second-round effects: productivity gains, cost reductions, new products, and organizational redesign.

The paper focuses mainly on the first two. The third remains outside its scope because long-term productivity effects are still too early and too difficult to measure reliably.

That boundary is essential. Many AI debates jump straight from capex to productivity. The paper says: not so fast. In the national accounts, AI first appears as spending and service production. Productivity is a slower, second-round outcome that depends on adoption, workflow redesign, complementary investment, labor reallocation, and the creation of genuinely new products.

For firms, that means AI economics should be assessed through cash conversion, not narrative density.

A cloud provider must ask whether high GPU rental prices can survive capacity expansion. An AI lab must ask whether model revenue can cover compute costs without relying indefinitely on investor funding or strategic subsidies. An enterprise buyer must ask whether AI subscriptions and API bills are replacing labor costs, expanding output, improving quality, or simply adding another software line item to the budget.

The paper does not answer those firm-level questions directly. But it gives managers a better diagnostic frame:

This is where the title’s “rented it” matters. The AI economy increasingly runs on rented capacity. That can be efficient. It can also hide fragility. Renting compute turns fixed infrastructure into flexible usage for customers, but it turns demand uncertainty into a capacity-planning problem for providers.

Somebody has to own the machines.

Depreciation is not the only risk; reinvestment is the quieter one

The paper’s forward-looking discussion is balanced in a way that many AI-market debates are not.

One popular concern is depreciation. AI hardware may become obsolete quickly as more powerful chips arrive. Thermal stress and heavy utilization may also shorten the physical life of GPUs. If firms depreciate hardware too slowly, accounting profits may look healthier than economic profits. That is a legitimate concern.

But the authors argue that the depreciation worry may be overstated if payback periods are also short. If a fully utilized data center can recover its investment in a few quarters or around a year, rapid technological obsolescence is less fatal. Older GPUs may also retain value for less demanding workloads, extending their economic life.

The more robust concern is not depreciation alone. It is the reinvestment cycle.

If AI demand keeps rising, firms may need to keep spending aggressively just to maintain competitive capacity. This can sustain gross revenue while suppressing free cash flow. In other words, the income statement may look exciting while the cash-flow statement quietly asks for a chair and a glass of water.

That distinction matters for valuation. A business with high revenue growth, high utilization, and high reinvestment needs can be valuable. But it is not automatically a free-cash-flow machine. Investors need to distinguish between operating profitability and distributable cash after replacement and expansion capex.

The paper also emphasizes two-sided demand risk.

If firms overinvest, excess capacity can compress GPU rental prices and cloud margins. Smaller AI labs and specialized infrastructure providers may be vulnerable. Larger conglomerates may absorb the shock better because they have diversified revenue streams, though their margins could still suffer.

If firms underinvest, service quality may deteriorate. Users may face throttled API access, weaker models, higher prices, or unreliable performance. That creates competitive risk, especially if alternative providers can offer capacity when incumbents cannot.

The uncomfortable point is that demand forecasting is extremely difficult. AI adoption has been unusually fast. The paper notes that by the end of 2025, only three years after ChatGPT’s launch, generative AI services had reportedly surpassed one billion users globally, while the internet took far longer to reach comparable penetration. Rapid adoption makes capacity planning both urgent and imprecise.

So the risk is not simply “bubble or no bubble.” The risk is capacity mismatch under radical demand uncertainty.

That is a much better sentence than “AI bubble,” and therefore less likely to appear in a headline.

What the paper directly shows, and what Cognaptus infers

A disciplined reading should separate evidence from interpretation.

What the paper directly shows: AI-related technology investment strongly supported US demand in 2025, especially through information-processing equipment, software, and data-center-related capex. However, after accounting for imported hardware, the direct contribution to US GDP growth is much smaller than gross investment figures suggest. The paper also shows that technology-intensive value-added sectors and computer-service exports provide evidence for an emerging production and revenue channel.

What the paper argues through accounting logic: Data centers are the backbone of the AI economy because they convert capital spending into computational services. These services can enter GDP through several channels, depending on whether they are sold to final users, used for R&D, bought by government, exported, or used as intermediate inputs.

What Cognaptus infers for business use: AI strategy should be evaluated less like a software adoption story and more like an infrastructure-utilization, service-pricing, and cash-conversion problem. The key questions are utilization, import exposure, service revenue, depreciation, reinvestment intensity, demand durability, and where value added is actually captured.

What remains uncertain: The paper does not estimate long-run productivity effects. It cannot fully isolate AI from broader IT categories. It does not have granular data on how AI services are split between final demand and intermediate use. Its data-center payback calculation depends on current pricing and high utilization, both of which could change quickly. It also does not resolve how profits and value added should be attributed across multinational firms and global supply chains.

Those boundaries do not weaken the paper. They make it usable. A map that marks unknown terrain is better than a motivational poster with arrows.

The practical framework: follow the invoice

The best way for business readers to use the paper is to follow the invoice.

When money moves through the AI economy, ask where it lands.

If it lands in imported servers, the domestic GDP effect is smaller than the spending headline. If it lands in US construction and power infrastructure, it supports domestic investment. If it lands in cloud services sold to final users, it may appear quickly in GDP. If it lands in AI services used as intermediate inputs, the direct GDP effect is less visible, while productivity effects may or may not emerge later. If it lands in exports of computer services, the United States captures external demand through digital infrastructure.

This invoice-following method also helps firms avoid three common mistakes.

First, do not treat AI capex as proof of productivity. Capex is spending. Productivity requires output improvements relative to inputs.

Second, do not treat cloud revenue as pure AI revenue. Public-cloud categories mix conventional cloud workloads with AI workloads, and the paper explicitly notes that hyperscalers rarely disclose clean AI-versus-non-AI splits.

Third, do not treat fast payback as permanent economics. Scarcity pricing can change. Utilization can fall. Enterprise discounts can matter. Hardware can age. Power constraints can bind. Competition can arrive. The economics of renting intelligence are still economics. Very unfair, I know.

The conclusion: AI is visible before it is transformative

The paper’s most useful message is not that AI is overhyped. That would be too easy, and also wrong.

AI is already macroeconomically visible. It has supported investment. It has reshaped import flows. It has lifted technology-intensive sectors. It is expanding cloud-service revenue and computer-service exports. It is forcing firms, investors, and policymakers to think about data centers as central economic infrastructure.

But visibility is not the same as transformation.

For now, AI appears in GDP mainly as investment, imports, data-center services, R&D, subscriptions, cloud bills, and exports. The productivity revolution, if it arrives, will be slower and harder to measure. It will require firms to redesign workflows, reallocate labor, change products, and convert rented computation into durable output gains.

That is why the accounting lens matters. It prevents us from mistaking a construction boom for a productivity miracle, while also preventing us from dismissing AI infrastructure as mere speculative excess. The current wave is neither magic nor nothing. It is a capital-intensive industrial build-out whose value depends on where the machines are made, where the services are produced, who rents the capacity, and whether the revenue can survive the next reinvestment cycle.

AI did not save the economy.

It rented a very large amount of machinery, imported a good portion of it, and started selling the output by the hour.

That is less romantic than intelligence explosion. It is also closer to how economies actually work.

Cognaptus: Automate the Present, Incubate the Future.