Fed AI Study Weighs Data Centers Against Productivity
Federal Reserve researchers have published a Fed AI study highlighting a widening gap between visible infrastructure investment and productivity gains that remain difficult to measure. Buildings, chips, and construction commitments offer concrete evidence of spending. They do not establish how much more efficiently the broader economy produces goods and services.
That distinction matters for the companies financing computing capacity and the businesses buying AI tools. Both need useful applications to justify their commitments. Their evidence, however, arrives on different schedules: a construction project becomes visible before its eventual customers demonstrate lasting improvements in everyday work.
The central tension is infrastructure investment versus realized economic value. Earlier research by Erik Brynjolfsson and colleagues offers a historical reference: technology spending and organizational improvements do not necessarily appear together. The question is how convincingly businesses can connect the former to the latter.
What the Fed AI study actually establishes
The paper evaluates current evidence; it does not deliver a final verdict on AI’s economic potential.
The November 2026 AI economy research comes from Nick Carollo, Leland D. Crane, Daniel Garcia Molina, and Joseph Nichols. Its official page was last updated November 12. It describes expanding investment and adoption alongside aggregate productivity effects that appear small so far.
The publication belongs to the Finance and Economics Discussion Series, a working-paper collection intended to support research and discussion. Its findings represent the authors’ views. They do not establish agreement by the Federal Reserve Board or other staff.
That boundary is especially relevant when a headline begins with “Fed study.” Readers should not interpret the paper as a monetary policy forecast. Nor does its publication mean policymakers have settled how AI will affect future growth, employment, or inflation.
The timing also deserves care. Most analysis was completed in June 2026, although the paper includes some later information. Its November appearance does not make every underlying observation a fresh November measurement.
For readers, the useful change is the analytical framework. The paper brings physical investment, business adoption, input costs, employment, and productivity into the same discussion. That encourages a more demanding interpretation of familiar AI headlines.
A planned facility indicates an intention to build. An operating facility indicates available capacity. A customer contract indicates demand under particular terms. A demonstrated improvement in completed work indicates something else again.
These observations can reinforce each other, but none automatically substitutes for the others. Treating them as interchangeable makes an investment story sound more conclusive than its evidence permits.
Consider a hypothetical company that purchases access to additional computing capacity. That purchase establishes that management expects a benefit. It does not establish whether employees finish useful work faster, customers receive better service, or operating results improve.
The distinction also works in reverse. An employee can experience a substantial benefit without creating a visible movement in national statistics. A successful application remains relevant even when it represents a small part of the economy.
This is why the paper should sharpen reporting rather than end debate. Optimistic claims need a bridge from adoption to outcomes. Skeptical claims need to explain why limited aggregate evidence rules out benefits in specific settings.
For enterprise buyers, that changes the most useful question. Asking whether AI “works” invites an answer too broad to guide a purchase. Asking which workflow improves, under what conditions, produces something a business can evaluate.
For infrastructure providers, the corresponding question concerns dependable demand. Customers experimenting with a tool create a different foundation from customers repeatedly purchasing it because it improves their operations.
The study’s importance lies in making those distinctions harder to ignore. It places the burden of explanation between the spending announcement and the eventual economic result.
Announced projects are not completed investment
The largest infrastructure numbers describe different stages of development, and those stages cannot be added together casually.
The data center analysis identifies announced plans approaching $300 billion in 2024 and exceeding $1 trillion in 2025. These are project plans, rather than spending already completed.
Using plans available in May 2026, the authors’ middle scenario projects investment rising from about $100 billion in 2024 to over $880 billion by end-2027. That scenario exceeds 2% of gross domestic product, the value of goods and services produced domestically.
The projection depends on assumptions about new plans, abandonment, and construction timing. Roughly one-third of investment goes to structures, with most remaining spending directed toward high-tech equipment.
The authors estimate an import share around 50%, including capitalized services. They also identify power, zoning, materials, labor, and financing constraints. These qualifications prevent a simple reading of projected investment as either guaranteed construction or an equivalent domestic output contribution.
The practical implication is that every infrastructure headline needs a clearly identified measurement. Announcements reveal ambition. Construction activity reveals execution. Equipment installation brings a project closer to service. Customer use determines whether that capacity supports durable economic activity.
For a hypothetical development, the same project might appear in several announcements as it moves through those stages. Counting each announcement as fresh investment would exaggerate the scale of new activity.
A companywide capital expenditure figure creates another interpretation problem. Capital expenditure means spending on assets expected to serve a business over time. Such a figure can cover several activities and does not automatically isolate AI.
An accurate comparison therefore needs consistent boundaries. A project estimate should be compared with another project estimate. Actual spending should be compared with actual spending. A scenario should retain its assumptions whenever it appears.
The GDP distinction requires similar discipline. Domestic production and imported equipment enter national accounting differently. A large purchase can support an American facility without representing an equally large amount of American production.
That does not make imported equipment economically irrelevant. A business can use it to provide valuable services for years. It means the purchase and its later productive contribution answer different questions.
Readers should also resist converting a projected investment level into an identical contribution to economic growth. A level describes an amount at a point or period. Growth describes change relative to another period.
Those differences matter when a large number moves from a research chart into a headline. Removing the measurement period or scenario label can change the claim more than shortening the number.
For infrastructure developers, the resulting pressure is operational. An announced project needs a credible path through approvals, construction, equipment delivery, and customer service. Public attention does not complete any of those steps.
For buyers, the useful question concerns available capacity under actual delivery terms. A distant project announcement offers less assurance than a service a business can use when its workflow requires it.
The investment pipeline remains economically meaningful. Its meaning becomes clearer when each stage retains its own label, rather than being presented as proof that the entire pipeline has already delivered.
Buyers need completed work, not just more capacity
Infrastructure providers and enterprise buyers face the same underlying test: whether additional computing produces outcomes worth sustaining.
The Fed AI study puts that test into focus without resolving it for every company. The following implications concern how businesses should interpret the gap, rather than findings about particular vendors’ returns.
Start with an enterprise buyer considering an AI writing assistant. Producing a draft faster can be useful. The business outcome depends on what happens after the draft appears.
If a reviewer must reconstruct the argument, verify every factual claim, and rewrite the conclusion, generation speed measures only part of the workflow. The relevant comparison includes those later steps.
Alternatively, a draft might require little revision and help an employee complete work that previously remained unfinished. That is a different result, even if the tool produces similar amounts of text.
The purchasing decision needs to distinguish these possibilities. Usage measures engagement with the product. Completed, acceptable work measures progress toward the business objective. Neither should be silently substituted for the other.
A hypothetical customer-support team illustrates the difference. An assistant might help agents locate an approved answer more quickly. Its value depends partly on whether the answer resolves the customer’s problem without creating additional follow-up.
For a software team, generated code faces a similar chain. A proposed change must fit the existing application, pass review, and remain maintainable. Faster generation can coexist with either faster or slower delivery.
These examples do not establish that AI generally helps or hinders. They explain why the same purchasing category can produce different results across organizations.
Context is a plausible dividing factor. A tool working with incomplete instructions faces a different task from one receiving relevant documentation and clearly defined acceptance criteria.
For teams examining that problem, a searchable knowledge base offers a related workflow to consider. Organizing source material is distinct from demonstrating an AI productivity gain, but it helps define what information a workflow requires.
Infrastructure providers encounter the customer side of this challenge indirectly. A business that obtains repeatable value has a clearer reason to keep buying services. A business that cannot identify the benefit has a weaker internal justification.
The same logic helps compare OpenAI and Anthropic as model providers with Microsoft and Amazon as cloud providers. These are different positions in the computing supply chain, rather than interchangeable businesses.
The analytical point is to follow the customer relationship. A model provider needs demand for its capabilities. A cloud provider needs customers for infrastructure services. An enterprise buyer needs a useful result from its own application.
Success at one layer does not automatically establish success at every other layer. An infrastructure supplier can receive payment before an enterprise customer has established whether its application improves operations.
This makes the pressure more specific than a debate over AI enthusiasm. Buyers need evidence that survives review and renewal. Suppliers need demand that survives customers’ evaluations.
A capacity announcement can support expectations about that demand. It cannot replace the customer outcomes on which those expectations ultimately depend.
Workplace evidence supports gains, with boundaries
Research showing a benefit in a defined workplace is stronger than a general promise, but its boundaries remain part of the finding.
Brynjolfsson, Danielle Li, and Lindsey Raymond studied an AI conversational assistant used by customer-support agents. Their published workplace productivity study reports evidence from 5,172 agents.
Access to the assistant increased issues resolved per hour by 15% on average. Effects varied across workers. The paper appeared in the May 2025 issue of the Quarterly Journal of Economics.
This offers a concrete example of productive use rather than an inference from infrastructure spending. It also establishes a specific outcome: customer-support issues resolved per hour, within the studied setting.
That outcome should remain attached to the percentage whenever it travels. It is not a measured increase in every worker’s productivity. It does not establish an equivalent change across every business function.
The distinction helps reconcile positive workplace evidence with limited aggregate effects. A local study asks whether access to a tool changes performance under defined conditions. National analysis asks about outcomes across a much broader economy.
Those questions differ in coverage, timing, and measurement. A clear answer to the first can coexist with an uncertain answer to the second.
For an enterprise buyer, the customer-support research is therefore useful as evidence of possibility and as an example of careful measurement. It does not remove the need to evaluate the buyer’s own workflow.
A comparable evaluation would identify the completed outcome before examining the result. Otherwise, a team risks choosing whichever measure makes the implementation look most favorable.
Suppose a hypothetical assistant increases the number of responses an employee drafts. That result becomes more informative when paired with whether recipients accept the responses and whether problems stay resolved.
The same principle applies when a tool changes the work itself. An employee might produce more detailed documentation, pursue additional options, or take on tasks previously avoided.
A time-saving measure alone can miss those benefits. Conversely, counting additional output without assessing usefulness can overstate them. The evaluation needs to match the reason the business adopted the tool.
Historical research offers another perspective. Brynjolfsson, Daniel Rock, and Chad Syverson’s productivity J-curve research examines how complementary investments affect the measured payoff from general-purpose technologies.
Their framework includes organizational changes, new processes, and human capital. It explains why building the surrounding capabilities can precede measured productivity improvements.
That is a reason to examine implementation costs carefully. It is not a guarantee that every implementation eventually succeeds. A delayed return and a failed project can initially look similar.
The distinction requires follow-through. If a company explains weak results as an adjustment period, later evidence should show what changed and whether performance improved.
For buyers, patience becomes more defensible when it attaches to an observable implementation process. An indefinite promise that benefits will appear later offers much less guidance.
The strongest optimistic interpretation therefore remains conditional. Workplace research demonstrates that gains are achievable. The next challenge is reproducing useful outcomes across more tasks and organizations without assuming identical results everywhere.
Neither weak statistics nor old experiments settle the case
The skeptical case deserves scrutiny too, especially when historical findings become claims about current tools.
METR’s July 2025 developer productivity experiment examined experienced open-source developers working on their own repositories. The study included 16 developers and 246 tasks.
Developers allowed to use AI took 19% longer to complete tasks. The tools reflected early-2025 capabilities, and the researchers explicitly warned against generalizing the result to all software development.
The finding challenges a simple assumption that assistance necessarily reduces completion time. It also demonstrates why user impressions and measured performance deserve separate attention.
But its scope matters as much as its direction. Experienced contributors working in familiar repositories face a particular mix of tasks and standards. Other developers and projects require their own evidence.
Later developments further complicate using that result as a current verdict. In a February 24, 2026 experiment design update, METR said its newer experiment produced an unreliable signal of current effects.
The researchers identified selection problems, including developers declining participation because they did not want to work without AI. They believed newer tools likely provided more acceleration, while emphasizing uncertainty about its size.
This update does not erase the earlier experiment. It changes how confidently that experiment should be applied beyond its original period and setting.
The broader lesson concerns changing technology and changing users. When tools evolve, evidence needs a date. When users change which tasks they attempt, comparisons need to account for that behavior.
Consider a hypothetical developer who uses AI to attempt a task previously considered impractical. Comparing completion time with an earlier workflow becomes difficult because the earlier workflow never produced the same result.
That does not prove the new workflow is productive. It explains why the comparison requires more than asking whether the developer feels faster.
The opposite problem also exists. A user might feel productive because an assistant generates visible activity, while review and correction consume the apparent savings.
A sound evaluation needs to distinguish these situations. It should identify what work was attempted, what met the required standard, and what human effort supported the result.
The Fed AI study warrants the same discipline at a broader scale. Limited measured gains justify caution about sweeping claims. They do not establish that every local benefit is imaginary or that future gains are impossible.
Equally, uncertainty should not become a blanket defense for disappointing results. A company cannot demonstrate value merely by arguing that existing measurements are imperfect.
The reasonable position is demanding in both directions. Positive claims need evidence tied to useful outcomes. Negative claims need evidence broad and current enough to support their conclusions.
For readers, this means treating a dramatic percentage as the beginning of an investigation. Its task, population, comparison, and measurement period determine what it can actually explain.
The disagreement becomes more useful once those details are visible. Instead of arguing over whether AI is universally productive, researchers and buyers can identify where results differ and investigate why.
Watch delivery, customer outcomes, and broader productivity
The next evidence should connect operational progress with customer value before supporting claims about economywide transformation.
Over the next several months, the first signal to watch is delivery against existing infrastructure commitments. Project updates matter most when they distinguish revised plans from facilities and services becoming available.
Evidence of completed delivery would strengthen the execution side of the investment story. Delays or reductions would weaken assumptions that announced capacity arrives on its original schedule.
Neither result alone settles the productivity question. Delivery establishes that an input exists. The next question concerns how customers use it.
The second signal is customer evidence that extends beyond access and experimentation. Watch for clearly defined results involving accepted work, service quality, repeat use, and continued purchasing.
A company reporting an improvement should explain the comparison. The useful disclosure identifies the workflow, the outcome, and whether review effort forms part of the measurement.
Repeatable improvements would strengthen the case that infrastructure spending supports durable demand. Reports focused only on accounts created or content generated would leave that bridge less developed.
This is also where comparisons among products become more informative. A buyer needs to know which tool fits its task and operating conditions, rather than which vendor announces the largest general ambition.
A product that performs well in a narrow workflow can deserve continued use. A broadly capable product can still require more implementation work than a particular buyer can justify.
The third signal is broader productivity evidence, interpreted over time. Aggregate productivity measures output relative to inputs across the economy. A stronger trend would matter, but attributing it to AI requires additional analysis.
A favorable national reading would not establish that every AI investment succeeded. A weak reading would not invalidate a well-designed workplace experiment. The measurements address different levels of activity.
What would strengthen the overall argument is consistency across those levels. Operational delivery, useful customer outcomes, and broader improvements would form a more persuasive chain than any isolated announcement.
What would weaken it is a persistent separation. Capacity delivered without dependable demand would pressure suppliers. Widespread use without measurable benefits would pressure buyers to reconsider their implementations.
These are observation signals rather than predictions of what the next reporting cycle will show. Their purpose is to make future claims testable.
The paper’s lasting value is the discipline it brings to that process. It asks readers to distinguish what has been built, what has been adopted, and what has improved.
Before treating the next AI announcement as evidence of an economic payoff, identify which part of that chain it measures. Look for a completed outcome, a meaningful comparison, and a clear accounting of the effort required. Keep successful workplace evidence in view, while preserving its limits. Read disappointing results with equal attention to their dates and settings. The Fed AI study makes a useful starting point because it separates visible investment from benefits still requiring demonstration. Follow the next project update, customer evaluation, and productivity release with that distinction in mind. Which new evidence actually connects additional computing capacity to better work?



