What to take from it
Stimulus in 2009. A second Recovery Act of the same size would have raised real GDP by about 1.4 percent at its peak in the third quarter of 2010 (1.8 percent, at the end of 2010, with forward-looking expectations) and lowered unemployment by about 0.7 point, with the Fed at zero throughout; core inflation rises by only about a tenth of a point. Federal debt is about 2.6 percent of GDP higher at its 2011 peak and less than 1 percent higher by 2015, as the stronger economy pays back part of the cost. Reacting by its inertial rule, the Fed would have edged rates up from 2012, to about 0.45 percent by late 2014, rather than holding them near zero until the end of 2015. That is near the low end of the Congressional Budget Office’s estimates for the actual act: on the model’s reading, the economy had room for much more support.
The 2017 tax cuts. With the Fed on its actual path, by 2025 the tax cuts had raised real GDP about 1.6 percent and lowered unemployment about 0.7 point, with federal debt about 3.5 percent of GDP higher. A Fed following the inertial rule would have offset much of the boost, with rates about half a point higher, leaving GDP up about 0.7 percent at most and debt about 8 percent of GDP higher.
The 2021 Rescue Plan. Removing the entire plan raises unemployment by about 0.9 point at the peak and lowers core inflation by less than 0.2 point (a little over 0.2 point with forward-looking expectations). Removing half does about half as much.
The Fed in 2021. A Fed following the Board’s inertial rule from mid-2021 raises unemployment by about 0.9 point at its peak, around the end of 2022, and lowers core inflation by about 0.1 point; the balanced-approach and Taylor (1993) rules raise unemployment by about 2 points and lower core inflation by at most a third of a point, and the first-difference rule raises unemployment by as much as 3.7 points by 2025 and lowers core inflation by up to 0.8 point.
Ukraine. In the model the war’s energy and food price shock raised headline PCE inflation by about two-thirds of a point at its peak in the spring of 2022 and then, as those prices fell back, lowered it by about half a point a year later: it moved inflation around in time more than it raised it. It did almost nothing to core inflation (under 0.05 point), unemployment (about 0.05 point) or GDP (about 0.1 percent), with either kind of expectations, and a Fed on the inertial rule would have set rates within a tenth of a point of the same path, because in FRB/US energy and food prices barely spill over into wages and other prices. Estimates that find bigger effects have that spillover.
The data center boom. The buildout by itself raised real GDP by only about 0.3 percent by the third quarter of 2026, and lowered unemployment by about 0.1 point, because about three-quarters of the spending was imported. Adding the AI stock-market gain, treated as investors expecting higher earnings, brings the total to about 0.6 percent on GDP and 0.3 point on unemployment, with the funds rate about 0.35 point higher.
Tariffs. Without the 2025–26 tariffs, core PCE inflation would have been about 0.6 to 0.7 point lower at its peak at the end of 2025 (the model was set to match the published estimates) and the federal deficit about 0.9 percent of GDP larger. Output and jobs would have been a little lower, not higher, in the short run: with the tariffs, real GDP is up to 0.3 percent higher and unemployment up to 0.1 point lower by mid-2026, when the data end, because at first FRB/US moves spending from imports to U.S.-made goods. The costs come later. Run for ten years on the Econvitals simulator (backward-looking expectations, taxes fixed), the same construction leaves GDP about 0.3 percent lower for a lasting 10 percent tariff under the balanced-approach and inertial rules and the holds, as dearer imported equipment holds back investment and productive capacity; under the Taylor (1993) rule GDP is about back to baseline by then, and under the first-difference rule about 2 percent lower. A Fed on the inertial rule would have had rates about 0.4 point lower by mid-2026. The model has no retaliation and no uncertainty, the channels through which tariffs most plausibly cost output sooner.
How the replays work
Reproducing history. FRB/US is run over the actual data with tracking residuals: every equation gets the adjustment that makes it reproduce what happened, quarter by quarter, so with nothing changed the model’s path is history exactly. Then one input is changed and the model is solved again. Everything else that hit the economy in those years—the pandemic, the supply shocks, the other fiscal bills, the tax rates—stays as it was. In the years shown, the Board’s rule that raises taxes when the debt rises is switched off (except in the data center case, described below), so tax policy stays as it was; the only movement in tax rates is the model’s automatic response to the business cycle (the average tax rate falls when incomes fall), and the change named is the only policy difference between the two lines. With forward-looking expectations the rule switches on after the last quarter shown (see “Expectations” below). “Actual” is the data in the Board’s package, which end in the second quarter of 2026; in the tariff case the quarters after that are the projection in the Board’s data package, which follows the FOMC’s Summary of Economic Projections, while the data center case uses BEA and BLS data through the third quarter of 2026.
The Fed. Either the Fed keeps to what it actually did, or from the quarter named the funds rate follows the Board’s inertial rule, still subject to the model’s 0.125 percent floor; in the data center case the Fed’s response to the boom follows the inertial rule. When the Fed is on a rule, the charts also draw the economy with the Fed on the same rule and nothing else changed (dashed), so the gap between that line and the counterfactual is the effect of the change alone, not of the change plus the switch to the rule. Under forward-looking expectations, “its actual path” means the actual path through the last quarter shown; after that the Fed reacts to whatever the change has left by the inertial rule, in every case, so markets are not asked to believe in a fixed rate for decades.
The 2009 Recovery Act, twice as large. A second Recovery Act is added on top of the real one—another $787 billion over 2009–13, with the same mix and the same quarter-by-quarter timing as the Bureau of Economic Analysis measured for the actual act’s spending and tax cuts. Household tax cuts lower the personal tax rate, and business tax breaks and payments to businesses lower the corporate rate. Unemployment benefits, food stamps, the $250 checks and other payments to people raise government transfers, as does the extra Medicaid money for the states. Education aid keeps teachers on state and local payrolls, highway and other aid raises state and local purchases, and federal purchases rise dollar for dollar. The Fed either keeps the funds rate on its actual path or reacts to the extra growth and inflation by the Board’s inertial rule. That rule called for a rate below zero in 2009 and 2010 (about minus 0.3 to minus 0.8 percent, since it moves only 15 percent of the way each quarter toward a balanced-approach rate that was near minus 5 percent), so the Fed stays at the floor until the rule turns positive, and the rest of the Fed’s actual decisions are left as they were; because this version keeps the Fed’s actual decisions, the comparison is with history. With forward-looking expectations, people learn of the bigger act when it passes in 2009, not before. The scenario assumes no change to the fiscal legislation that followed the act, so the second act comes on top of those later measures. FRB/US’s spending effects are modest—about a dollar of output for a dollar of aid to states, and about a third of that for transfers and tax cuts—and the model has no channel for the credit crunch of 2009, when many households and businesses could not borrow; studies of that period often find larger effects.
The 2017 tax cuts never happen. From 2018 personal and corporate taxes are higher by the Tax Cuts and Jobs Act’s revenue cost, year by year. Personal taxes are higher by 0.85 percent of GDP in 2018, rising in equal steps to 1.2 percent by 2025, a straight line between the official estimates: the Joint Committee on Taxation estimated the individual provisions at 0.9 percent of GDP in 2019, the first full year, and the Congressional Budget Office’s later estimates of what they cost a year are 1.2 to 1.3 percent of GDP, as inflation and income growth raised the value of the lower rates. Corporate taxes are higher by the Joint Committee’s estimates for the business provisions, by fiscal year: about 0.85 percent of GDP in the first three quarters of 2018, since the 2018 fiscal year’s cost is spread over the quarters the law was in force, and about 0.6 percent in 2019, shrinking to under 0.1 percent by 2023 as expensing phased down and research costs had to be spread over five years. The international provisions are left out; most of their revenue was the one-time tax on companies’ past foreign earnings, which fell on old profits rather than changing a tax rate. The total is about 1.5 percent of GDP in 2019 and 1.3 percent in 2025. The differences stay at their 2025 size after 2025, since the individual provisions were made permanent that year. The investment incentives, expensing and the lower statutory rate, enter only through the average rate, which understates them. The Fed is either on its actual path or on the inertial rule from 2018Q1. The model only incorporates the demand side of these changes, not the ways in which business investment would affect the capital stock and output; for the longer run that matters, and Barro and Furman (2018) estimate it directly.
A smaller American Rescue Plan. The plan’s roughly $1.9 trillion is treated as transfers to households, removed on a front-loaded profile: $500 billion in 2021Q2, $300 billion in 2021Q3, $250 billion in 2021Q4, $200 billion in each of the first two quarters of 2022, then tapering to $50 billion a quarter by late 2023. Aid that went to state and local governments is treated the same way, a simplification. Half or all of it is removed, with the Fed either on its actual path or on the inertial rule from 2021Q2. With forward-looking expectations, people learn of the cut when the first payments would have been removed, in the second quarter of 2021, not when the plan passed in March. Treating the whole plan as transfers to households oversimplifies, since it also included aid to states and businesses. And because FRB/US’s inflation responds little to demand—the reason it missed the 2021–22 inflation—it may understate how much the plan added to inflation.
The Fed starts raising rates in 2021. From the quarter chosen, the funds rate obeys a policy rule instead of the FOMC’s actual decisions. The rules use core PCE inflation over the past four quarters, the 2 percent target and the output gap; all but the first-difference rule also use the neutral real rate (1.2 percent in the Board’s data). Inertial rule: 85 percent of last quarter’s rate plus 15 percent of the balanced-approach prescription, the Board’s own rule. Balanced approach: the neutral rate plus inflation plus half the inflation gap plus the full output gap. Taylor (1993): as the balanced approach but with half the weight on the output gap. First difference: last quarter’s rate plus half the inflation gap plus half the change in the output gap over the past year. The first-difference rule reacts to the four-quarter swing in the output gap after the 2020 collapse and takes the rate to about 12 percent in 2022; it is kept because it is one of the Monetary Policy Report’s rules, and the chart shows why few would follow it mechanically. The same weak link between slack and inflation means the model may understate how much an earlier tightening would have lowered inflation; it also has no role for the Fed’s credibility or for the supply-chain disruptions of 2021–22.
No energy and food price shock from the invasion of Ukraine. From March 2022, the month after Russia invaded, oil follows the path the futures market expected in mid-February—about $93 a barrel for March delivery, easing to $87 by June and staying there—and consumer energy and food prices rise only as fast as other prices instead of jumping. Each returns to what actually happened once actual prices come back down to it: oil in September 2022, energy in December 2022. U.S. food prices never fully came back down, so food is held only from March to June 2022, the months world food prices were above their prewar level; after that the model’s own equation for food prices lets the gap close, and it is gone by the third quarter of 2024. At the peak, in the second quarter of 2022, oil is about $19 a barrel cheaper than it actually was, consumer energy prices are 11 percent lower and food prices 2 percent lower (food stays as much as 2.5 percent lower into early 2023). All of the post-invasion rise in oil and energy prices is put down to the war. That is a smaller oil shock than the $45 a barrel that Federal Reserve Board economists attribute to oil supply disruptions in the first half of 2022, though economists at the Dallas Fed argue that global demand, not the war, drove most of that year’s oil prices. Nothing before the invasion is removed, including the run-up in prices on fear of it. The Fed is either on its actual path or on the inertial rule from 2022Q1. FRB/US has no supply chains or separate industries, so energy and food costs barely spill over into other prices and wages; models and studies with that spillover find larger effects on core inflation.
No data center boom. “With” is what happened. “Without” is what happened minus FRB/US’s estimate of the buildout’s effect: the investment in computers, communication equipment and data-center construction above its 2015–2022 trend, the imports that came with it and, unless you leave it out, the AI companies’ stock-market gain. The two lines are close until the buildout starts, which by this measure is the middle of 2024. The details are below.
No tariffs in 2025 and 2026. The average tariff actually paid on imported goods—customs duties from the Treasury’s monthly statements, counted in the month after the goods arrive, divided by the Census Bureau’s figures for goods imports—was 2.7 percent in 2024. It rose to 3.4 percent in the first quarter of 2025, when the first increases took effect, peaked at 11.4 percent in the third quarter and fell to about 7.5 percent by mid-2026, after the Supreme Court struck down the emergency tariffs in February and lower replacements took their place. In the counterfactual the rate stays at 2.7 percent from 2025 on. The extra duties—about $350 billion through the third quarter of 2026, at a peak of about 0.9 percent of GDP—come out in four places. The Treasury does not collect them. Consumers paid half of them, so consumer prices are lower by half the extra duties, phased in the way Federal Reserve Board economists found tariffs pass through: about half within three months and all within about seven. FRB/US on its own takes years to pass a change in costs into consumer prices, so this schedule is imposed rather than left to the model. “Half” is the share that makes the model match the published estimates that the tariffs added 0.4 to 0.8 point to core PCE inflation by early 2026, among them the Board economists’ 0.8 percent on core prices through February 2026 and Chair Powell’s “between a half and three-quarters” of a point that March. Businesses paid half of the duties on imported equipment, which made up 29 percent of goods imports and took 28 percent of the extra duties (Census Bureau), so without them equipment costs less, entered as a lower price of business investment on the same schedule; FRB/US measures that price against the price of output, which the lower consumer prices move, so it is otherwise held on its actual path. Company profits absorbed the rest and keep it, so companies bore half of the duties in all. And imports cost less, so more of them come in. The refunds of the struck-down tariffs, about $130 billion paid back to importers from May 2026, are left out, and the early-2025 rush to import ahead of the tariffs stays in history. Two things that probably made the tariffs more costly are not in the model at all: other countries’ retaliation (exports move only with the dollar and foreign demand) and the uncertainty from the on-and-off announcements, which may have held back business investment. Models with richer trade structures handle these better. The data end in the second quarter of 2026; after that the “actual” line is the projection in the Board’s data package, which follows the FOMC’s Summary of Economic Projections and already assumes the tariffs, and the fourth quarter of 2026 holds the third quarter’s tariff rate. The Fed is either on its actual path or on the inertial rule from 2025Q1.
Expectations. The default is the Board’s backward-looking (VAR) version, in which people learn about a policy only as it shows up in the data. The forward-looking version is offered too: financial markets and wage and price setters see the changed policy coming the moment it starts. The data center case adds a third setting, in which people are backward-looking until the start of 2026 and forward-looking from then on. The forward-looking runs follow the same rules as the forward-looking runs of the simulator on Econvitals. The model is solved 50 years out from the quarter the change starts, far enough that where the solution ends barely moves the years shown: in the tax-cut and tariff cases, solving 75 years out instead moves no line on the charts by more than 0.05, except stock-market wealth in the tariff case, by about 0.2 percent. The Fed’s rule lets its estimate of the neutral real interest rate adjust, moving 5 percent of the way toward the actual real funds rate each quarter; in the backward-looking runs it stays at the Board’s 1.2 percent. And after the last quarter shown, the Board’s rule that raises taxes when debt rises is switched on, so that a lasting change in taxes or spending settles down instead of adding to the debt forever; inside the years shown, tax policy stays as it was. The data center case lets the Fed’s rule learn the neutral rate in the same way and has the tax rule on throughout; its forward-looking runs are all solved to the end of 2072, 50 years from 2023, so the runs that turn forward-looking in 2026 are solved 47 years out. Every combination solves under every setting.
The data center boom in more detail
- The usual count overstates the buildout’s contribution to growth about three times over, and more if the stock gain is left out. Counting the gross spending, the buildout accounted for 0.58 percentage point of the 2.05 percent growth of real GDP in 2025 (28 percent of it) and about 1 point of the 2.6 percent expected for 2026 (38 percent). That count leaves out the imports, the crowding out and the Fed. In FRB/US, with three-quarters of the spending imported and the AI stock-market gain included, the net contribution was 0.17 point in 2025 and 0.32 point in 2026, about 9 to 12 percent of growth. Leave the stock gain out and the spending alone contributed 0.06 and 0.24 point.
- About three-quarters of the boom was imported. The central case gives each kind of spending a plausible direct import share: about 90 percent for computers, 80 percent for communication equipment and 11 percent for data-center construction, 75 percent overall. The alternatives are 85 and 50 percent. On its own, the buildout raised real GDP only about 0.3 percent by 2026Q3, because about 1 of its 1.3 percent of GDP went to imported equipment; crowding out of other spending barely shows up.
- The stock market is the largest channel. The market value of Nvidia, Microsoft, Alphabet, Amazon, Meta, Broadcom and Oracle rose about $11 trillion more than the rest of the market from the end of November 2022 to September 2026, about 10 percent of the market’s total value, and households own about $7.6 trillion of that. The gain is treated as investors expecting higher earnings from these companies, not as investors accepting lower returns on stocks in general, so it works through household wealth rather than by making it cheaper for businesses to raise money. In FRB/US it raises consumption about 0.5 percent by 2026Q3, and the extra demand raises other business investment about 0.3 percent. With the buildout, real GDP ends up 0.63 percent higher and unemployment 0.28 point lower, and the Fed’s rule has the funds rate 0.34 point higher.
- The boom kept the trade deficit from narrowing. By 2026Q3 net exports are 1.2 percent of GDP lower because of the buildout, and the current account deficit is wider by about the same amount. The real dollar barely moves, because the model’s exchange rate responds to the foreign debt the boom builds up as well as to interest rates.
- Forward-looking expectations change the answer a lot. If markets had foreseen in 2023 that the boom would run at its 2026 pace through 2031, real GDP would be only 0.14 percent higher by 2026Q3, the 10-year yield 0.32 point higher and housing 1.5 percent lower. Add a believed productivity payoff of 0.5 point a year from 2027 and GDP ends up a little below where it would have been without the boom (0.14 percent lower), while core inflation is 0.4 point higher, the 10-year 0.75 point higher and housing 3.6 percent lower. Nobody expected the boom in 2023, so this setting bounds the answer from one side.
- Expectations that turn forward-looking only in 2026. The whole extra rise in long rates arrives at once—the 10-year yield is about 0.4 point higher by 2026Q3 (0.75 point if people also expect a productivity payoff of 0.5 point a year), a little more than if markets had seen the boom coming from 2023—but there is not yet time for it to show up much in the economy: real GDP is still 0.51 percent higher, against 0.63 percent with backward-looking expectations, and housing is about 1.2 to 2.5 percent lower. Turning forward-looking in 2026 trims the boom’s effect at every import share: at 50 percent imported, GDP is up 0.81 percent by 2026Q3 instead of 1.0, with the 10-year yield 0.47 point higher instead of 0.18.
- After 2026. If the buildout holds at its late-2026 share of GDP and the stock gain holds, the effect on real GDP builds a little further and levels off: by the end of 2028 real GDP is about 0.75 percent higher and unemployment about 0.45 point lower, with the funds rate about 0.7 point higher. The buildout alone gives about 0.45 percent and 0.25 point; if it keeps growing by half a percent of GDP a year, about 0.7 percent and 0.35 point. When people see a bigger boom coming, the 10-year yield rises right away and the extra lift is smaller: by the end of 2028 a buildout that keeps growing leaves GDP only about 0.1 percent higher than one that holds steady with forward-looking expectations, against about 0.25 percent with backward-looking ones.
How the data center case is built
The buildout. Three series, each measured as its excess over a log-linear trend fitted to 2015–2022: BEA’s private fixed investment in computers and peripheral equipment, BEA’s investment in communication equipment, and the Census Bureau’s construction spending on data centers. Software is excluded. The excess was about zero through early 2024, because computer spending fell below its trend after the pandemic surge, then rose to 0.7 percent of GDP at the end of 2025 and 1.3 percent in 2026Q3. Computers are most of it. As a check, the five largest cloud companies’ capital spending rose about $760 billion a year from 2022 to 2026Q3, more than the $510 billion rise in these three series, as it should, since theirs includes data centers abroad, land and other spending.
The imports. Each kind of spending gets a plausible direct import share, rather than one measured from the trade data, since both sides of that comparison are noisy and nothing can be more than fully imported: about 90 percent for computers, 80 percent for communication equipment and 11 percent for data-center construction (shares of 85, 75 and 10 percent scaled so the total over 2023–2026 is 75 percent). BEA’s 2017 input-output tables put the average import content of these categories at about 40 percent for equipment and 4 to 7 percent for structures, but that average includes domestic installation, systems work and retail and wholesale markups, which matter much less for a boom that is mostly server hardware bought straight from assemblers. The trade data point higher still: imports of data-center goods rose 1.1 times as much as the investment, net of the matching rise in exports, which is why 85 percent is an alternative. The low alternative is 50 percent. The overall share rises over time with the mix, because computers grew fastest.
The stock-market gain. Each company’s market value is its shares outstanding, from its SEC filings, times its closing price. The rest of the market is the Federal Reserve’s Financial Accounts measure of the market value of all U.S. corporate equity, less the seven. The gain is what the seven are worth above what they would have been worth had they moved with the rest of the market since November 30, 2022. The gain is treated as investors expecting higher earnings from these companies, not as investors accepting lower returns on stocks in general. So in the counterfactual investors expect slower growth of dividends, by just enough to make household stock wealth lower, quarter by quarter, by that share of market value than in the same case with only the buildout taken out, and that lower expectation stays after 2026. In FRB/US expected dividend growth enters only household stock wealth, so the gain works through the wealth effect on consumer spending and does not make it cheaper for businesses to raise money; with forward-looking expectations it is what investors believe about earnings, not earnings that arrive. With the switch on, the without case therefore has exactly the measured gain taken out. Attributing all of it to the buildout is generous, since it also reflects the companies’ other businesses and AI expectations in general, which is why there is a switch.
Expectations and the productivity payoff. Forward-looking agents expect the boom to continue at its late-2026 share of GDP through 2031 and fade by 2035; if they expect it to end in 2027 instead, the forward-looking effects shrink to near zero. In the setting that turns forward-looking in 2026, the model runs backward-looking through 2025 and the news arrives at the start of 2026. A productivity payoff raises trend productivity growth from 2027 and is foreseen from the moment expectations turn forward-looking. In FRB/US, a payoff that is believed in advance makes the years before it slightly weaker: forward-looking wage setters build the coming gains into wages before they arrive, so wage growth and inflation run higher, interest rates rise and housing falls. The runs with the payoff and the stock gain together are solved in small steps, and every combination solves. The forward-looking runs let the Fed rule learn the neutral interest rate (the Board’s option), because with it fixed the productivity cases mostly do not solve and the rest drift.
The counterfactual. The model reproduces the Board’s data package exactly from 2023Q1 onward: history through the second quarter of 2026 and, after that, the projection in the Board’s data package, which follows the FOMC’s Summary of Economic Projections. The estimated effects are then applied to actual data through the third quarter of 2026. The without case takes out the excess investment, its imports and the chosen extras from 2023Q1, with the Fed’s response to the boom following the Board’s inertial rule and fiscal policy following the Board’s rule that raises taxes as debt rises. Unlike the other cases, this one keeps that rule on throughout, the setting the Board’s package ships with, since its paths run past 2026; it makes little difference, because with backward-looking expectations the boom moves the federal budget balance by at most about 0.04 percent of GDP through 2026, leaving the rule almost nothing to react to. The buildout alone can be approximated on the Econvitals simulator with its business-investment shock at 75 percent imported, the inertial rule, the debt rule and backward-looking expectations (the simulator’s business-investment chart includes the added investment itself; this case charts other business investment). Three changes to the Board’s model matter. In FRB/US, business investment and non-oil imports are equations that close gaps toward desired levels, so a shock added to either is partly taken back; each is split here into a data-center piece, set from outside the model, and an “other” piece that runs the Board’s equation unchanged. FRB/US builds real GDP with fixed estimated weights, which for a mostly imported shock would cut the domestic content of each dollar from 25 cents to 14 to 16; the data-center pieces enter at current shares instead. And data-center capital is not counted as productive capacity, so potential output does not rise automatically; a productivity payoff is a separate choice.
After 2026. The runs are extended through 2028 with the buildout either held at its 2026Q4 share of GDP or rising by half a percent of GDP a year. With backward-looking expectations the stock gain is held at its latest share of market value; with forward-looking ones investors’ higher expectation of earnings growth is held at its late-2026 size, so the gain fades a little. Forward-looking agents see the chosen path coming: from 2023, or from the start of 2026. These lines are projections, drawn dotted after the last quarter of data.
The data. Actual data through 2026Q3. For 2026Q4, real GDP, unemployment and core inflation come from the median forecasts in the Philadelphia Fed’s August 2026 Survey of Professional Forecasters; the funds rate and the 10-year yield are quarter-to-date averages; the dollar is October’s; net exports are held at their 2026Q3 share; and the investment series continue their past year’s growth. This snapshot, as of November 9, 2026, uses the August survey, the latest then.
- The import share is the biggest judgment. With the stock gain, the buildout raises real GDP by 2026Q3 about 0.5 percent at 85 percent imported, 0.6 percent at 75 percent and 1.0 percent at 50 percent. If BEA undercounts the investment, the true share is lower and the contribution larger.
- What the model cannot see: electricity prices and power capacity, the very fast depreciation of servers and chips, bottlenecks that have driven up the prices of chips and other hardware, and any productivity gains already showing up in the data before 2027.
Sources
Federal Reserve Board, FRB/US Project, the Python package (pyfrbus 1.1.1) and the data package (vintage of November 4, 2026, history through 2026Q2); the Board’s policy rules page; Brayton, Laubach and Reifschneider, “The FRB/US Model: A Tool for Macroeconomic Policy Analysis” (FEDS Notes, 2014). For the Recovery Act: the Bureau of Economic Analysis, “Effect of the American Recovery and Reinvestment Act on Selected Federal Government Sector Transactions” (revised August 2013). For the invasion of Ukraine: the Energy Information Administration’s NYMEX futures and WTI spot prices; BEA, PCE price indexes (Table 2.8.4); the FAO Food Price Index; Presno and Prestipino, “Oil Price Shocks and Inflation in a DSGE Model of the Global Economy” (FEDS Notes, 2024); Kilian and Patel, Federal Reserve Bank of Dallas (2023). For the tariffs: the Treasury’s Monthly Treasury Statement and Daily Treasury Statement; the Census Bureau’s trade data; Minton, Ray and Somale, “Detecting Tariff Effects on Consumer Prices in Real Time, Part II” (FEDS Notes, 2026); Chair Powell’s press conference of March 18, 2026. For the supply side of the 2017 tax cuts: Barro and Furman, “Macroeconomic Effects of the 2017 Tax Reform” (Brookings Papers on Economic Activity, Spring 2018). For the data center case: Bureau of Economic Analysis (national accounts table 5.5.5U, private fixed investment in equipment by type, and international transactions); Census Bureau (construction spending, data centers, and trade by end-use category); Bureau of Labor Statistics; Federal Reserve Board (H.15 interest rates, the broad real dollar index, the Financial Accounts of the United States); Federal Reserve Bank of Philadelphia, Survey of Professional Forecasters; company filings with the SEC. The model is public domain; the replay assumptions are mine. Every chart downloads as a PNG or as the data behind it.
Replication package (zip, 4.4 MB): the code and the Federal Reserve’s model to rebuild every case on this page.