Spend where the lift is: a $381B AI bet, a 5-to-58-point GDP range, and one discipline for your budget.

The five biggest U.S. tech firms spent $381B on AI infrastructure in 2025, and they guide to roughly $755B in 2026, about double in a single year [1]. A new NBER working paper by Jessica Wachter and Jonathan Wachter reads that ramp as a rational bet and asks a simple question back: if the firms putting up the money are not fools, what must they believe? Invert the spend, and the answer is a productivity jump of about 2.7x for the AI sector with each boom (Section 3.3). Run that through their two-sector model and you do not get a forecast. You get a range: 5 to 58 percentage points of extra cumulative GDP growth by 2030 (Table 5).

That range is the whole story. The gap between the moderate case (5.4 points) and the singularity case (58.2 points) is more than tenfold. Across 10,000 simulated paths to 2050, the single most likely outcome is no boom at all, 44% of paths (Section 4). This is a calibrated model, not peer-reviewed, and the authors are explicit that the figures are scenarios, not calendar-year GDP. The operator takeaway is not “AI is coming, spend now.” It is the discipline that follows when the upside is this wide: treat your own AI dividend as uncertain, measure incremental lift instead of assuming it, and fund the pillars that compound regardless of when the boom lands.

The paper, read properly

Start with what this paper is, because the temptation to over-read it is strong. It is a macro-finance calibration exercise, not a sales study, not a field experiment, and not a causal analysis. There is no human-subject sample, no benchmark dataset, and nothing about pipeline, win rate, or any go-to-market outcome. The “data” are the consolidated annual capital-expenditure figures of five firms (Amazon, Alphabet, Microsoft, Meta, Oracle), pulled from SEC filings: 10-K cash-flow statements for 2022 to 2025, 8-K earnings guidance for 2026, and the authors’ own bottom-up estimates for 2027, since no firm has issued formal 2027 capex guidance.

The research question is narrow and sharp. If you assume firms invest rationally and the marginal AI dollar is a break-even (zero-NPV) bet, then the size of the spend reveals the productivity gain those firms must expect. That is textbook q-theory: under constant returns, observed investment is a clean signal of the expected return on installed capital [2]. The authors invert it. Cumulative 2024-to-2027 capex of $2,452B (226 + 381 + 755 + 1,090) against a pre-boom AI capital stock of about $620B implies a boom multiple of e to the xi, which works out near 2.69 (Section 3.3, equation 13).

That 2.7x then feeds a two-sector open-economy model with rare productivity booms, and three scenarios fix the payoff. The constructs matter, so read the definitions carefully. “Cumulative GDP growth from AI” is the one-time level increase in GDP relative to a pre-boom baseline, holding non-AI output fixed to isolate the AI contribution. It is not a forecast of any calendar year’s GDP. The “AI share of the economy” is AI-sector output over total output. With those definitions, the three scenarios land as follows (Table 5): one boom (the moderate case) lifts the AI share to 8.0% and adds 5.4 points of cumulative growth; two booms (the authors’ middle case) reach 19.0% share and 19.7 points; three booms (the singularity case) reach 38.7% share and 58.2 points.

The capex ramp behind all this is genuinely steep. AI-infrastructure spend rose from 0.6% of GDP in 2022 to 1.2% in 2025, with the model projecting 2.4% in 2026 and 3.4% in 2027 (Table 2). For scale, the late-1990s telecom buildout peaked near 1.5% of GDP. As a share of all U.S. gross private fixed investment, the figure climbs from 3.3% in 2022 toward a projected 19.2% by 2027. The authors add an Epstein-Zin asset-pricing layer that maps the boom into the risk-free rate and the equity premium, but the headline asset-pricing numbers (risk-free rate up about 0.5 points, equity premium up about 3 points at risk aversion of 3) are the Panel-B absorbing-limit case, where AI is treated as if it already dominates the economy. That is a ceiling, not a center estimate. Panel A, where AI stays permanently small, moves rates far less.

The numbers, worked

Here is the move worth stealing. Not the GDP figure. The method. The paper takes a spend and backs out the bet hiding inside it. An operator can run the identical test on an outbound program, not to predict anything, but to decide whether the spend is rational at all before approving it.

Read step four out loud before you approve anything. Committing $500,000 against 2,000 accounts is a rational, break-even bet only if the program produces at least a 0.6% account-to-closed-won rate. If your own history shows outbound converting at 1.2%, the program clears its hurdle with room. If history shows 0.4%, the spend is below break-even on its own stated assumptions, and you are implicitly betting on a conversion lift that has not yet shown up. That is the same honest tension the paper flags about AI capex: the spend is rational only if a not-yet-observed jump arrives.

Now stress the economics. Keep N and C fixed, and vary the deal math. Halve the contribution per deal, by a smaller ACV or a thinner margin, and the required win rate doubles.

Both halving routes (half ACV, half margin) land on the identical 1.2%, because only the product of ACV and margin drives contribution. The “effect half as strong” row is the operator mirror of the paper’s own caution: cut the assumed payoff in half, and the spend now needs twice the performance to be rational. State the bet before you fund it, and the budget conversation changes from “do we believe in AI” to “do we believe in a 1.5x lift, and what does our history say.”

A trap to name honestly. The 0.6% required win rate is a per-account hurdle rate, not the probability that any single account closes. It is the break-even average across the whole list of 2,000. Do not read “0.6%” as “each account has a 0.6% chance.” And the threshold is arithmetic, not a prediction that you will hit it. Validate it against your own closed-won history before you trust it.

Why the bet self-justifies

The economic mechanisms the paper applies are well grounded. Revealed-preference q-theory says investment reveals the expected payoff [2] (strength: established; the equivalence result is foundational, though its application to AI is the paper’s assumption-laden step). Rare-events asset pricing says a small per-year probability of a large jump is enough to move the equity premium and the risk-free rate [3] (strength: established; the magnitude here is calibration-dependent and explicitly an upper bound).

But the mechanism a revenue leader needs is behavioral, and the paper does not study it. It is the reason teams over-commit to an uncertain AI dividend in the first place. Once an organization has visibly committed budget, headcount, and reputation to an AI buildout, the prior spend itself biases the decision to keep funding it, independent of forward expected value. People continue an endeavor because of what they have already put in, partly to avoid looking wasteful. That is the documented sunk-cost and escalation-of-commitment effect [4] (strength: plausible as a transferred read; the effect is established in its own literature, but applying it to AI-budget discipline is our inference, not a finding of either paper).

The bridge is exact. The paper’s own caution is that if the boom fails to materialize, the buildout would be the largest misallocation of capital in history. Escalation of commitment is the behavioral path to exactly that outcome: each round of spend makes the next round feel inevitable, and the wide payoff distribution gets quietly collapsed into a single optimistic number nobody re-checks.

The discipline antidote is the whole point of this piece. Treat the dividend as a wide distribution. Hold out a control. Measure the incremental lift before you scale the spend.

Why a revenue team should care

Your AI tooling line is on the same curve as the macro number, doubling year over year, and very often without a single attribution model behind it. The sunk-cost mechanism predicts that the more you have already spent, the less likely you are to ask whether the marginal dollar still clears its hurdle. That is precisely the dollar the zero-NPV test is built to interrogate.

This is where Pipeline Architecture earns its keep, and it sits on two pillars. Market is where you place the bet: which segment, which motion, which spend. Measurement is how you prove it paid. Before a program is funded, state the conversion bet it implicitly makes (the 0.6% in the worked example). Then check that bet against your own history. If you cannot tie the spend to a number, you are escalating commitment, not architecting pipeline.

What to do Monday morning

One move per rung of the maturity ladder, each startable Monday, each tied to Market (where you place the bet) and Measurement (how you prove it paid).

Manual. Run a two-week holdout on your single highest-volume outbound play. Split the next 200 accounts: 100 worked with your current AI assist, 100 worked the old manual way. Track reply rate and meetings booked for both. You now have a baseline for what AI actually adds before you spend another dollar scaling it. Most teams have never measured this and are budgeting off vibes.

Assisted. Audit your AI tool spend against booked pipeline, line by line. List every per-seat license and usage fee. Next to each, write the incremental meetings or opportunities you can attribute to it this quarter. Freeze anything you cannot tie to a number. The paper’s logic applies to your stack: an investment is rational only if the marginal dollar is near zero-NPV, and you cannot know that without the attribution.

Orchestrated. Stand up a standing holdout cohort, not a one-off test. Permanently hold 10% of one well-defined segment out of your AI-orchestrated sequences and keep it on the prior process. Report the lift gap monthly alongside pipeline. This converts “we think the AI motion is working” into a recurring incremental-lift number your CFO can read.

Autonomous. Set a kill-switch threshold before you expand any autonomous agent fleet. Write the rule down: if attributed cost per booked meeting from the autonomous motion exceeds a set figure for two consecutive months, pause and revert. The model’s depreciation assumption is brutal, with AI capital carrying a four-year useful life (delta of 0.25, Table 3). Your autonomous tooling depreciates too. Decide the abandon condition in advance, while you are calm, not after the budget is committed.

Where this lands by industry

Five parameterized archetypes. The parameters are stated assumptions, not claimed facts about any named firm.

Limits and caveats

Be honest about what this does not establish. It is a calibrated model plus revealed-preference inference, not an experiment, and it carries no causal identification. The 2.7x boom size is inferred only by assuming the marginal investment is exactly zero-NPV. The authors note that positive economic profits would imply a larger boom, and that a failed boom would be a historic misallocation. The boom size also depends on chosen inputs: a 0.25 depreciation rate, a 0.15 required return, a 0.50 window probability, a 0.04 post-window probability. Change those and the boom size changes.

The GDP figures are scenarios, not forecasts, and hold non-AI output fixed to isolate the AI contribution. Do not quote any single number as “the AI growth forecast.” The long-run ranges come from 10,000 Monte Carlo paths where the modal outcome is no booms. Model bands are not observed history. The paper is also a non-peer-reviewed NBER working paper, and one author is affiliated with Point72, which did not verify the contents. The transfer risk is the largest caveat of all: this paper has zero go-to-market content, so it cannot support any claim about selling, win rate, or pipeline. The worked example borrows only the method, and the arithmetic is the operator’s own.

What would change the read. For the macro claim, observed economic profits at the named firms (which would imply a larger boom) or a documented failure of the buildout to produce productivity gains (which would expose it as misallocation). For your own spend, a holdout that shows the AI motion clears, or fails to clear, the break-even hurdle you stated up front. The point is not to predict either. It is to measure.

What you learned

The capex ramp is real: $381B in 2025, a forecast $755B in 2026 (Table 1). The implied bet is a 2.7x productivity jump (Section 3.3). The payoff is a coin-flip-wide range, 5 to 58 points of GDP by 2030, with no boom as the single likeliest path (Table 5, Section 4). The transferable tool is revealed preference: a $500,000 program on 2,000 accounts is rational only above a 0.6% win rate, a 1.5x bet on a 0.4% baseline (derived). And the discipline is the same at both scales: name the not-yet-observed jump your spend is betting on, hold out a control, and measure the incremental lift before you scale.

Operators who want that hold-out-and-measure discipline in their inbox every week can Get the Weekly Research.