Wharton paper warns: AI must boost productivity by 2.7 times to prevent tech company bankruptcies

```

A new study quantifies the tech giants’ AI gamble as a matter of life and death arithmetic: either achieve an unprecedented surge in productivity, or face bankruptcy risk.

According to a new working paper released by the National Bureau of Economic Research (NBER), Wharton School financial economist Jessica Wachter and Point72 hedge fund’s Jonathan Wachter jointly modeled that Amazon, Alphabet, Microsoft, Meta, and Oracle’s combined capital expenditure will reach $381 billion in 2025, and is expected to rise to about $755 billion in 2026—more than double the 2024 level. The authors estimate that by 2027, this number will further climb to around $1.1 trillion.

The core conclusion of the paper points directly to the inner logic of this gamble: This scale of expenditure is only rational if the AI industry achieves about a 2.7-fold leap in productivity; otherwise, these companies will “face bankruptcy risk.” Meanwhile, AI investment now accounts for an estimated 14% of total private fixed investment in the US, and 2.4% of GDP, surpassing the roughly 1.5% peak during the late 1990s telecom boom. Its magnitude is now sufficient to sway the entire economic trajectory.

Unprecedented Scale of the Gamble

The paper uses a “rare productivity boom” model to reverse-engineer the spending behavior of tech giants, trying to clarify the expectations implied behind this investment frenzy. The results show that for a 2.7-fold productivity increase to be realized in about five years, it would surpass any comparable period in economic history.

The paper cites two historical references as illustration: It took about 60 years in the US railroad era to nearly triple per capita GDP; during the full IT boom from 1995 to 2005, productivity only increased by 1.5 times. By this measure, the speed and magnitude of tech giants’ current bets are both unprecedented.

If the gamble pays off, the model predicts that by 2030, cumulative US GDP growth will be boosted by an additional 5 to 58 percentage points. The wide range itself reflects the high uncertainty of the prediction.

AI Investment Has Become an Economic Pillar

Data in the paper shows that AI investment’s penetration of the US macroeconomy has far exceeded common market expectations. In the second half of 2025, AI contributes roughly one-fifth of real GDP growth; if AI-related expenditures are excluded, overall business equipment investment would turn negative.

In terms of share, AI investment in US private fixed investment has jumped from 3.3% in 2022 to a projected 14%, and at 2.4% of GDP has surpassed the peak of telecom investment in the previous tech bubble. This means any substantial shrinkage in AI investment would bring a quantifiable downward impact to the overall economy.

The authors clarify the scope of applicability for their conclusions, and investors should note two important caveats when interpreting this paper.

First, the $1.1 trillion capital expenditure estimate for 2027 is the authors’ own, based on a bottom-up approach; no company has released 2027 capital expenditure guidance, and this number does not come from official company disclosures.

Second, “bankruptcy risk” is an argument based on “revealed preference”—that at the current level of expenditure, a substantial productivity jump is a necessary rationale for these outlays—not a prediction that these companies will actually go bankrupt. In other words, the paper describes the inner logic of the gamble rather than judging the outcome.

Cheap Alternative Models Intensify Pressure on Frontline Labs

Against this macro backdrop, frontline AI labs are also facing direct challenges from low-cost competitors. An independent test compared DeepSeek V4 Pro and GPT-5.5 Pro on precision tasks, and DeepSeek won by 38 to 33 in four text tests, using Grok as the evaluation model.

However, the methodology of this test faced widespread criticism on Hacker News: critics pointed out the experiment only covered four crudely designed tasks, lacked reproducible procedures, and the evaluation model was obsolete. Still, the persistent user-reported cost gap remains a more substantial signal—one commenter noted that in another vulnerability scanning test, DeepSeek’s operating cost was about one-tenth that of GPT Pro.

This controversy reflects deeper structural pressures in the market: as corporate clients increasingly seek to cut token costs and are more inclined to adopt open-source alternatives like DeepSeek, frontline labs’ pricing strategies and investor narratives are being challenged. The productivity threshold revealed in the Wharton paper is the ultimate quantitative expression of this pressure—if the assertion that cheap models are “good enough” stands up to scrutiny, the justification for frontline labs keeping a high premium faces a fundamental challenge.

Risk Warning and DisclaimerThe market has risks; investment should be cautious. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article fit their particular circumstances. Investing accordingly is at your own risk. ```