San Francisco Fed: AI Spending Under Weak Total Factor Productivity, Growth from Labor and Capital Inputs During Internet Boom Clearly Better
Bloomberg cites updated total factor productivity estimates from the San Francisco Federal Reserve, indicating poor results; in contrast, growth from labor and capital inputs during the internet boom was clearly better. This comparison circulates among macro accounts, juxtaposing the path after the release of ChatGPT with the information technology cycle driven by browsers at that time.
The San Francisco Fed's May 2026 economic letter notes that labor productivity (output per hour) has seen a solid increase in recent years, but total factor productivity, excluding equipment and technology inputs, has only recorded relatively modest growth. As of Q4 2025, the model estimates the probability of being in a high productivity growth state at approximately: 57% for labor productivity and 21% for total factor productivity. Neither is strong enough to conclude that the economy has entered a sustained high productivity regime.
The letter likens the current divergence where labor productivity is outpacing total factor productivity to the early period of the late 1990s information technology boom: at that time, investment was already underway, but official efficiency figures had not fully caught up. The authors write that AI-related investments may enable workers to use better tools through new software and computing power, but broader efficiency gains have yet to materialize. Analysis cites four quarters up to Q2 2026: general total factor productivity increased by about 1.1%, while utilization-adjusted total factor productivity decreased by about 0.44%, with capital inputs growing at an annual rate of about 3.2% during the same period.
Demand-side figures appear ahead of supply-side ones. Tracking by the St. Louis Fed shows that generative AI has entered GDP through investments in hardware, software, R&D, and data centers; by August 2025, about 55% of Americans and 37% of workers had used such tools. Mark Zandi of Moody's Analytics states that the broad technology sector accounts for about 10% of GDP but represents about half of the S&P 500 market value, contributing approximately two-thirds of the stock price increase since the emergence of ChatGPT; growth is more driven by demand-side construction, while productivity has not yet shown a synchronous increase.
Corporate surveys provide a micro-level comparison: AI adoption among typical U.S. companies is widespread but shallow, with most adopters renting upstream intangible capital rather than building it themselves; executives perceive productivity to be higher than measurable gains, with recent net employment impact being small, and layoffs more concentrated in routine clerical positions at large firms, with jobs shifting towards technical roles. The Fed's background paper also states that generative AI has characteristics of a general-purpose technology and "the invention of methods of invention," but diffusion to measurable labor productivity will take a long time.
In market mechanisms, buyers treat capital expenditures as future efficiency from cloud vendors, model labs, and thematic stock funds, while sellers are upstream providers of accelerators, server rooms, and electricity. The event is driven by updates to total factor productivity and comparisons to the internet. Funds flow from software subscriptions, corporate human resource budgets, and venture capital into chips and data centers. Beneficiaries include upstream entities that can bill and valuation accounts, while those measuring whether the macro account has grown using total factor productivity are under pressure.
Source: Public information
ABAB AI Insight
Total factor productivity writes the "how much more was produced with the same labor and capital" as a residual. The residual not rising does not mean no one is making money; it only indicates that profits are occurring in identifiable factor sellers. In the 1990s, Greenspan did not trust the then-weak official productivity figures because IT investments had already spread, and later revisions moved the acceleration point forward. This time, investment categories are more concentrated: accelerators, cabinets, substations, but the adoption method is to rent upstream intangible capital, with capital deepening recorded in Nvidia and cloud vendors, rather than in every industry using chat windows.
The path of money is thus replacement rather than exogenous creation. Companies change their original payments to SaaS and budget allocations to model interfaces and cloud bills; venture capital reclassifies project funding as pre-paid computing power; public companies reclassify buyback limits as capital expenditures. Upstream gross margins and market values expand first, while downstream wages and total factor residuals lag behind. Moody's writes technology demand as a growth engine, while the San Francisco Fed describes total factor productivity as an unconfirmed regime shift; both sets of accounts exist simultaneously: one for stocks and one for potential output.
Similar lags appear when electricity replaced steam, computers entered offices, and the internet entered retail: equipment is installed first, followed by organizations, processes, and complementary skills, with the Solow paradox recurring. The software industry has contributed significant labor productivity and total factor acceleration since 2017, but that segment comes more from traditional software capital rather than generative interfaces themselves. The current industry is in a phase of investment expansion and measurement lag, not yet at the stage of confirming high growth regimes for labor and total factors together as in the late 1990s.
The essence is the transfer of pricing power. Segments that can invoice tokens and cabinets first take revenue, while segments unable to write efficiency into national accounts temporarily only see cost structures relocating. The mechanism is: when new tools primarily enter the production function in rental form, it statistically first appears as an increase in capital services, and the residual will only rise once organizational restructuring is complete; until then, the most certain macro fact is that the weight of cake distribution has changed, rather than the cake having been measured to grow larger.
ABAB News · Law of Cognition
Capital expenditures can precede productivity, and market value can even more so.
Rented intelligence records efficiency in the profits of the lessor.
Before the cake has been measured to grow larger, the re-slicing has already completed the first round.