Apollo’s “Zero Financial Payoff” Note Turns S&P 493 Charts Into an AI Verdict, but the Flat Productivity Line Rests on a 1.5-Point Adjustment

TechCrunch’s In Brief of Sunday, October 25 said AI “hasn’t benefited the bottom line of any big corporations outside the tech industry,” citing a note by Apollo’s chief economist, Torsten Slok, and its headline widened that to “most companies.” The note is “No Signs of AI Payoff Outside the Tech Sector,” published October 24 in Apollo’s Daily Spark. It rests on five charts: revenue per worker and profit margins for the S&P 493 against the Magnificent 7, productivity growth, the concentration of AI spending, and the share of top AI spenders that are venture-backed. Two earlier notes, from October 20 and October 22, carry the token-spending chart and the first appearance of the venture-capital chart.

I read all three notes, decoded the data behind the productivity chart from the chart PDF, rebuilt its numbers from the San Francisco Fed and Bureau of Labor Statistics series, and checked the spending and adoption claims against Ramp’s public index, the Census Bureau and an NBER executive survey. Where I read a value off a chart image, I say so. Three results:

• The productivity zero is one of three measures. The chart’s flat line is utilization-adjusted total factor productivity (TFP, the part of output growth that more capital, hours and labor quality do not explain), which the San Francisco Fed’s own summary puts at -0.07 percent for the past four quarters. The same summary line gives 1.42 percent before the adjustment, and BLS output per hour has grown at or above 2.0 percent a year in each of the last ten quarters.
• The pullback is two flat months. Apollo’s chart of Ramp data shows token spending up 1 percent in August and 0.2 percent in September, in a series that ran from -2 to 58 percent over the previous 18 months. Ramp’s own page shows total token spending up 23.0 percent in the week ending October 11 against the same week a month earlier.
• Ramp is not the S&P 493. Its index covers more than 85,000 businesses on its own card and bill-pay platform, and its published token metrics come from a subsample that Ramp says “is not representative of Ramp’s total AI spend volume.” Ramp counts 55.9 percent of its businesses as AI adopters in September; the Census Bureau’s representative survey found 17 to 20 percent of US businesses using AI between December and May.

None of this shows that AI pays off. It shows that “zero” is a stronger word than these charts can carry.

What the note claims, and what the headline added

The table lists each claim, the chart behind it, and whether the data can be checked in public.

The strength of the claim changes between notes. On October 20 Slok wrote that returns on AI investments were “either arriving far slower than projected or failing to materialize at all,” and that “the AI payoff for firms has been slow to materialize.” On October 24 the first sentence became “zero financial payoff to corporate America outside of tech,” while the last one says only that “there is no evidence that revenue growth or margins are improving outside the tech sector.” A claim of zero and a claim of no evidence are different findings, and TechCrunch’s lede turned an aggregate into a universal with “any big corporations.”

The note also supplies its own limit. It says “as long as agentic AI and token spending are confined to a small part of the economy, the macroeconomic effects of AI will be very limited.” If use is that concentrated, flat economy-wide lines are what you would see even when the adopters gain, so they cannot tell no payoff apart from slow spread. One more detail: the note never lists the seven companies. TechCrunch does, defining the S&P 493 as “the S&P 500, minus Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla.”

Chart 3 rebuilt from public data

What the chart plots

Chart 3 is titled “No signs of AI boosting productivity growth” and plots two year-over-year lines, shown from 1985 to the second quarter of 2026. Its note says TFP is “the utilization-adjusted series from the San Francisco Fed (Fernald, 2014),” and it cites BLS for output per hour. The PDF holds this chart as vector paths, so I read the plotted values out of it: the paths carry 314 points per line, back to 1948, although the chart displays only 1985 onward.

The grey line matches BLS’s nonfarm business output per hour, series PRS85006091 (percent change from the same quarter a year ago), with a correlation of 0.9999 over the 38 quarters from 2017 Q1 to 2026 Q2 and a mean absolute difference of 0.03 points. The green line matches the series I rebuilt from the San Francisco Fed workbook (the four-quarter mean of its annualized quarterly rates) with a correlation of 0.999 over 314 quarters, against only 0.28 for the unadjusted TFP series, so the chart is plotting the adjusted one. The last few points differ by up to about 0.3 points: the chart ends near -0.3 where the workbook says -0.07. That looks like an older data vintage and not a different series.

The three measures, from the workbook

The workbook’s readme says it was produced on October 1, 2026, and its quarterly sheet ends with summary rows of average annual growth. The table shows them, with the utilization adjustment broken out, and the script below reproduces every figure.

Run the script with Python 3.13 after pip install pandas openpyxl requests (I used pandas 3.0.5, openpyxl 3.1.5 and requests 2.34.2). The BLS call uses the public v1 API, which needs no key.

"""prod_check.py: re-derive the productivity numbers behind chart 3 of Apollo's Oct 24 2026 note from public data."""
import io

import pandas as pd
import requests

UA = {"User-Agent": "Mozilla/5.0"}

# 1. SF Fed (Fernald) quarterly TFP workbook: annualized quarterly growth rates, percent.
url = "https://www.frbsf.org/wp-content/uploads/quarterly_tfp.xlsx"
raw = requests.get(url, headers=UA, timeout=60).content
q = pd.read_excel(io.BytesIO(raw), sheet_name="quarterly", header=1)

# The workbook ends with summary rows ("2004:4-2019:4", "Since 2019:4", "Past 8 qtrs", "Past 4 qtrs").
cols = ["dLP", "dtfp", "dutil", "dtfp_util"]
labels = {"2004:4-2019:4": "2004:Q4-2019:Q4", "Since 2019:4": "Since 2019:Q4", "Past 8 qtrs": "Past 8 quarters", "Past 4 qtrs": "Past 4 quarters"}
summ = q[q["date"].isin(labels)].set_index("date")[cols].astype(float).round(2)
summ.index = [labels[i] for i in summ.index]
summ.columns = ["output/hour", "TFP", "utilization adj.", "TFP util-adj."]
print("Average annual growth, percent (rows from the SF Fed workbook)")
print(summ.to_string())

# Labor productivity = TFP + capital deepening + labor quality (identity in the workbook's README).
last = q[q["date"].astype(str).str.match(r"^\d{4}:Q\d$")].tail(4)
for c in ["dY", "dhours", "dk", "dLQ", "alpha", "dLP", "dtfp"]:
    last[c] = last[c].astype(float)
cap = (last["alpha"] * (last["dk"] - last["dhours"])).mean()
lab = ((1 - last["alpha"]) * last["dLQ"]).mean()
print("\nPast 4 quarters (2025:Q3 to 2026:Q2): output/hour %.2f = TFP %.2f + capital deepening %.2f + labor quality %.2f" % (last["dLP"].mean(), last["dtfp"].mean(), cap, lab))

# 2. BLS nonfarm business output per hour, percent change from the same quarter a year ago (series PRS85006091), the grey line in chart 3.
body = {"seriesid": ["PRS85006091"], "startyear": "2024", "endyear": "2026"}
r = requests.post("https://api.bls.gov/publicAPI/v1/timeseries/data/", json=body, timeout=60).json()
rows = sorted((d["year"] + d["period"].replace("0", "", 1), float(d["value"])) for d in r["Results"]["series"][0]["data"])
print("\nBLS output per hour, % change from a year earlier")
print("  ".join(f"{p}: {v:.1f}" for p, v in rows))
print("mean of these %d quarters: %.2f, minimum: %.1f" % (len(rows), sum(v for _, v in rows) / len(rows), min(v for _, v in rows)))

It prints:

Average annual growth, percent (rows from the SF Fed workbook)
                 output/hour   TFP  utilization adj.  TFP util-adj.
2004:Q4-2019:Q4         1.48  0.49              0.01           0.49
Since 2019:Q4           2.29  1.10              0.27           0.83
Past 8 quarters         2.58  1.18              0.91           0.27
Past 4 quarters         2.51  1.42              1.49          -0.07

Past 4 quarters (2025:Q3 to 2026:Q2): output/hour 2.51 = TFP 1.42 + capital deepening 1.16 + labor quality -0.06

BLS output per hour, % change from a year earlier
2024Q1: 3.3  2024Q2: 3.2  2024Q3: 2.9  2024Q4: 2.3  2025Q1: 2.0  2025Q2: 2.1  2025Q3: 2.5  2025Q4: 2.5  2026Q1: 2.9  2026Q2: 2.2
mean of these 10 quarters: 2.59, minimum: 2.0

What the three measures say

The grey line is where the pickup shows. Output per hour grew 2.2 percent in the year to the second quarter of 2026 and was at or above 2.0 percent in each of the previous ten quarters, a mean of 2.59 percent. The workbook’s 2004 to 2019 average is 1.48 percent, so the recent pace is roughly a point faster.

The green line is the one at zero, and an adjustment puts it there. Before the utilization adjustment, TFP growth over the past four quarters is 1.42 percent, nearly three times its 2004 to 2019 pace of 0.49 percent. The adjustment then removes 1.49 points, against 0.01 points on average from 2004 to 2019. The San Francisco Fed’s page states both results in one sentence: “Over the past four quarters ending in the second quarter of 2026, Total Factor Productivity grew at a rate of 1.42% and Utilization-adjusted Total Factor Productivity fell at rate of -0.07%.” The chart’s note discloses that the series is adjusted. It does not say the adjustment is worth about 1.5 points a year right now.

Using the adjusted series is a defensible choice, and the page calls it an improvement over naive TFP as an indicator of technological change. But the page also says the adjustment follows Basu, Fernald and Kimball (2006) and that “other corrections that are not implemented in the quarterly series include allowing for deviations from perfect competition and for various reallocation effects.” A flat adjusted line is therefore a statement about that method as much as about AI.

Output per hour is not TFP either. The workbook’s identity for TFP (output growth minus the share-weighted growth of capital, hours and labor quality) rearranges so that labor productivity growth equals TFP growth plus capital deepening plus labor quality. For the past four quarters (2025 Q3 to 2026 Q2) the script splits the 2.51 percent into 1.42 from TFP, 1.16 from capital deepening and -0.06 from labor quality. Nearly half of the recent pace comes from more capital per hour worked, which is where a data center and equipment build-out would be expected to show up, so the output per hour line cannot be read as a measure of what AI use pays firms either.

Executives, asked directly, give AI a small credit. The revised NBER working paper “Firm Data on AI,” a survey of nearly 9,000 CFOs, CEOs and senior executives in seven countries in summer 2026, says they “report AI adoption has raised annual labor productivity growth by 0.2% over the last 3 years, a modest impact, but expect a substantially larger 0.6% per annum boost over the next 3 years.” That is a self-report, not a measurement. It is small but not zero, and it is much less than the roughly one-point gap in the table, so the aggregate pickup should not be credited to AI on the aggregate data alone.

So chart 3 supports a narrower sentence than the note’s: the utilization-adjusted TFP line shows no pickup, while labor productivity is growing about a point faster than its 2004 to 2019 average and the aggregate data cannot say how much of that is AI. For how hard attribution is at the level of a single technology, see our posts on Red Hat’s AI productivity measurement problem and the Harvard study of 718 firms.

Charts 1 and 2: what a flat S&P 493 line can and cannot show

Bloomberg and Macrobond data are licensed, so I could not rebuild these two charts. They are raster images in the PDF, so I read each line off the image by color, calibrated on the axis labels. The readings should be good to about three thousand dollars on chart 1 and 0.3 points on chart 2, and the year-start positions are approximate to within a quarter.

Flat is a fair word for a range. The S&P 493 margin has stayed between roughly 6.4 and 11.6 percent since 2015, sat between 9.2 and 9.5 percent at each year start from 2023 to 2026, and ends at about 10.1 percent, a bit under a point above those readings. S&P 493 revenue per employee ends near 525 thousand dollars, against a late-2022 high near 547 thousand and a 2019 start near 493 thousand.

What the lines do not show is more important. The chart’s note says only “Inflation adjusted to Q4 2019,” so the deflator and the weighting are unstated. The aggregate also includes banks, energy producers, utilities and retailers whose margins move for reasons largely unrelated to AI, and the note does not try to separate them. And the note’s own mechanism makes zero an odd target. It says mainstream companies are starting to view token spending “as a cost center that weighs on margins.” If AI costs weigh on margins, AI is affecting margins, and a flat line is also what offsetting gains and costs would produce.

The Mag 7 lines are not a clean read of AI payoff either. They put sellers of AI hardware and cloud capacity in the same aggregate as their customers, and the October 20 note’s second chart says return on invested capital for the five largest hyperscalers (Amazon, Google, Meta, Microsoft and Oracle) “is declining, as invested capital grows faster than the earnings it generates.”

The Ramp charts: concentration, the sample, and the pullback

Concentration is the baseline

Chart 4 gives, for each spend category, the split between the top 10 percent of firms and the bottom 90 percent. The slide does not label the unit, but each pair sums to 100, so I read the figures as percent of the category’s spend.

The three non-AI rows matter. The top decile already holds 84 to 92 percent of SaaS, advertising and CRM spend, so a top-decile share in the 90s is what spend looks like across firms of very different sizes. The AI rows sit about 4 points above the SaaS row for the two big-lab and cloud categories (95.6 and 95.7 against 91.8) and about 7 to 8 points above it for neocloud and model serving. That is more concentrated, but it is not a different kind of distribution.

The per-employee figures on Ramp’s public page carry more weight, because they remove firm size: “Among firms paying for AI, the top 1% spent a median of $9,567.53 per employee per month on AI, compared with $15.70 for the median firm.” AI use is concentrated, as the note says. The NBER survey finds the same shape in a different dataset, with “just 2% of firms accounting for over 50% of total spending.”

Who is in the sample

The note’s phrase is “mainstream corporate America.” Ramp’s methodology section describes its sample this way: “more than 85,000 American businesses and billions of dollars in corporate spend using data from Ramp’s corporate card and bill pay platform.” That is a large sample of one fintech’s customers, and Ramp’s post on how it built the index concedes the obvious objection: “our customers are more likely to adopt AI solutions anyway since they already use Ramp, an AI-oriented approach, as their business spend platform. It’s a fair criticism.” It is not a sample of the 493 largest public companies, and nothing on the page says how many of them are in it. Ramp also argues that surveys “may lead to underreporting of actual AI adoption and spend,” so the gap with the Census Bureau’s figure can have both causes. The three measures below span a wide range.

The gaps persist inside sectors and sizes. Ramp has 73.7 percent of Finance and Insurance businesses and 49.9 percent of Retail Trade businesses paying for AI in September, where the Census Bureau’s May figures were 33.9 percent and about 14 percent. For larger firms Ramp reports 65.7 percent in its large-business category, whose cut-offs the page does not give (it says size segments come from models using revenue and employee counts), against the Census Bureau’s 37 percent of firms with at least 250 employees. Definitions and months differ, since Ramp counts any paid AI transaction in a month and Census asks about use in the past two weeks, so these are not like-for-like. The narrower point is that how widespread AI use looks depends on the question and the sample, and none of the three is the S&P 493. The note’s “confined to a small part of the economy” is a statement about dollars, not about adoption. We hit the same denominator problem in OpenAI’s EU agentic report, and covered an earlier release of this index in OpenAI Narrows Anthropic’s Business AI Lead, New Ramp Data Shows.

Chart 5 and the circularity claim

Chart 5 says 80.4 percent of the top 1 percent of AI spenders are venture-backed, against 13.1 percent of the median spender. The October 22 note infers that “the startups buying the most compute are paying for it with venture capital rather than organic earnings.” The chart shows who the top spenders are by funding type, not what they pay with, so the inference is plausible and untested here. The same note says the pattern suggests “the circularity seen among AI hyperscalers and model providers” also reaches token spending; for that side of the story, see our post on Anthropic’s Akamai deal.

The pullback

The October 20 note’s chart is titled Ramp data shows corporate token spending slowing down and plots the monthly change in token spend. The table transcribes its bar labels, in percent.

August and September are weak, but the series has bounced after each earlier dip: 9, 21, 49 and 45 followed November 2025’s -2, and 22, 19 and 42 followed April 2026’s 6. Two flat months may be the start of a plateau. They are also what this series has looked like before.

The note says “mainstream corporate America is pulling back,” but the chart is not split by type of firm, and the only split in the pack is chart 5, where the heaviest spenders are mostly venture-backed. Ramp’s own page, which I read on October 25, reads differently for the latest period. For the week ending October 11 it says “Total token spending rose 23.0% compared with the same week last month,” and total token volume rose 52.4 percent. Adoption also rose, to 55.9 percent in September, “up 0.6 percentage points from August.” One number leans the note’s way: AI’s share of business spending excluding payroll slipped from 5.96 to 5.77 percent in September.

Spend rose less than volume, so the average price per token fell, by about 19 percent on my arithmetic (1.230 divided by 1.524, ignoring changes in model mix). That is consistent with the October 20 note’s argument that demand is “highly price-sensitive,” and it is not by itself a retreat from AI use.

Two caveats from Ramp’s own definitions cut across all of this. The token metrics “track token volumes, spend, and prices against AI models among businesses using Ramp’s Token Spend Management product,” and Ramp says that sample “is not representative of Ramp’s total AI spend volume, though results are directionally similar.” I cannot tell which Ramp series Apollo’s monthly chart uses, because its source line says only Ramp AI Index. Ramp also says it likely underestimates AI adoption and spend, because free tools and personal accounts leave no paid transaction. A shift to free tools or personal accounts could look like falling spend in this data without any drop in use, and a shift to cheaper models would lower dollars per token even as usage grows.

What would settle it

Three tests would do more than another aggregate chart. The first is to compare adopters with similar non-adopters inside the 493, before and after adoption, on margin and revenue per worker. Ramp has shown what that design looks like for hiring. Its June 30 letter reports a working paper that joins Ramp spend data to Revelio Labs workforce records for more than 21,000 US firms. It finds that “these gains are entirely driven by high-intensity adopters” while “low-intensity adopters see no statistically significant change,” that “firms don’t increase their headcount until 6-12 months following adoption,” and that “AI adopters are already larger, more engineering-intensive, more likely to be venture-backed, and faster-growing than non-adopters.” Those are headcount results, not margins, but they show why a whole-index line is a weak test: gains arrive late, concentrate in heavy users, and start from firms that were already different. The second test is to rerun the script above when the next quarter’s GDP and BLS releases land and see whether the utilization adjustment unwinds. The third is to watch Ramp’s October monthly change: one more flat month would make a plateau, and a rebound would not.

What I could not check

I could not rebuild charts 1 and 2 from licensed Bloomberg and Macrobond data, see their deflator, weights or membership rules, or tell how many of the 493 had reported the latest quarter. I could not see the data behind charts 4 and 5, only Ramp’s public summary, or tell which Ramp series the monthly token chart draws on. The chart’s last TFP points differ from the current workbook by up to about 0.3 points, which I attribute to data vintage without having the older file. Everything here is aggregate, so it cannot show that AI pays at any one firm. It shows that this note’s charts do not show zero.