Apollo’s ‘Zero Financial Payoff’ AI Claim, Explained | Zeniteq
Loading
Loading article
Apollo’s ‘Zero Financial Payoff’ AI Claim, Explained
Torsten Slok’s analysis shows a financial divide between the Magnificent Seven and the rest of the S&P 500, raising a harder question about enterprise AI returns.
Condensed Summary (98 words)
Apollo chief economist Torsten Slok argues AI has delivered "zero financial payoff" to corporate America outside technology. His analysis reveals a stark divide: the Magnificent Seven tech companies show rising revenue per worker and profit margins, while the remaining S&P 500 companies remain flat. Published in October 2026, Slok's findings challenge the enterprise AI investment narrative and suggest broader productivity gains remain absent. However, the comparison uses aggregate indicators that cannot isolate AI's specific contribution or evaluate individual deployments, leaving open questions about whether cheaper models can translate into measurable business improvements for non-tech companies.
Apollo chief economist Torsten Slok says AI has delivered “zero financial payoff” to corporate America outside technology. His evidence is a split in financial performance: revenue per worker and profit margins are rising for the Magnificent Seven, while the remaining S&P 500 companies show flat trends.
That is a consequential challenge to the enterprise-AI investment story. It is also a macroeconomic interpretation of group-level indicators, with a narrower reach than its headline suggests. The useful question for an AI budget owner is whether cheaper models can translate into measurable revenue or margin improvement.
Apollo’s Comparison Starts With Seven Companies
Slok divides the S&P 500 into the Magnificent Seven and everyone else. As TechCrunch explains, those seven companies are Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla. The other companies form what he calls the S&P 493.
This grouping matters. It compares seven giant companies with the remainder of a large-company stock index. It is not a clean separation between every technology company and every nontechnology company, or between businesses using AI and businesses avoiding it.
The two groups can also occupy different positions in the AI economy. Selling computing infrastructure, cloud services or AI-enabled products creates different financial opportunities from buying those services to improve an existing operation.
I’d read Apollo’s comparison as a challenge to the claim that AI’s financial benefits have already spread broadly. Using it to judge a particular insurer’s claims-processing system or a manufacturer’s maintenance tool would require deployment-level evidence that these charts do not provide.
Revenue per Worker and Margins Show the Divide
Apollo’s charts present a parallel pattern: the leading group is generating more revenue per worker and improving margins, while the comparison group is broadly flat.
Roughly $550,000 in 2022, rising to nearly $800,000
Around $550,000, with little change
Profit margins
Rising since 2022
Broadly flat, around 10%
The dollar figures and approximate margin level above come from Business Insider’s description of Apollo’s charts. The opened sources do not provide enough methodological detail about weighting or the underlying datasets to reconstruct the comparison independently.
Revenue per worker divides a company’s revenue by its workforce. A rise can reflect employees producing more valuable work, but it can also reflect higher prices, stronger demand, a different product mix or lower headcount. It is a useful business indicator without being a direct measurement of AI-generated efficiency.
Profit margins address a different question: how much revenue remains as profit after the relevant expenses. An AI deployment could improve a margin by reducing operating costs or supporting higher-value sales. Integration expenses and recurring model charges could push in the other direction.
Taken together, the charts support Slok’s observation that financial improvement is concentrated in the seven-company group. They leave the explanation open. Neither a rising line nor a flat one separates AI’s contribution from the other forces affecting a business.
Aggregate Productivity Measures Something Different
Slok’s third point concerns the wider economy. Apollo’s note says there are still no signs of AI boosting aggregate productivity growth.
Economy-wide productivity, commonly discussed as output per hour worked, is different from corporate revenue per employee. Revenue is measured in dollars and can rise with prices. Productivity measures aim to capture the amount of output generated from labor inputs. Profit margins measure another relationship again: profit relative to sales.
These indicators provide complementary perspectives, but they should not be treated as three interchangeable AI scorecards.
Successful deployments could be too small to move an aggregate measure, or their gains could be offset by weakness elsewhere. Conversely, a productivity improvement across the economy would still require analysis before attributing it to AI.
Those are possible explanations, not findings established by Apollo’s charts. The missing evidence is a link between deployment spending, operational changes and financial outcomes. Group averages cannot tell a buyer which workflow will pay back its implementation costs.
Token Spending Makes the Debate Concrete
Slok also argues that mainstream companies are pulling back and increasingly treating token spending as a cost that weighs on margins. That is his assessment in the Apollo note, rather than a separately documented spending survey in the opened evidence.
Token spending is the recurring usage expense associated with running models. Its financial effect depends on what the work accomplishes.
A system that answers more customer requests automatically might reduce the cost of serving each customer. A system that generates extra drafts, checks and retries could increase usage without changing the final business outcome. The distinction is what happens after the model produces its output.
Cheaper inference helps by reducing one input cost. It does not automatically:
Create additional customer demand.
Turn time saved into lower operating expenses.
Remove integration, supervision or correction work.
Improve the quality of a consequential business decision.
My view is that a lower model bill is best treated as room to improve a business case, not proof that the business case works. If an organization must spend heavily checking the output, token prices may be a relatively small part of the total expense.
That shifts the vendor conversation toward the cost of completing an acceptable task. Buyers ultimately need a useful result, not simply a larger allowance of model output.
Cheaper Models Need a Measurable Business Case
Slok’s forward-looking question is whether increasingly commoditized models can deliver higher revenue growth and profit margins, as Business Insider reports.
A practical implication of his analysis is to evaluate deployments against a business baseline. The following framework is an interpretation of the evidence, not a testing method prescribed by Apollo.
Proposed payoff
Baseline to establish
Evidence that would support the claim
Higher revenue
Revenue or conversion from a comparable workflow before deployment
An incremental improvement after accounting for pricing, demand and sales changes
Lower operating costs
Cost of completing an acceptable unit of work
Lower total cost with quality and service levels maintained
Better margins
Revenue and relevant expenses over a defined period
A sustained improvement after implementation and recurring costs
The measurement period matters. Setup costs can make an early pilot look expensive, while a short demonstration can omit the maintenance burden that follows. A credible evaluation should distinguish initial spending from recurring expenses and make its assumptions explicit.
The cost calculation should include model usage, integration, data preparation, oversight, corrections and ongoing maintenance. Existing software and infrastructure expenses also belong in the comparison when the deployment changes them.
Time savings require particular care. Saving an employee an hour may create capacity, but the financial consequence depends on how that capacity is used. More completed work, reduced overtime and lower contractor spending produce different outcomes from an hour that remains available but unused.
A comparable team or workflow can help distinguish an AI-related change from a wider business trend. Such comparisons have their own limitations, but they address attribution more directly than an index-level chart.
For investors, Apollo’s split raises a question about how widely AI gains are spreading. For buyers, it creates a concrete budgeting standard: show what changed, explain why AI deserves the credit, and count the full expense. Cheaper intelligence becomes financially meaningful when those numbers improve.
Frequently Asked Questions
3 questions
1
Does Apollo’s Analysis Prove Every Non-Tech AI Project Failed?
No. It compares broad company groups and aggregate indicators. Evaluating an individual deployment requires evidence about its costs, operational effects and financial results.
2
What Is the S&P 493?
It is Slok’s shorthand for the S&P 500 excluding Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla. It is not a separately defined nontechnology sector.
3
Would Cheaper AI Models Resolve the Payoff Problem?
They could improve the economics by lowering usage costs. A deployment still needs measurable benefits that exceed its full costs, including integration, oversight and maintenance.
