Apollo Economist Sees No Broad AI Profit Payoff Yet | Zeniteq
Loading
Loading article
Apollo Economist Sees No Broad AI Profit Payoff Yet
Torsten Slok’s analysis challenges claims of an economy-wide AI dividend, but its aggregate charts cannot settle whether individual companies are benefiting.
Apollo chief economist Torsten Slok argues that artificial intelligence has yet to deliver broad financial benefits beyond the Magnificent Seven tech companies. His October 2026 analysis shows rising revenue per worker and profit margins for those seven firms alongside flat metrics for the remaining S&P 500 businesses, challenging claims that AI investments will strengthen performance across the economy. While aggregate data cannot determine whether individual companies have made successful AI investments, the evidence suggests mainstream businesses increasingly treat token spending as a cost center rather than a profit driver, requiring specific evidence linking deployments to measurable financial gains.
Apollo chief economist Torsten Slok says AI’s financial payoff is still missing from broad corporate results beyond the Magnificent Seven. In an October 24, 2026 Daily Spark note, he points to rising revenue per worker and profit margins among those seven companies, alongside flat measures for the remaining S&P 500 businesses.
That divergence challenges a central enterprise-AI promise: that buying access to increasingly capable models will translate into stronger financial performance across the economy. Slok argues that mainstream companies are increasingly treating token spending as a cost center.
His evidence supports a narrower conclusion than a verdict on every deployment. The charts offer a test of whether AI’s benefits are becoming visible across large companies. They cannot tell us whether a particular insurer, manufacturer or retailer has made a successful AI investment.
What Slok Means by the S&P 493
The S&P 493 is Slok’s shorthand for the S&P 500 excluding Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla. TechCrunch’s October 25 coverage identifies those exclusions and attributes the financial-payoff argument to him.
This distinction matters because the S&P 493 is not a formally defined non-tech sector. Removing seven technology-linked giants still leaves other technology businesses in the comparison.
It also creates a more useful question than whether AI companies are growing: Are financial gains spreading beyond the seven companies most closely associated with the technology boom?
My view is that this is the strongest use of Slok’s comparison. Strong results among prominent AI suppliers and platform companies cannot, by themselves, demonstrate that their customers are earning attractive returns. Those are different financial outcomes.
The reverse needs care, too. Rising results for the Magnificent Seven cannot automatically be credited to AI. These businesses have other products, revenue streams and operating decisions affecting their performance.
Revenue per Worker and Margins Measure Different Things
Slok’s Apollo note describes two corporate financial comparisons, followed by a broader productivity indicator:
Work with Zeniteq
Promote your app with us
Reach readers exploring AI tools and workflows. Discover sponsorship and advertising options for your app.
Rising for the Magnificent Seven; flat for the S&P 493
Whether companies generate more real revenue relative to workforce size
Profit margins
Rising for the Magnificent Seven; flat for the S&P 493
Whether companies retain more profit from their revenue
Utilization-adjusted total factor productivity
Slok sees no sign of an AI-driven boost
Whether broader production efficiency is improving
Apollo credits Bloomberg and Macrobond for the corporate charts. The revenue-per-worker chart notes an inflation adjustment to the fourth quarter of 2019. That adjustment helps separate changes in real revenue from changes attributable to inflation.
Revenue per worker is a useful financial productivity proxy, but it is sensitive to more than employee effectiveness. It can rise because of stronger demand, changes in product mix, outsourcing or a smaller workforce. It is not a direct measurement of how much work an employee completes with AI.
Profit margins add another perspective. A company could generate more revenue per employee while spending heavily on software, computing, integration and oversight. Higher revenue efficiency would then coexist with little improvement in profitability.
The available source text provides directional chart descriptions and source notes, without readable numerical series, complete observation windows or weighting methods. It therefore supports discussing the divergence, but not calculating its size or judging how sensitive it is to the chosen period.
Reading the two financial measures together is useful: one tracks revenue relative to headcount, while the other asks how much revenue survives as profit.
Why Token Spending Can Become a Cost Center
Slok argues that venture-backed companies continue spending more on tokens while mainstream businesses are less willing to pay for the latest models. He interprets that split as evidence that AI spending is concentrated in a relatively small part of the economy, limiting its broader effect. These are his conclusions about purchasing behavior, not a company-by-company accounting of spending cuts or returns.
Tokens are the units models process when handling inputs and generating outputs. For usage-based services, token consumption is a billable input. More consumption can indicate heavier use, but it says little by itself about the value created.
An AI system could process more customer-service requests while failing to resolve more of them. A coding assistant could generate more code while leaving review and maintenance costs unchanged. These are illustrative possibilities, not findings from Apollo’s charts.
Cheaper tokens can improve the economics of a useful workflow. They cannot guarantee that the workflow produces something customers value or removes a cost the business actually pays.
Slok also questions whether model capability translates readily into business judgment:
“AI may be a great calculator, but in business, intelligence is mostly judgment.”
He connects that argument to context, accountability, trust and organizational culture in the original note. Those are his explanations for why cheaper intelligence might struggle to produce better financial results; the charts do not directly test them.
I’d separate model purchasing from deployment value. A business can sensibly decline a newer, more expensive model if an existing system already meets its quality requirements. Lower spending in that situation could reflect disciplined procurement, even if AI remains useful.
Apollo’s third chart uses utilization-adjusted total factor productivity, with source notes referencing the San Francisco Fed and the US Bureau of Labor Statistics. Total factor productivity concerns output relative to combined inputs such as labor and capital. The utilization adjustment seeks to account for cyclical changes in how intensively those inputs are used.
That makes it a different measure from corporate revenue per worker. It offers a broader view of efficiency, while corporate ratios describe financial performance.
Neither measure cleanly isolates AI’s contribution. Consider three analytical problems:
Gains and losses can offset each other. Successful deployments at some companies may disappear inside a flat aggregate.
Other business conditions affect the same metrics. Demand, wages, pricing, financing costs and investment can move revenue or margins independently of AI.
The missing comparison is what would have happened without AI. Flat margins could reflect no benefit, or a benefit offset by unrelated pressure.
These possibilities do not refute Slok’s interpretation. They explain why the evidence is better suited to challenging claims of a visible, broad financial uplift than declaring that all individual investments have failed.
The same standard applies if the charts improve. Rising margins across the S&P 493 would strengthen the case that financial gains are spreading, but attributing those gains to AI would still require evidence linking deployments to outcomes.
What Would Make the Financial Case Stronger?
Slok’s comparison gives executives and investors a useful starting point: ask where AI spending appears in the business result.
A practical evaluation should follow a traceable chain:
Define the workflow. Identify the task being changed, such as resolving a support case or processing an invoice.
Measure useful output. Track completed work at an acceptable quality level, including errors and rework.
Count the full cost. Include model usage, software, integration, supervision and ongoing maintenance.
Connect the change to finances. Show whether it creates revenue, avoids a real expense or expands capacity that the business can use.
Use a credible comparison. Account for changes in workload, staffing and demand before assigning the improvement to AI.
This framework is analysis derived from the gap in Apollo’s evidence, not a methodology Slok reports having applied.
Time saved deserves particular care. An employee finishing a task faster may create valuable capacity, but that capacity becomes a financial return only through some further outcome: more work completed, better service, additional sales or reduced expenditure.
For investors, the distinction is equally practical. Token consumption and deployment counts measure activity. Revenue per worker and margins measure business outcomes. Evidence connecting the two is what would turn an adoption story into a persuasive return-on-investment case.
Frequently Asked Questions
3 questions
1
Does Slok Say Every Non-Tech AI Project Has Failed?
He makes a broad argument about absent financial payoff outside tech. His aggregate evidence cannot determine the returns of every individual company or project.
2
Is the S&P 493 a Non-Tech Index?
No. It is shorthand for the S&P 500 excluding the Magnificent Seven, so it still includes other technology companies.
3
What Would Strengthen the Case for AI Returns?
Broader improvements in revenue per worker and margins would help. Credible attribution would also require evidence connecting specific AI deployments to measurable gains after their full costs.
