🔍 Read the full analysis: What The AI Investment Surge Reveals About Real-World Returns on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Apollo economist Torsten Slok’s October 24 analysis found that productivity and financial gains remain hard to see across most non-tech U.S. companies, even as investment in AI infrastructure expands. The finding describes aggregate trends, not proof that AI produces no returns: some workplace trials show gains, while the costs, benefits and financing risks vary by company.
Apollo chief economist Torsten Slok said in an October 24 research note that available data show no visible financial payoff from AI outside the technology sector, even as companies make large commitments to data centers, computing equipment and cloud services. The finding is about broad corporate averages, not proof that AI has no value: productivity data have not yet shown a clear economy-wide lift, while some workplace studies report measurable gains in specific tasks.
Apollo’s note, titled “No Signs of AI Payoff Outside the Tech Sector,” compares the seven largest technology companies with the other 493 firms in the S&P 500. Its charts show revenue per worker and profit margins rising for the seven, while the corresponding measures for the other companies remain broadly flat. Apollo’s reading of the figures puts the seven near $780,000 in inflation-adjusted revenue per employee in the third quarter of 2026, compared with about $525,000 for the other 493. The figures show a gap between groups; they do not establish that AI caused it.
The note also points to San Francisco Federal Reserve productivity measures that, according to Apollo, show no clear sign of an AI-driven productivity increase. Separate Ramp data covering more than 85,000 U.S. business customers suggest AI-related spending is concentrated among a small group: the top 10% of firms account for most spending on model inference, neocloud services and products from OpenAI and Anthropic. Those figures describe Ramp’s customers and should not be treated as a survey of all U.S. businesses.
Corporate disclosures show the scale of the investment side. Meta’s October 22 filing listed roughly $825 billion in future commitments, including leases for facilities and contractual purchases of cloud capacity, servers and network equipment. The amount is not all AI spending, and it is not a bill due at once. Meta also reported $34.91 billion in capital expenditures in the third quarter and forecast $135 billion to $145 billion in capital spending for 2026. Its free cash flow was negative $2.12 billion in the quarter, compared with positive $10.6 billion a year earlier.
These commitments sit alongside reported financing activity elsewhere. The source material says SpaceX borrowed $40 billion to buy AI chips, but does not provide further transaction details. Meanwhile, workplace evidence offers a more mixed picture than the economy-wide indicators: cited studies found customer-support workers resolved 14% more issues per hour with AI assistance, with a 34% gain for novice workers, while retail sales experiments ranged from no increase to a 16.3% rise depending on the workflow. These results concern specific settings, not whole-company returns.
Infographic 01 · The “zero” in numbers
What “zero financial payoff” actually measures.
Apollo’s headline rests on five charts. Here are their numbers, and what they can and cannot prove.
Who pays for AI? Share of spending by the top 10% of firms
Ramp customer data via Apollo · non-AI categories in grey for comparison
Spending is concentrated in every category, not only AI; AI is the most extreme.
What the five charts do and do not show
shown not shown
- Shown: revenue per worker and margins rising for the 7 tech giants, flat for the other 493
- Shown: no AI lift yet in US productivity statistics (San Francisco Fed TFP, output per hour)
- Shown: AI spending sits in few, mostly venture-funded firms (Ramp, 85,000+ US businesses)
- Not shown: that AI caused the gap; rates, tariffs and wages move the same lines
- Not shown: gains inside single functions, which can cancel out in an average
- Not shown: the S&P 493 itself in the Ramp data; those are Ramp’s customers
“The AI story will be decided in the S&P 493, not the Mag 7.”Torsten Slok, Apollo chief economist, 24 Oct 2026 (opinion)
Sources: Apollo Daily Spark (24 Oct 2026) with charts by Bloomberg, Macrobond, SF Fed, BLS and Ramp; Apollo (20 Oct 2026); Ramp AI Index. Values with ≈ are read off Apollo’s charts. Assessment: Thorsten Meyer.
Powered by Thorsten Meyer AI
The Gap Between AI Costs and Returns
The gap matters because companies are committing money to infrastructure before returns are readily visible in broad productivity or financial measures. For executives and investors, the key issue is not simply whether AI can help with a task, but whether savings or additional revenue exceed the cost of computing, software, implementation and staff changes. Apollo’s note argues that many companies have yet to show that balance in their aggregate results.
The financing question grows as spending plans expand. A large commitment can support future capacity, but it can also leave a company with costs that persist even if demand or commercial benefits fall short. The source material reports that Meta’s five-year credit default swaps reached 114.79 basis points, a record, according to Fortune. That figure signals a reported change in the cost of insuring its debt; on its own, it does not establish that Meta faces a default or that its AI investments will fail.
For workers and customers, the results suggest that effects may arrive unevenly. A defined task with clear performance measures can show gains before those gains appear in company-wide productivity statistics. Conversely, a successful trial does not prove that an AI tool will generate a net return after its full costs are counted. The distinction between task-level performance and firm-wide returns is central to interpreting both the optimistic investment figures and Apollo’s caution.
Products Worth Considering
Apple iPad Air 13-inch (M4), 128GB, Wi-Fi 7, Space Gray
Apple Mac mini Desktop Computer with M5 Pro Chip (2026)
Dell UltraSharp 32 U3226QER 4K Thunderbolt Hub Monitor
As an affiliate, we earn on qualifying purchases.
What Apollo’s Comparisons Measure
Apollo’s October 24 note was reported by TechCrunch, PYMNTS and Bloomberg the following day. Its central comparison is between the so-called Magnificent Seven—Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla—and the remaining S&P 500 companies. The comparison describes differences in revenue per worker and margins; it does not isolate AI as the cause. Company size, business models, market position and other economic conditions can also affect those measures.
Slok has framed the issue in terms of whether the marginal cost of computing exceeds the benefit delivered to users. In an October 20 note, he called that tension a central weakness in the optimistic investment case; on October 22, he described venture-funded spending on AI tokens as “circularity.” These are his interpretations of investment and spending patterns, not settled findings about every AI business or deployment.
The investment totals also need careful reading. Meta’s filing combines lease obligations, non-cancelable contractual commitments and potential cloud purchases extending across multiple years. Some commitments relate to infrastructure that supports more than AI, including Reality Labs hardware. Forecasts for other major technology companies’ capital spending are projections, not signed commitments. Comparing a long-term commitment total with a single year’s revenue or capital spending can give a misleading impression of immediate cash costs.
“No Signs of AI Payoff Outside the Tech Sector”
— Torsten Slok, Apollo chief economist, in the October 24 Daily Spark note
What Aggregate Data Cannot Show
Apollo’s “zero” framing is stronger than what the cited aggregate measures can establish on their own. A flat average among 493 companies could conceal firms that gained from AI and others that did not, with the results offsetting each other. The data cited do not separate AI’s effects from other influences such as labor costs, interest rates or tariffs. Apollo’s note supports the narrower conclusion that a broad payoff is not yet visible in those measures.
It is also unclear how quickly task-level gains will become lasting improvements to revenue, margins or output per hour. The cited workplace results differ by job and workflow, and the source material does not provide enough information to compare their study designs, costs or duration. Nor does it establish whether those gains persist when systems are used at scale.
Several financial details remain incomplete. The source material gives a $40 billion SpaceX borrowing figure but no terms, schedule or additional transaction documentation. It also mentions rising insurance costs for debt tied to Meta and CoreWeave without providing a full series for comparison. The size and timing of future AI revenue, infrastructure utilization and eventual returns remain uncertain.
Evidence to Watch in Company Results
The next evidence will come from company filings and operating results that show whether large infrastructure plans translate into sustained revenue growth, lower costs or higher output. Investors and readers will need to distinguish planned spending from cash already spent, and broad company commitments from AI-specific costs. Meta’s forecast of $135 billion to $145 billion in 2026 capital spending makes its subsequent quarterly cash flow and explanations of infrastructure use especially relevant.
Researchers and companies will also need to track productivity over longer periods and across more kinds of work. For individual deployments, useful evidence includes the task being measured, the comparison group, the cost of AI tools and the duration of the result. Until such evidence accumulates, isolated gains should not be treated as proof of economy-wide returns, and aggregate stagnation should not be treated as proof that no firm benefits.
Financing conditions for data-center operators and specialized cloud providers will be another indicator. The source material reports that Meta described fundraising as increasingly challenging and that analysts have raised concerns about scarce capital for neocloud businesses. Further disclosures could clarify how much infrastructure is financed through borrowing, leases or customer contracts. For now, the confirmed picture is one of rapidly expanding commitments and uneven, difficult-to-measure returns—not a settled verdict on AI’s long-term economic value.
Key Questions
Did Apollo prove that AI produces no financial returns?
No. Apollo’s analysis says no payoff is visible in the aggregate measures it examined outside the technology sector. Those comparisons do not isolate AI’s effect or rule out gains at individual companies.
What does Meta’s roughly $825 billion figure represent?
It combines future leases, contractual commitments and potential cloud purchases reported in Meta’s filing. The total covers multiple years, includes some items not specific to AI and is not an amount due all at once.
Are there examples of AI improving work performance?
Yes. The source material cites customer-support research finding 14% more issues resolved per hour, with a 34% gain for novice workers, and retail experiments ranging from no sales increase to a 16.3% increase. These are task- and workflow-specific results, not a measure of economy-wide returns.
Why are company-wide productivity measures not showing a clear AI boost?
The cited data show no clear economy-wide increase so far, but they do not explain why. Gains may be limited to particular tasks, too small to move broad measures yet, or offset by other changes. The available figures do not isolate AI from other economic factors.
What evidence would clarify whether AI spending is paying off?
Useful evidence would connect AI-specific costs to sustained changes in revenue, margins, output or task performance. Results should identify the time period and comparison baseline, and account for implementation and computing costs. Future company filings and longer-term productivity studies may help answer that question.
Source: ThorstenMeyerAI.com
