The Rising Cost Of AI And The Search For Returns

🔍 Read the full analysis: The Rising Cost Of AI And The Search For Returns on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

• Fast, free delivery for dorm and study essentials
• Prime Video and Amazon Music included
• Member-only deals

Try Prime for Young Adults →Free trial for eligible 18–24 year olds

As an affiliate, we earn on qualifying purchases.Hide

Apollo economist Torsten Slok says recent productivity and company data show no clear financial payoff from AI across most major U.S. companies, even as spending on data centers, chips and cloud capacity grows. His October 24 analysis compares seven large technology companies with the other 493 S&P 500 firms; it describes an aggregate pattern, not proof that AI has produced no gains at individual companies.

Apollo’s charts show inflation-adjusted revenue per worker rising among the Magnificent Seven — Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla — while the other 493 S&P 500 companies have remained broadly flat. The note also shows margins near 30% for the seven and around 10% for the rest. Those comparisons describe company performance; they do not isolate AI as the cause of the difference.

The cost of building capacity is visible in company disclosures. Meta reported $34.91 billion in capital expenditures in the third quarter and said it expects to spend $135 billion to $145 billion on capital projects in 2026. Its third-quarter operating cash flow was $32.79 billion, leaving free cash flow at negative $2.12 billion, compared with positive $10.6 billion a year earlier.

Meta’s October 22 filing lists about $355.73 billion in leases that have not started and $454.52 billion in non-cancelable contractual commitments, mainly for cloud capacity, servers and network equipment. It also lists up to $14.72 billion in contingent cloud purchases. The figures total roughly $825 billion in future commitments, but are not all AI spending; the filing also includes items such as Reality Labs hardware. Separately, the source reports SpaceX borrowed $40 billion to buy AI chips, a sign of the scale of financing being sought for infrastructure.

The Cost of Building AI Capacity

The contrast between substantial investment and limited aggregate evidence of returns matters to companies deciding how quickly to adopt AI, and to investors and lenders financing the infrastructure. If spending rises faster than revenue or productivity gains, it can weigh on margins and increase dependence on borrowing. Meta’s negative free cash flow in the quarter is one concrete measure of the pressure, although the company remains a large, profitable business and the figure alone does not establish that AI investment is uneconomic.

Evidence from particular uses is more mixed than the broad company averages. The source cites customer-support research showing 14% more issues resolved per hour, with a 34% gain for novice workers, and retail experiments reporting sales changes from zero to 16.3%, depending on the workflow. It also reports that 83% of nearly 9,000 executives said AI had not yet affected productivity, while the average estimated effect was 0.2% a year. These figures have different scopes and methods; they should not be treated as a single, directly comparable measure of return.

What Apollo’s Data Covers

Apollo’s October 24 Daily Spark note, titled “No Signs of AI Payoff Outside the Tech Sector,” draws on five charts. Alongside company revenue and margins, it cites San Francisco Federal Reserve productivity measures that, according to Apollo, show no sign of AI boosting productivity growth. The note also uses data from Ramp, a corporate-spending platform with more than 85,000 U.S. business customers. Ramp’s figures indicate that the top 10% of firms account for most recorded spending on model inference, neocloud services and named AI providers. Those are Ramp-customer spending patterns, not a census of the S&P 500.

The interpretation has drawn attention because the investment commitment is large and continuing. Meta’s filing details leases and purchasing obligations over multiple years, while forecasts cited in the source put spending by Amazon, Alphabet, Meta, Microsoft and Oracle at nearly $780 billion in 2026 and above $1 trillion annually from 2027 through 2030. Those are projections, not amounts already spent or binding commitments. Apollo’s findings concern visible economy-wide results so far; they do not establish what future returns will be.

“Artificial intelligence has delivered zero financial payoff to corporate America outside of tech.”

— Torsten Slok, Apollo chief economist, in an October 24 Daily Spark note

What the Return Figures Cannot Show

A flat average among 493 companies can mask firms that are gaining from AI and firms that are not, with their results offsetting each other. Apollo’s comparisons also cannot separate AI’s effect from other forces such as interest rates, tariffs or labor costs. The Ramp data describe the platform’s customer base, and the productivity indicators may not capture gains in individual jobs or newly adopted workflows.

It remains unclear how much of Meta’s listed commitments will be used, when the associated capacity will generate revenue, and what share is specifically attributable to AI rather than other products and infrastructure. A negative free-cash-flow quarter does not by itself establish a lasting trend. The available material also does not show whether the reported workplace trial results will persist at larger scale or translate into lasting profits.

Evidence Investors Will Watch

Investors and companies will have to compare subsequent earnings, cash-flow reports and productivity data with the infrastructure commitments already disclosed. For Meta, future filings and capital-spending updates should clarify how spending develops against its 2026 outlook and whether the new capacity supports revenue growth. For the wider sector, the key test is whether measurable gains spread beyond a small group of technology firms and selected workplace tasks.

Financing conditions are another signal to watch as AI providers seek capital for chips and data centers. The source reports that insurance costs on debt for Meta and CoreWeave have risen, but does not give enough detail to establish how broad or persistent that change is. Apollo’s analysis is a snapshot of available data, not a verdict on eventual returns; the timing and distribution of those returns remain open.

Key Questions

What did Apollo mean by saying AI has delivered “zero financial payoff” outside tech?

Slok was describing aggregate trends: revenue per worker and margins rose for seven large technology companies while the other 493 S&P 500 firms were broadly flat. That comparison does not prove that no individual non-tech company or work process has benefited.

Does Meta have $825 billion in AI spending?

No. The roughly $825 billion figure is a total of future commitments listed in Meta’s filing, including leases and contractual purchases. The source says the commitments are mostly for data centers, servers and cloud capacity, but notes that not all of them are AI-related.

Is there evidence that AI can improve work outcomes?

The source cites task-specific results, including more customer-support issues resolved per hour and sales changes in retail experiments. Results varied by task, and the figures do not show that every company will achieve the same gains or that those gains will exceed implementation costs.

What remains unknown about AI’s financial returns?

It is not yet clear how quickly company-wide productivity gains will appear, how widely they will spread, or whether revenue and cost savings will cover investment in chips, data centers and cloud services. Aggregate figures also cannot identify AI’s effect separately from other economic factors.

Source: ThorstenMeyerAI.com