AI Gains Confined to Tech? Non-Tech Industries See Little Payoff

Over the past two years, generative AI has dominated the global business narrative. From chip shortages to the large-model arms race, tech giants have proven through quarterly earnings that AI is a good business. But a key question is surfacing: Are AI's economic gains staying within the tech sector? A analysis published by Apollo Chief Economist Torsten Slok offers a fairly clear judgment—outside the tech sector, there are currently no signs of quantifiable returns from AI. The piece sparked over 280 comments on Hacker News, suggesting it struck a chord with many people's real experiences.

A Sober Assessment: AI Gains Have Not Spread

Apollo's analysis has a blunt title: "No Signs of AI Payoff Outside the Tech Sector." It doesn't deny AI's technological potential but points to an empirical fact—based on macroeconomic and corporate earnings data, the productivity gains and revenue growth from AI are highly concentrated in a handful of tech companies.

What does this mean? If AI is truly a "general-purpose technology"-level transformation as many predict, its impact should gradually permeate retail, manufacturing, healthcare, finance, and other industries—just like electricity and the internet—showing up in their profit margins, employment structures, and growth curves. But current data shows no sign that this diffusion has occurred.

This judgment deserves attention precisely because it is an observation based on earnings and macroeconomic data, not an emotional statement about technological prospects.

General Purpose Technology (GPT) is a core concept in economics, referring to technologies that can widely permeate almost all industries and continuously spawn complementary innovations, such as the steam engine, electricity, and the internet. These technologies share common features: massive initial investment with delayed returns, ultimately reshaping the entire economy's production function. One key indicator of whether a technology reaches GPT status is whether its benefits spread from the originating industry to downstream sectors. Apollo's analysis borrows this framework: if generative AI is merely an arms race among tech companies, it has not yet passed the GPT test; only when profit margins and employment structures in retail, manufacturing, healthcare, and other industries show changes attributable to AI will the GPT attribute be truly realized. This also explains why "whether returns spill over" is a better measure of AI's economic significance than "whether the technology is advanced."

Why Haven't Returns Spilled Over?

There's a Clear Lag Between Input Costs and Output

Corporate spending on AI is currently skewed more toward "preparation" than "harvest." Infrastructure procurement, data governance, model selection, employee training—these are all upfront costs. The enterprise AI project failure rates repeatedly cited by consulting and research firms also reflect the huge gap between pilot programs and scaled deployment. For non-tech companies, AI is currently more like a capital expenditure that hasn't yet paid off.

The lag in returns on capital expenditure has a clear theoretical counterpart in macroeconomics. When companies purchase computing power, build data pipelines, and restructure workflows, they are essentially accumulating intangible capital, whose returns only materialize after organizational processes complete their adaptation. Economist Brynjolfsson's "productivity J-curve" describes this phenomenon: in the early stages of general technology adoption, productivity may even temporarily decline due to learning costs and process restructuring, before accelerating upward. The history of electrification is a classic example—it took decades after factories were electrified for the full efficiency gains to be released through redesigned production layouts. By comparison, non-tech companies' AI spending is currently in the downward segment of the J-curve: costs on the books, with output not yet ramping up. Apollo's observation can therefore be read as "the right half of the J-curve hasn't arrived yet," rather than "AI has no value."

The Real Beneficiaries Are Highly Concentrated

Looking at the value chain, the most certain winners in the AI wave are the infrastructure layer providing computing power, models, and cloud services. These companies are themselves tech firms, and their revenue growth reflects other companies' AI procurement, not efficiency gains in end-user industries. In other words, AI's "revenue" circulates within the tech sector and has not yet translated into "output" in the broader economy.

The "infrastructure layer" of the AI value chain mainly refers to companies providing computing chips (such as GPUs), cloud service platforms, and foundation models. This layer's business model has extremely strong economies of scale and winner-take-all characteristics: training frontier models requires billions of dollars in capital expenditure, and only a few giants can sustain such investment, while their computing power and models are charged downstream on a token or subscription basis. As a result, every time a downstream company purchases AI capabilities, the revenue is recorded on the infrastructure layer companies' financial statements—and these companies themselves belong to the tech sector. This creates a statistical "internal loop": total AI spending is growing, but the accounting attribution of that growth is highly concentrated, while efficiency improvements in end-user industries are difficult to reflect in current financial data. Understanding this layered structure is key to understanding "why returns aren't spilling over."

Transforming Traditional Businesses Is Harder Than Expected

Tech companies can embed AI directly into products and charge per token; manufacturing, retail, and professional services must graft AI onto existing processes, compliance requirements, and organizational culture. The data quality and process restructuring costs involved in this transformation are often underestimated by early optimistic forecasts.

Divergent Views on Hacker News: Just Too Early, or Overhyped?

The discussion on Hacker News provides a valuable cross-perspective. Some commenters agree with Apollo's observation, comparing AI to the dot-com bubble—infrastructure investment comes first, and application-layer returns take longer to materialize. Others are more pointed: they question whether many companies' AI projects are essentially "AI for AI's sake," lacking a clear business闭环 (closed loop), so seeing no returns is a natural outcome.

Still another view holds that macroeconomic statistics may themselves underestimate AI's impact—efficiency gains are dispersed across daily work in various roles and difficult to measure separately like new product revenue. This is a reasonable reminder: there is a gray area in measurement between "no visible returns" and "no returns." But regardless, if AI's impact is large enough, it should eventually appear in some form in industry-level data.

The measurement blind spot of macroeconomic statistics for dispersed efficiency gains deserves separate mention. GDP accounting and industry productivity indicators mainly capture tradable product and service output, while many AI benefits manifest as individual employee time savings, improved document quality, or faster decision-making—these "soft" improvements are difficult to monetize and don't appear separately in corporate financial statements. This is the contemporary version of the so-called "Solow paradox": in 1987, economist Solow quipped, "You can see the computer age everywhere but in the productivity statistics." History shows that such paradoxes are eventually resolved as statistical methods adjust and new business models mature, but before that, macroeconomic data may indeed systematically underestimate the breadth of the transformation. However, if the impact is large enough, it will eventually leave traces through indirect indicators such as industry profit margins and employment structures.

Implications for Businesses and Professionals

The most practical takeaway from Apollo's analysis is to make companies re-examine how they measure AI investment. If an AI initiative cannot specify which concrete process it will change, how much cost it will reduce, or how much conversion it will improve, then it is likely just a cost item.

For decision-makers in non-tech industries, a more pragmatic path may be to start with narrow use cases: customer service ticket triage, contract review, coding assistance, document processing—the returns in these scenarios are easier to quantify and more likely to create positive feedback within the organization. Conversely, the narrative of "deploy a large model and achieve an efficiency leap" often sounds compelling during budget approval but looks bad in year-end reviews.

For investors, this judgment also provides an observation framework: whether the AI narrative can spread from the tech sector to the broader market is itself a signal worth tracking. The moment non-tech companies' financial statements begin to consistently reflect AI-driven margin improvements is when AI's gains truly materialize.

Frequently Asked Questions

Has AI really brought no benefits to non-tech industries?

Apollo's analysis points to the absence of quantifiable macroeconomic return signs, not a blanket denial. Some individual companies have had successful AI deployments, but these have not yet formed an industry-level trend.

How long will this situation last?

The original article does not provide a clear timeline. Historical experience shows that the productivity gains from general-purpose technologies often take years or even longer to fully materialize, and the process is uneven.

Should companies stop investing in AI now?

Not stop, but calibrate. Tying investment to measurable business outcomes is more sustainable than chasing concepts.

One-line Summary

Apollo analysis indicates that AI's quantifiable economic returns remain concentrated in the tech sector and have not yet spilled over to other industries.

Paragraph Summary

Apollo Chief Economist Torsten Slok's analysis argues that, based on macroeconomic and corporate earnings data, the quantifiable returns from generative AI are highly concentrated within the tech industry, with no clear signs yet in non-tech sectors such as retail, manufacturing, healthcare, and finance. The article explains why returns haven't spilled over from three angles: corporate AI spending is skewed toward upfront preparation rather than harvest; the real beneficiaries are concentrated in the infrastructure layer of computing power and models; and transforming traditional businesses faces high process and compliance costs. Hacker News discussion reveals divergent views: some see delayed returns as normal, some question whether projects lack a clear business closed loop, and others remind that macroeconomic statistics may underestimate dispersed efficiency gains. The article suggests companies tie AI investment to measurable business outcomes, start with narrow use cases, and view margin improvements in non-tech companies' financial statements as the signal that AI's gains are truly materializing.