Corporate Profit Margins Stay Flat Despite Heavy AI Spending

The Missing Returns

Artificial intelligence investments by major corporations outside the technology sector have yet to produce measurable improvements in financial performance, according to recent analysis by Torsten Slok, chief economist at Apollo Global Management. His examination of 493 large US-listed companies - the S&P 500 index excluding seven tech giants - shows flat profit margins and stagnant revenue per worker, indicators that productivity gains from AI deployment remain elusive.

The finding arrives at a moment when enterprise software vendors, consulting firms, and cloud infrastructure providers have built multi-billion-dollar businesses on the promise that generative models and machine-learning tools will reshape operations across manufacturing, finance, healthcare, and retail. Yet Slok's data suggests that promise has not yet translated into bottom-line results for the corporations writing the cheques.

Apollo Global Management, a private equity firm with more than $650 billion in assets under management, excluded Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla from the analysis - precisely the companies building and selling AI infrastructure. What remains is a cross-section of industrials, consumer goods, pharmaceuticals, banks, and energy firms that represent the intended customer base for AI tooling.

Intelligence Versus Judgement

At Opentechwire, we've tracked the gap between capability demonstrations and operational deployment for years. The core tension Slok identifies is structural: artificial intelligence excels at pattern recognition, classification, and retrieval, but corporate value creation often hinges on context-dependent decisions that blend technical analysis with institutional knowledge, risk appetite, and stakeholder relationships.

According to Apollo's analysis, the technology may function as a sophisticated calculator, yet business decisions require judgement rooted in accountability, organisational culture, unwritten norms, and ethical trade-offs. These are domains where large language models and computer-vision systems remain brittle. A model can summarise a legal document or flag anomalies in financial filings, but it cannot weigh reputational risk against short-term margin expansion, nor can it navigate the internal politics that determine whether a new initiative secures executive sponsorship.

This distinction matters because much of the AI investment narrative has rested on the assumption that cheaper, faster inference - compute costs have fallen sharply over the past two years - would automatically flow through to higher revenue growth and wider margins. Slok's scepticism centres on the question of conversion: even if intelligence becomes abundant and inexpensive, turning that resource into incremental profit requires changes in workflow, incentives, and decision rights that many organisations have struggled to implement.

Deployment Challenges Across Sectors

The flat productivity data aligns with findings from labour economists and field studies conducted over the past 18 months. Early pilots in customer service, software engineering, and document review have shown time savings for individual tasks, but those gains often fail to compound at the departmental or company level. Reasons include integration friction, hallucination risks that require human review, and the reality that many workflows are not bottlenecked by the speed of information retrieval.

In regulated industries - banking, insurance, pharmaceuticals - compliance and audit requirements add layers of verification that slow adoption. A generative model might draft a response to a regulatory query in seconds, but legal and compliance teams still need to review every line, erasing much of the time advantage. In manufacturing and logistics, where margins are thin and systems are deeply embedded, the cost and risk of replacing legacy software with AI-enabled alternatives can outweigh projected benefits.

Meanwhile, the technology sector itself has captured the financial upside. Nvidia's data-centre revenue has grown at triple-digit year-on-year rates, hyperscale cloud providers have seen accelerated consumption, and software companies with embedded AI features have commanded valuation premiums. The question Slok's analysis raises is whether this concentration of returns is temporary - a natural lag as enterprise buyers experiment and learn - or structural, reflecting fundamental limits on how much value AI can unlock outside the platforms that produce it.

The Lag Hypothesis

One interpretation is that corporate America is simply early in the adoption curve. Historical parallels - electrification in the early 20th century, enterprise software in the 1990s - suggest that transformative technologies often take a decade or more to reshape productivity metrics, because they require complementary investments in training, process redesign, and organisational structure.

Under this view, the current phase is one of experimentation and capability building. Firms are training employees, running pilots, and identifying use cases, but have not yet reached the scale or maturity needed to move company-level metrics. Revenue-per-worker and margin improvements will materialise in the next three to five years as deployment broadens and best practices diffuse.

Slok appears less convinced. His note emphasises the qualitative mismatch between what AI does well and what drives corporate performance. If intelligence were the binding constraint, cheaper and faster intelligence would indeed unlock growth. But if the constraint lies elsewhere - in coordination, in risk management, in customer relationships, in regulatory navigation - then model improvements and cost reductions may have limited impact.

What the Data Reveals

Apollo's analysis does not disclose the specific time period examined or the granularity of sector breakdowns, but the broad conclusion is clear: across nearly 500 large non-tech companies, there is no observable uptick in profit margins or labour productivity that coincides with the surge in AI investment and deployment since late 2024.

This does not mean AI is delivering zero value. Individual teams and functions may be seeing meaningful benefits that do not yet aggregate to company-level financial metrics. It also does not preclude future gains. But it does challenge the narrative that AI is already reshaping corporate economics outside the technology sector.

For investors and executives, the implication is that AI spending should be evaluated with the same rigour as any other capital allocation decision. Enthusiasm about the technology's potential is warranted; assumptions about near-term returns may not be.

Regional and Competitive Dynamics

The concentration of AI financial returns within a small number of US technology companies also has implications for competitiveness in Asia and Europe. If the current pattern persists - infrastructure providers capture value, enterprise users see flat returns - then the regions and countries with strong positions in semiconductors, cloud platforms, and foundation models will extract disproportionate economic benefit.

South Korea, Taiwan, and Japan have semiconductor fabrication and design capabilities that position them in the supply chain, but fewer large-scale foundation-model developers. China has invested heavily in domestic AI platforms, but export controls on advanced chips have constrained access to cutting-edge training infrastructure. Southeast Asian and South Asian firms are largely buyers of AI services rather than builders, raising questions about where value will accrue as the technology matures.

European regulators have focused on governance and risk frameworks, which may slow deployment but could also create competitive advantage if safety and transparency become differentiators. The EU AI Act, which entered into force in 2024, imposes obligations on high-risk systems that add compliance cost but may also build trust with enterprise customers wary of reputational exposure.

What Comes Next

Slok's analysis will likely fuel debate among economists, technologists, and business leaders about the pace and distribution of AI's economic impact. Optimists will point to early-stage adoption and the long tail of diffusion; sceptics will highlight the structural barriers to converting intelligence into profit.

For companies outside the technology sector, the data suggests caution. Investments in AI should be tied to specific, measurable outcomes rather than broad assertions about transformation. Pilots should be evaluated on whether they deliver financial returns, not just technical feasibility. And organisations should be prepared for the possibility that the most hyped technology of the decade may take longer, and require more organisational change, than many forecasts assume.

The funding rounds we've followed across the region over the past two years have been premised on the idea that every company will become an AI company. Apollo's numbers suggest that transition is proceeding more slowly, and less profitably, than the pitch decks promised.