Goldman Sachs Analysis Finds AI Inference Costs Comparable to Corporate Payroll Expenses

A Goldman Sachs research note circulating among institutional investors has crystallized a trend that CFOs across the S&P 500 are grappling with: at a growing number of companies, the cost of running AI models in production — known as inference — is approaching the same order of magnitude as their total payroll expenses. The finding has sparked debate about whether corporate America is entering a structural shift in how it allocates spending between human labor and machine intelligence.

The Numbers Behind the Claim

The core of the Goldman Sachs analysis rests on a comparison that would have been unthinkable three years ago. At major AI-native companies, compute expenses — encompassing both training and inference — now consume 54% to 62% of total operating costs, while staff spending accounts for less than 25% [1]. At Anthropic, for instance, total compute spending in 2025 reached $6.8 billion (42% for R&D compute, 28% for inference), compared to $2.9 billion for staff — meaning the company spent more than twice as much on running and training models as it did on paying people [1]. Minimax and Z.ai show similar patterns, with compute representing 57% and 62% of costs respectively [1].

These ratios are most extreme at AI-first companies, but the trend is spreading. Enterprise generative AI spending jumped from $11.5 billion in 2024 to $37 billion in 2025 — a 320% increase — and is projected to reach $55 billion in 2026 [2]. The average enterprise AI budget has grown from $1.2 million per year in 2024 to $7 million in 2026, with 45% of organizations now spending over $100,000 per month on AI [2]. Goldman Sachs projects that AI companies collectively may invest more than $500 billion in infrastructure during 2026 alone [3].

Source: Epoch AI

Data as of Dec 1, 2025CSV

The Inference Cost Paradox

The most counterintuitive aspect of this story is that inference costs per unit have collapsed even as total spending has skyrocketed. The cost of querying a model at GPT-3.5 performance levels fell from $20 per million tokens in November 2022 to $0.07 by October 2024 — a more than 280-fold reduction in roughly 18 months [4]. Epoch AI's analysis of price trends across multiple benchmarks found median annual decline rates of 50x, with post-January 2024 declines accelerating to 200x per year [5].

Source: AnalyticsWeek / Industry Reports

Data as of Jan 1, 2026CSV

This is a textbook illustration of Jevons' Paradox — the 19th-century observation that making a resource cheaper increases rather than decreases total consumption. As inference became 1,000 times cheaper, companies deployed it in customer service automation, real-time data analysis, personalized content generation, and continuous agentic AI systems running around the clock [2]. The per-unit savings were overwhelmed by demand that the lower prices unlocked.

Goldman Sachs CIO Marco Argenti has warned about "token sticker shock" — the moment when executives realize that despite dramatically lower unit costs, their aggregate AI bills are growing faster than almost any other budget line item [6]. The 2026 turning point, where inference spending now exceeds training spending for the first time — accounting for 55% of the $37.5 billion in AI cloud infrastructure spending — confirms that production deployment, not research, is now the primary cost driver [7].

Which Industries Hit the Threshold First

The payroll-parity claim holds most cleanly in technology and financial services. Among S&P 500 companies, AI-related stocks have accounted for 75% of index returns, 80% of earnings growth, and 90% of capital spending growth since ChatGPT launched in November 2022 [8]. These are also the sectors spending most aggressively on inference.

But adoption remains uneven. As of March 2026, fewer than 19% of U.S. establishments have adopted AI, though this is expected to rise to 22.3% within six months [9]. Large firms (250+ employees) show a 35.3% adoption rate, while small firms lag at 21.5% [9]. Goldman Sachs itself estimates that 80% of companies are not yet using AI meaningfully [9].

This means the "inference costs rival payroll" comparison applies to a relatively small vanguard of companies — primarily hyperscalers, AI labs, and major financial institutions — rather than the median S&P 500 firm. For most enterprises, AI spend remains a single-digit percentage of operating costs. The question is how quickly the rest of corporate America follows the leaders into production-scale deployment.

What Counts as 'Inference Cost'

One significant complication in the Goldman Sachs analysis is the absence of standardized accounting treatment for AI inference expenses. The costs bundled under "inference" can include cloud compute fees paid to providers like AWS, Azure, or Google Cloud; amortized capital expenditures on proprietary GPU clusters; software licensing fees for model access; and energy costs for running data centers [10].

Companies report these expenses inconsistently. Some classify AI compute under R&D; others fold it into cost of goods sold or general technology spending. Only 51% of organizations can confidently measure their AI return on investment, creating what one analysis calls an "accountability gap" in spending that continues to accelerate [2].

This matters because the payroll comparison depends on what you include. A narrow definition — just cloud API fees — produces a smaller number. A broader definition — encompassing capital depreciation on GPU clusters, energy, cooling, and internal engineering time spent on model optimization — can dwarf the narrow figure. Without consistent reporting standards, cross-company comparisons remain imprecise.

The Labor Displacement Question

If inference costs are functionally replacing payroll line items, the implied question is whether workers are being displaced. The evidence is mixed but directional.

In a survey of 1,000 U.S. business leaders, 39% reported conducting layoffs in 2025, with 58% expecting further layoffs in 2026 [11]. Corporate America cut more than 1.17 million jobs in the first 11 months of 2025 — a 54% increase from 2024 [11]. High-salary roles are being targeted first for immediate payroll savings, with workers lacking AI-related skills facing the highest risk [11].

Specific categories face acute pressure: 80% of customer service roles are projected to be automated, potentially displacing 2.24 million out of 2.8 million U.S. jobs in that sector [12]. Paralegals face an 80% automation risk by 2026, and legal researchers face 65% by 2027 [12]. Mid-layer management — including HR, talent acquisition, and compliance roles — is being disproportionately targeted [11].

Source: Epoch AI / Stanford AI Index

Data as of Oct 1, 2025CSV

Total nonfarm payrolls stood at 158.6 million as of March 2026, up only 0.2% year-over-year [13], while average hourly earnings rose 3.5% to $37.38 [14]. The labor market remains broadly stable, but the composition is shifting. Goldman Sachs' own research warns that AI-driven job losses could have "lasting costs," with displaced workers facing pay cuts and slower career growth [15].

Historical precedent offers a partial guide. ATMs did not eliminate bank tellers — the number of tellers per branch fell, but the lower cost per branch led to more branches being opened, and tellers shifted toward sales and relationship roles [12]. ERP systems eliminated many data-entry positions in the 1990s but created new categories of work in systems administration and business analysis. The net employment effect of prior automation waves has generally been positive over 10-to-20-year horizons, though the transition period imposes real costs on displaced workers [12].

Women face disproportionate exposure: 79% of employed women are concentrated in occupations at high risk of AI automation, compared to 58% of men — a 1.4x differential [12].

Goldman Sachs' Conflicts of Interest

Goldman Sachs occupies an unusual position in this analysis. The firm is simultaneously a producer of research about AI infrastructure spending, a major financial advisor to the hyperscalers driving that spending, and an investor in AI infrastructure itself. Goldman's stock soared 56% in 2025, driven in part by AI-related deal flow [16].

Goldman formally discloses these conflicts: "Goldman Sachs does and seeks to do business with companies covered in its research reports. As a result, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of this report" [16]. The firm's business model creates a structural incentive to normalize large AI capital expenditures — every upward revision to AI capex estimates generates advisory fees, trading commissions, and asset management opportunities [16].

This does not mean the analysis is wrong. But it does mean the framing — presenting inference costs reaching payroll parity as a natural and inevitable progression rather than a potentially unsustainable spending pattern — serves Goldman's institutional interests. When a bank that profits from AI infrastructure deals publishes research arguing that massive AI infrastructure spending is rational, the audience should weigh that context.

Goldman is also transforming its own operations with AI through its "OneGS 3.0" initiative, embedding generative and predictive AI across the firm [6]. CFO Denis Coleman has described it as a multi-year overhaul of Goldman's operating model [6]. The firm is both analyst and subject in this story.

The International Picture

The payroll-parity claim is most applicable to U.S. companies, where AI infrastructure investment is concentrated. The international picture differs substantially.

China has created a structurally lower cost environment through aggressive state subsidies. Chinese provinces like Gansu, Guizhou, and Inner Mongolia have slashed cloud providers' power bills by as much as 50% for AI workloads [17]. Electricity in China's western provinces costs as little as $0.05 per kilowatt-hour, compared to $0.25 in Beijing and Shanghai or $0.40 in parts of the U.S., and power makes up roughly 35% of inference costs [17]. Chinese AI labs like Kimi offer inference at one-quarter the price of comparable U.S. models [17]. Beijing is also directly subsidizing API access and model licensing [17].

The European Union faces the opposite problem. EU fiscal rules constrain member states to deficits below 3% of GDP, limiting the kind of public investment that could subsidize AI infrastructure [18]. The EU produced only about three large foundation models in 2025, compared to roughly forty in the U.S. and fifteen in China [18]. European companies face higher compliance costs under the EU AI Act's risk-based regulatory framework [19]. The payroll-parity dynamic is largely absent in Europe — not because companies are spending less on AI relative to workers, but because they are spending less on AI altogether.

These differences suggest that the Goldman Sachs analysis describes a phenomenon concentrated in the U.S. and, in modified form, in China — not a universal global trend.

Tax, Governance, and the Balance Sheet

If AI inference is functionally substituting for labor, the implications extend beyond operating budgets into corporate tax strategy, governance, and shareholder relations.

Payroll spending carries significant tax overhead — employer-side Social Security and Medicare contributions, unemployment insurance, workers' compensation, and benefits costs. Compute spending carries none of these. A dollar shifted from payroll to inference is worth more to the company on a post-tax basis, creating a structural incentive to automate even when the pre-tax costs are roughly equal [20].

Proposals to address this imbalance are emerging. The "Transition Economy Act," a bipartisan framework under discussion in Congress, would fund direct transfers to displaced workers through a tax on AI inference compute [20]. Whether such legislation gains traction depends on how quickly the payroll-to-compute substitution becomes visible in employment statistics.

At the board level, the shift is beginning to register. Companies now spend an average of 36% of their digital budgets on AI, with digital budgets themselves growing from 7.5% of revenue in 2024 to 13.7% in 2025 [7]. Some compensation committees are starting to evaluate AI spending alongside headcount in workforce planning, though formal treatment of AI as a payroll-substitute line item remains rare.

Source: BLS / Bureau of Labor Statistics

Data as of Mar 1, 2026CSV

What the Productivity Data Actually Shows

For the payroll comparison to be economically meaningful — rather than merely arithmetically interesting — AI inference spending needs to generate productivity gains that justify the outlay. The early evidence is encouraging but incomplete.

Enterprise workers using AI tools report recovering 40 to 60 minutes per day, according to Goldman Sachs' own survey data [9]. Academic studies show a 23% average productivity uplift, while individual company reports indicate gains closer to 33% [9]. Seventy-five percent of AI-equipped workers say they can now complete tasks they previously could not do at all [9].

But only 51% of organizations can confidently measure AI return on investment [2]. The gap between perceived productivity gains and measurable financial returns remains wide. And the agentic AI systems that are driving the fastest inference cost growth — autonomous agents running 24/7 — are too new to have established clear ROI benchmarks.

The Bottom Line

The Goldman Sachs analysis identifies a real and accelerating trend: at AI-intensive companies, compute costs have reached and in some cases surpassed payroll as a share of operating expenses. The comparison is most stark at AI-native firms like Anthropic, where compute consumes 70% of costs versus 30% for staff [1], and it is spreading into financial services, technology, and customer-facing industries as deployment scales.

But the framing requires several caveats. The comparison applies to a small fraction of companies today — fewer than one in five U.S. establishments have adopted AI at all [9]. The "inference cost" category lacks standardized accounting treatment, making precise comparisons unreliable. The Jevons' Paradox dynamic — plummeting unit costs driving exploding total spend — means the trend could reverse if usage growth slows or if further cost reductions outpace demand expansion. And Goldman's own financial interests in normalizing large AI capital expenditures should inform how much weight readers place on the analysis.

What is not in dispute is the direction. Enterprise AI spending tripled in a single year [2]. Inference now exceeds training as a share of AI infrastructure budgets [7]. Companies are beginning to treat AI compute as a functional substitute for certain categories of labor. Whether that substitution produces net economic gains or net displacement — and over what time horizon — is the central economic question of this decade.

Source: FRED / Bureau of Labor Statistics

Data as of Mar 1, 2026CSV