AI Infrastructure Has Led the Boom. The Productivity Trade Could Be Next
AI infrastructure stocks have dominated the market's attention and the earnings scoreboard. The next wave of gains may land somewhere far less glamorous.
Infrastructure drove this earnings season
S&P 500 earnings per share grew 31% year-over-year in the second quarter, excluding one-time income tied to private investment stakes. AI infrastructure stocks, including hyperscalers, accounted for roughly half of that growth, with their earnings up 54% year-over-year.
The median S&P 500 company outside energy still grew earnings 14%, a strong showing with little help from AI adoption. The gap between the infrastructure winners and everyone else reflects how the market has behaved all year.
Investors have paid up for companies with clear, near-term earnings tied to AI buildout: chips, data centers, networking, power. The companies that might eventually benefit from AI-driven productivity have largely been ignored.
The earnings proof is still thin
Goldman Sachs strategist Ben Snider has been tracking how often S&P 500 companies actually quantify AI's impact, not just mention it. The results are still early and messy.
Only 11% of S&P 500 companies quantified AI productivity for a specific use case, such as coding or customer support, during earnings calls. Just 2% quantified AI's direct impact on earnings, the same share as in the first quarter.
"Q2 results showed a small and statistically insignificant difference in earnings growth between the companies quantifying AI productivity gains this quarter and other S&P 500 companies."
Ben Snider, Goldman Sachs
That is the core tension. Companies are spending on AI and talking about it constantly. The savings and efficiency gains are real enough to keep budget allocations moving. But they have not bent the earnings curve in a way markets can score cleanly.
Spending is accelerating from a low base
The clearest forward-looking signal comes from enterprise AI spend data. The Ramp AI Index shows median monthly AI spend per employee rose from $5 in January to $12 in July.
Top-decile spenders went from $240 to $650 over the same stretch. Goldman estimates AI inference costs still represent under 0.5% of S&P 500 revenues.
That low cost base means companies can experiment without damaging their income statements, and the upside can scale before the expense line becomes a problem.
Roughly two-thirds of companies fund AI spending by reallocating existing budgets. The biggest reallocation targets are software at 18% and labor at 11%.
Companies with heavy software use and meaningful labor exposure are best positioned to convert AI into margin expansion.
Where Goldman is looking next
Goldman ran a screen for Russell 1000 companies most likely to benefit from AI adoption, focused on labor cost sensitivity and AI automation exposure.
The screen excluded AI infrastructure and disruption-risk names and required that companies mentioned AI in the context of productivity or efficiency on their second-quarter earnings calls.
The top names span a wide range of industries. CoStar Group, Dollar Tree, and eBay appear alongside insurance brokers Arthur J. Gallagher, Brown & Brown, and Aon.
The list also includes Axon Enterprises, Trade Desk, Airbnb, Iron Mountain, CBRE, RTX, Boeing, and Expedia. What ties them together is high labor costs relative to revenue and significant potential to automate that work with AI.
Goldman separately projects hyperscaler capital spending could reach roughly $1.1T in 2027, above the ~$920B Wall Street consensus. That infrastructure spend continues to support earnings for chip, networking, and power companies.
Goldman also flags that valuations in that part of the market have expanded rapidly, increasing volatility risk.
The productivity trade is still an expectations market. The companies in Goldman's screen have not demonstrated a statistically meaningful earnings edge yet.
The thesis is that accelerating enterprise AI deployment will make that edge visible in coming quarters. Until it does, the position is patience on software-heavy, labor-leveraged names, waiting for the moment proof replaces promise.