The AI payoff is real. The timeline is the problem

Goldman Sachs estimates the growth boost from AI infrastructure spending will peak in late 2026 or early 2027, while Apollo finds revenue per worker and margins rising among the Magnificent Seven but flat across the other 493 S&P 500 companies. Surveys indicate that only about 6% of companies report AI contributing at least 5% of EBIT, and 40% of surveyed CEOs said AI hype had led to overinvestment. The article argues that AI’s long-term productivity potential remains real, but adoption and financial returns are lagging investment, raising the risk of overbuilt data-center capacity and a near-term spending recalibration.

Analysis

The key market risk is a duration mismatch: infrastructure cash flows are being financed ahead of evidence that enterprise adoption can produce durable, broad-based returns. If customer deployments remain pilots, the first adjustment is more likely to be lower utilization and delayed orders than an abrupt collapse in AI demand. That would pressure the economics of incremental data-center capacity and eventually challenge financing assumptions; the spillover could reach lenders and asset managers if projects depend on refinancing or contracted power commitments. The article does not establish the exposure of any named firm to those projects, so this is a sector-level risk, not a company-specific earnings call.

For GS, APO, BLK and JPM, executive optimism is not proof of near-term monetization. JPM and GS could benefit over time from internal process automation, but regulated workflows and verification costs may defer savings. APO and BLK warrant monitoring for any disclosed data-center financing, valuations, or impairments rather than assuming direct losses. The contrarian point: flat broad-market productivity does not disprove long-run AI value; it does argue against pricing broad productivity gains on the same timetable as capex.

Over days, positioning and narrative can drive volatility. Over 1–3 months, watch hyperscaler capex guidance, data-center utilization/lease commitments, and enterprise software spending. Over 6–18 months, the test is measurable EBIT improvement net of token, integration, and rework costs. A broad-based rise in productivity and earnings would reverse the caution; continued capex without utilization or customer ROI would strengthen it.

Key Decisions for Investors

• Avoid adding to undifferentiated AI-infrastructure exposure solely on capex guidance. For a relative-value expression, consider a modest long in an S&P 500 equal-weight ETF versus a short in a cap-weighted technology-heavy index, sized as a hedge rather than a high-conviction standalone bet. Reassess over 1–3 months as capex and adoption evidence arrives; broad earnings upgrades among large technology firms would falsify the relative-value thesis.
• Do not infer company-specific credit or impairment risk for APO or BLK from this article. Flag their disclosures for data-center and digital-infrastructure financing, asset marks, project concentration, and contractual power exposure; escalate only if those data show material exposure or weakening counterparties.
• For JPM and GS, track disclosed AI-related cost savings against implementation, oversight, and remediation costs over the next 2–4 quarters. Treat announced pilots or productivity claims as non-confirmatory until reflected in expense ratios or operating efficiency; rising costs without measurable savings would weaken the longer-term beneficiary case.
• Maintain an alert on data-center supply and financing: order cancellations, lower utilization, delayed project delivery, or revised lease commitments would be near-term downside catalysts for infrastructure-linked assets. Conversely, sustained enterprise deployment and measurable customer ROI would argue against extending the underweight.