The depreciation clock Goldman just started

Goldman Sachs Research published a note this week projecting that AI capital expenditure will compress return on equity at the top seven US tech companies by an average of seven percentage points in 2027. The headline number is doing the work in the coverage. The mechanism underneath it is more interesting, and slightly more awkward, than the headline suggests.

What was actually published. Goldman's economists model the top-seven mega-cap tech aggregate, which ran ROE (return on equity, profit divided by shareholder equity) of 44% in the first quarter against an S&P 500 record of 22%.¹ Against that base, they project a 7pp drag in 2027. The driver is not the capex commitment itself but what happens to the capex once it sits on the balance sheet. Depreciation on AI assets is projected to rise from 7% of revenue in 2022 to roughly 12% by 2027.¹ That five-point swing, applied to a high-margin revenue base, is most of the ROE story.

The capex inputs themselves are well-rehearsed. Goldman now projects total AI infrastructure investment of roughly $770bn across the ecosystem in 2026 once data-centre providers and utilities are included, with $320bn coming from the five core hyperscalers, Microsoft, Meta, Alphabet, Amazon, Apple, up from about $230bn in 2025.²

The depreciation lag is the structural feature. Capex commitments are a flow; depreciation is the same flow rephrased as accounting reality once assets enter service. Servers and data-centre fit-out depreciate on schedules of roughly four to six years. The kit ordered through 2024 and 2025 will be hitting the income statement through 2026 and 2027 regardless of what AI revenue does in those years. Goldman's 7pp ROE figure is, in large part, an arithmetic consequence of decisions already made.

This matters for how the note should be read. It is not a forecast about whether AI will work commercially. It is a near-deterministic projection about how assets already on the ground will flow through the P&L. The bear case on ROE assumes nothing controversial about AI demand; it assumes only that the depreciation schedule does what depreciation schedules do.

Where the note gets awkward. Goldman is simultaneously embedding a 1.5pp AI productivity tailwind to S&P 500 EPS (earnings per share) in its 2027 base case.¹³ The bank's headline projects compression on the equity side of the ROE ratio while its earnings forecast assumes the productivity thesis lands on the numerator. These are not strictly contradictory, ROE compression and EPS growth can coexist if equity grows faster than earnings, but the framing tension is real. The note that gets quoted as a warning about AI capex is, in its own base case, still long the productivity story.

Capex at 100% of operating cash flow is the line that matters. The other number worth dwelling on is that top-seven aggregate capex in 2026 is running at roughly 100% of operating cash flow.¹ The hyperscalers have funded infrastructure from internal cash for most of the cloud era. A 100% ratio means free cash flow after capex is approximately zero before any return of capital to shareholders.

This is not circular financing in the sense the term usually gets thrown around — these companies are not financing themselves with their own paper. But it is the threshold at which balance sheet quality starts mattering again. Any shortfall in AI revenue, any acceleration in spend, any buyback the boards want to maintain, has to come from debt issuance or dilution. The hyperscalers can absorb this comfortably for now: investment-grade ratings, low leverage, cheap access to debt markets.⁴ The risk is contingent, not imminent. But the buffer that defined the previous decade is gone.

The correlation collapse is the market-structure tell. Goldman notes in passing that correlations among AI-related stocks have fallen from roughly 80% to 20%.¹ This is the line that should travel further than the ROE number. From late 2022 through early 2024, owning AI exposure was sufficient — the basket moved together. At 20%, the market is underwriting names on individual ROI (return on investment) evidence rather than thematic membership.

The inference-economics frame predicts this transition. Once the binding question shifts from "who benefits from AI?" to "which AI investments are generating returns?", undifferentiated exposure stops paying. That is the regime the correlation number says we are now in. The implication for capital allocators is straightforward: the trade that worked from late 2022 through mid-2024 has structurally changed shape, and the analytical work required to make money in AI equities just got considerably harder.

What this is the shape of. The aggregate capex number, $770bn in 2026, is now large enough to move S&P 500 returns metrics at the index level, not just individual company margins. That is the AI performativity frame operating at macro scale. The spend, independent of delivery, is reshaping the financial landscape it claims to be predicting. ROE compression at the mega-cap level pulls down index ROE; depreciation flowing through hyperscaler income statements compresses aggregate margin; the productivity tailwind Goldman embeds is needed to offset the headwind Goldman is also flagging.

The honest reading of the note is that Goldman has produced a piece of research where the headline finding and the embedded assumption point in opposite directions, both presented as base case. That is not a criticism of the analysis. It is a description of where the market actually sits in mid-2026: committed to a spend that mechanically compresses returns, dependent on a productivity case that has not yet shown up in margins, and now being differentiated at the name level by investors who have stopped buying the basket.

The depreciation clock runs whether or not the productivity thesis lands. That is the asymmetry worth holding in mind.

Footnotes

¶