Goldman Forecasts a Seven-Point ROE Hit From AI Capex

Goldman Sachs warned on June 12 that the Magnificent 7 are about to discover what happens when a software business turns into a power plant: return on equity is projected to fall an average of seven percentage points next year as $770 billion in 2026 AI capex — roughly 100% of hyperscaler operating cash flow — forces rising depreciation, heavier debt, and dilution. Microsoft is down more than 21% year-to-date. Micron is up 338% over the past year. Wall Street is not abandoning AI. It is abandoning the companies writing the checks.

The Goldman math: Portfolio strategist Ben Snider's June report frames the problem as asset intensity, not demand failure. Hyperscalers — Amazon, Microsoft, Alphabet, Meta, and Oracle among them — are projected to spend $770 billion on AI infrastructure in 2026. That sum equals about 100% of their collective cash from operations. Analyst estimates imply depreciation and amortization will climb from 7% of hyperscaler revenue in 2022 to 12% by 2027. Net debt across the group has risen $170 billion since the start of 2025. Share counts are creeping up as buybacks slow. Goldman projects 2027 hyperscaler capex could reach $1.1 trillion — well above the $920 billion Wall Street consensus.

Why investors are selling: The selloff is not a vote against AI adoption. Azure grew 40% last quarter. Amazon guides to $200 billion in 2026 capex. Alphabet raised its range to $175–$185 billion. Meta lifted guidance to $125–$145 billion, with CEO Mark Zuckerberg citing memory pricing as a cost driver. Microsoft projects $190 billion — a 61% jump — even as gross margins compress to their lowest level since 2022. The market is repricing what those numbers do to returns, not whether customers want the product.

The rotation in motion: Capital is flowing out of AI spenders and into AI suppliers. Microsoft has fallen more than 30% from its all-time high. Meta and Nvidia are down double digits from recent peaks. Meanwhile Micron has surged 338% over the past year, Western Digital is up 466%, and the VanEck Semiconductor ETF has gained roughly 60% year to date. SK Hynix is sold out through 2026. Micron is sold out through 2027. TSMC CEO C.C. Wei told shareholders on June 4 that advanced-node demand exceeds capacity by 25–30% with no relief expected until 2027 at the earliest. The bottleneck moved from GPUs to memory — and investors followed the scarcity.

The IPO clock: This repricing lands as OpenAI and Anthropic prepare trillion-dollar listings and SpaceX has already absorbed enormous public-market capital. Oracle's June earnings beat could not stop a 25% selloff once investors digested $70 billion in planned capex and $40 billion in fresh financing. Goldman notes S&P 500 ROE hit a record 22% — but the AI buildout is now a headwind to the mega-cap cohort that carried that number. Public markets will soon ask the same question bond desks are already asking: who earns durable margins in a cycle where the buyers are asset-heavy and the sellers have pricing power?

What to watch:

• Q2 hyperscaler earnings for capex guidance revisions and margin compression
• Memory contract pricing as the SK Hynix/Micron sold-out window extends into 2027
• Whether Goldman’s seven-point ROE decline materializes across Microsoft, Amazon, Alphabet, and Meta
• OpenAI and Anthropic IPO timing as capital markets absorb SpaceX’s debut

Convina's view: Goldman just named the trade Wall Street had been acting on without admitting it. For two years, the smart money rode the Magnificent 7 because AI looked like pure software economics — high margins, low capex, infinite leverage. That era is over. The companies building frontier models still need the infrastructure, but the infrastructure bill is converting asset-light giants into depreciation machines that borrow to grow. The winners in this phase are not the labs writing press releases. They are the suppliers with multi-year order books and the pricing power to match. Every enterprise AI buyer should plan for a world where compute stays expensive, capacity stays rationed, and the vendors underwriting your model access are fighting shareholders over who pays for the buildout.