Goldman Sachs forecasts AI infrastructure spending will climb toward $1.1 trillion annually, even as valuations of pure-play AI infrastructure companies are compressing.

AI infrastructure stocks have become cheaper relative to their projected earnings even as technology companies continue to increase spending on data centers and computing capacity, according to Goldman Sachs research. The median stock in the group traded at 22 times expected earnings, down from 32 times in April—a decline that suggests investors are questioning how long today's earnings growth can last.

The highest expectations persist in optical networking, electrical grid and liquid cooling stocks. Valuations for semiconductor and utility stocks, by contrast, reflect more cautious forecasts.

Memory chip stocks present a stark picture, trading at about four times earnings projected two years ahead—roughly half their 15-year average, despite gross margins near 80%. Goldman said earnings would have to fall approximately 50% for their valuations to return to the historical average. Chip designers that outsource manufacturing show a different balance: their margins remain closer to historical levels, and a roughly 10% decline in earnings would bring valuations back to normal. Goldman maintained Buy ratings on Nvidia and Broadcom and a Neutral rating on Micron Technology.

Large cloud providers also traded at their smallest valuation premium to the median S&P 500 company in more than a decade. Slower growth in capital spending and clearer prospects for free cash flow could support those stocks.

The scale of investment remains the central question. Goldman estimated that U.S. cloud providers are on track to spend $800 billion on capital projects this year, rising to $1.1 trillion in 2027. The firm calculated that providers would need about $300 billion in annual AI revenue to break even on their 2026 and 2027 investments. Estimated AI revenue remains below that level, although major cloud providers' combined backlogs exceed $1.7 trillion.

For those investments to deliver strong returns, Goldman estimated that applications built on AI infrastructure would ultimately need to produce roughly $1 trillion in annual revenue.