Wall Street has spent the past few weeks piling back into AI stocks like the doubts from earlier this year never happened. Goldman Sachs just handed those investors some math that is harder to wave away than a single bad earnings quarter.

The bank’s own strategists are not calling the AI trade a bubble. They are simply pointing out how far hyperscalers still have to travel before their spending actually pays for itself, and the distance is bigger than most investors probably realize.

The $300 billion break-even number

Goldman Sachs strategist Ryan Hammond crunched the numbers and landed on a specific bar hyperscalers need to clear. Companies like Amazon, Oracle and Microsoft will need roughly $300 billion in annual AI revenue in the next few years just to break even on their investments.

That is not a small ask relative to where things stand today. Hyperscaler cloud revenues have accelerated sharply this year, annualizing at about $70 billion above the pre-AI trend in the second quarter of 2026. That leaves a meaningful gap between current results and the $300 billion target, according to Yahoo Finance.

More Goldman Sachs:

  • Goldman Sachs sees writing on the wall for Eli Lilly stock
  • Goldman Sachs doubles down on oil price forecast for 2026
  • Goldman Sachs gives major reset to YETI Holdings stock target

The pipeline of future business looks healthier than current revenue does. Announced backlogs across the hyperscaler group already exceed $1.5 trillion, giving companies a substantial pipeline of future business still waiting to convert into recognized revenue.

Hammond was careful to frame this as a real gap rather than a reason to panic. He noted that market participants are skeptical about the durability of AI infrastructure returns. That skepticism is already showing up, at least partially, in how the market prices these stocks, as reported by Investing.com.

Why the math gets even bigger from here

Breaking even is only the first hurdle Goldman describes. For hyperscalers to generate what Hammond calls solid returns on investment, and for the application layer built on top of that infrastructure to earn strong profit margins, Goldman estimates AI users would need to spend roughly $1 trillion annually on AI applications.

The spending required to get there keeps climbing too. Goldman now expects the five largest U.S. hyperscalers to lift AI infrastructure capital expenditures to $1.2 trillion in 2027 –– above the Wall Street consensus of $1.1 trillion, and a sharp jump from the roughly $800 billion the group is on pace to spend this year, Bloomberg reported.

Goldman’s own strategists put that spending scale in historical context. Based on consensus estimates, capital expenditures in 2027 are on track to reach a larger share of GDP than any technology investment cycle since the railroad buildout of the late 1800s.

That pace of growth is not expected to continue indefinitely. Goldman’s model shows hyperscaler capex growth stepping down to 54% in 2027 and then to just 12% in 2028, when total hyperscaler capital spending is projected to reach $1.4 trillion.

The pipeline of future business looks healthier than current revenue does.

Michael M. Santiago / Getty Images

The buyback squeeze behind the spending

All of that capital has to come from somewhere, and Goldman’s research shows exactly where hyperscalers have been pulling it from. Buybacks among the five largest hyperscalers fell 64% year over year in the first quarter as companies redirected cash toward data centers, chips and related infrastructure instead.

Goldman has described this pattern as a broader rotation from buybacks to capital expenditures and research spending, noting that the five AI hyperscalers accounted for roughly 32% of all S&P 500 capital expenditures in 2025 alone.

The financing pressure is starting to show up in how companies fund themselves rather than just how they spend. Goldman Sachs credit strategists project hyperscalers will finance more than a third of their 2027 capital expenditures with debt. Alphabet has also sharply scaled back buybacks, repurchasing no stock at all in one recent quarter, compared to $13.2 billion in buybacks the year before, reported by TheStreet.

The broader market has not felt the same squeeze yet. Goldman still expects total S&P 500 buybacks to exceed $1 trillion this year, with other sectors stepping in to keep the overall total climbing even as the biggest AI spenders pull back their own repurchase activity.

What it means for investors

Investors have shown little appetite to wait for the math to catch up before buying back in. The Roundhill Magnificent Seven ETF, which tracks the largest AI hyperscalers, is up 8% in the past month, compared with a far more modest gain for the S&P 500 over the same stretch.

Goldman itself is not arguing that enthusiasm is misplaced, only that the spending cycle will take time to translate into returns. That patience is echoed elsewhere on Wall Street too. Morgan Stanley is separately estimating hyperscaler spending will reach nearly $1.1 trillion in 2027, even as political resistance in some states could reshape where new data centers actually get built.

According to TheStreet, the bank has pointed to AI infrastructure stocks posting cumulative earnings estimate increases of 59% since January 2025, compared with just 9% for the S&P 500 overall. Showing how sharply expectations for the group have already moved higher even before the $300 billion revenue bar is cleared.

For now, the AI trade remains a bet that revenue eventually grows into spending rather than the other way around. Goldman’s own numbers suggest that the bet still has years to run before investors find out whether it pays off.

Related: Morgan Stanley has a strong message for worried AI stock investors 

​