5 hours ago
11
Wall Street’s favorite question about AI, “but when does it actually make money?,” keeps getting the same answer from Goldman Sachs: it doesn’t matter yet, because the spending isn’t slowing down.
Goldman’s 1-Delta desk, led by Rich Privorotsky, is making the case that AI infrastructure investment remains firmly intact even as broader markets have wobbled and skeptics have questioned the return on all those GPU purchases.
The numbers behind the conviction
Goldman’s research projects that hyperscaler capital expenditures will land somewhere between $527 billion and $765 billion in 2026. That’s not a typo, and it’s not a cumulative figure. That’s a single year.
For context, the previous estimate for 2026 hyperscaler capex sat at $465 billion. Goldman revised it upward to $527 billion on the low end, a bump of more than 13%.
Cumulative AI infrastructure spending is expected to reach $7.6 trillion by 2031, according to Goldman’s research.
The catalyst for this latest round of bullishness came from an unlikely corner: memory chips. Micron Technology reported increased demand for high-bandwidth memory, the specialized DRAM that sits on top of AI accelerators and feeds them data at blistering speeds. When Micron says orders are strong, it means someone downstream is building a lot of AI servers.
Open-source models are changing the math
Goldman’s desk flagged that open-source models like Gemini 3 Flash and DeepSeek V4 Pro have slashed tokenization costs by as much as 75%. In English: the price of actually running AI inference, the process of getting a trained model to produce useful outputs, has cratered.
The Goldman view is that lower costs drive more usage, which drives more infrastructure demand. Make something cheaper and people use dramatically more of it, not less.
If open-source models continue to close the gap with proprietary ones, the pricing power of companies that sell AI-as-a-service could erode over time. But the companies selling the infrastructure underneath may actually benefit from the resulting surge in total compute demand.
What this means for investors
When hyperscaler capex estimates get revised upward by tens of billions of dollars, it creates a floor under the companies that supply the physical backbone of AI. Chipmakers producing GPUs and HBM, server manufacturers, data center REITs, and power infrastructure companies all sit in the direct path of this spending.
For context, the entire global semiconductor industry generated roughly $527 billion in revenue in 2023. Goldman is projecting that hyperscaler capex alone could match or exceed that figure in a single year by 2026.
Goldman’s desk is essentially arguing that we’re still in the “build it” phase, not yet at the “prove it” phase, and the build cycle has more runway than the market is pricing in. A 75% reduction in tokenization costs is a seismic shift. If that trend continues, the winners won’t be the companies charging premium prices for AI access. They’ll be the ones selling the infrastructure that everyone, proprietary and open-source alike, needs to operate.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·