In a dramatic sign of artificial intelligence’s (AI) rapid growth, Meta and OpenAI say they are encountering “significant capacity constraints” in their supply of graphic processing units (GPUs), according to a note from Morgan Stanley analysts at the company’s Tech, Media & Telecom 2025 event.
The GPUs required to build AI models — most of which are produced by NVIDIA — enable the models to perform the ultra-fast data crunching needed for generative AI. Adding to the bottleneck, these supply-constrained GPUs are required for both inferencing and training AI models.
The report counters concerns among industry observers that enterprise AI investment is fueled by hype rather than real demand, and that the lack of real profit will soon derail AI’s growth. That both these companies are short of GPUs indicates that AI demand is running in excess of supply.
Indeed, OpenAI reported that its GPUs are “completely saturated,” according to the Morgan Stanley note. Meta said that “multiple teams are still waiting for GPUs.” And in case skeptics were wondering if AI is on track to generate robust profits, OpenAI said that there has never been a period in which “it can’t sell out access to its GPUs at reasonable margins.”
The Meta and OpenAI chip shortages carry heightened significance as the firms are part of a small group of pioneering AI model builders. Because developing AI models is so exorbitantly expensive, only a few companies build their own, with a far larger cohort of firms purchasing GPUs for secondary AI deployment and development. The tight supply even among Meta and OpenAI — with their primary focus on AI development — is a sign of a robust demand for AI services.
Indeed, Meta has even begun building its own in-house custom silicon for AI modeling, an expensive venture that indicates that the social media giant has a rock solid belief in its need for AI.
“For Meta and OpenAI, this isn’t merely about higher costs—it’s an existential ceiling on their ability to train next-generation models and maintain competitive positioning,” said Nick Patience, VP and practice lead for AI, The Futurum Group. “We’re witnessing a massive capital reallocation across the industry, with AI compute infrastructure now commanding investment priority over traditional chip segments.”
The constraints “are forcing both companies to make difficult strategic choices,” he added. “OpenAI must carefully allocate its Microsoft-backed compute resources, and is now seeking alternatives to supplement that. And Meta faces pressure to accelerate its custom silicon initiative despite the significant technical challenges involved. The companies that secure reliable access to advanced AI compute—either through partnerships, custom silicon or superior capital deployment—will gain sustainable advantages in the AI race that could last for years.”
These reports of shortages also address concerns raised by DeepSeek, which surprised the AI community when it unveiled an AI model built with lesser GPU resources than US-based models. Some experts opined that DeepSeek’s model would deflate the need for leading-edge GPUs, and NVIDIA’s stock plunged in the aftermath. Yet it now appears that cutting-edge AI development continues without a hint of slowdown. If anything, DeepSeek is likely to boost AI’s growth – and therefore the demand for GPUs – by democratizing access to AI model building.
Anticipating this growth, Morgan Stanley research predicts that revenue from generative AI could reach $1.1 trillion in 2028, an astounding leap from 2024’s $45 billion of revenue.
To begin this 20-fold increase over three years, “software and internet companies are expected to see a positive return on their AI investments as soon as this year, as the expanding functionality of GenAI prompts broader use and triggers a new technology cycle.” This exponential growth will require vast amounts of GPUs, so shortages could continue in the months and years ahead.
These GPUs will be deployed in a growing number of use cases. For instance, Morgan Stanley predicts that enterprises will spend $400 billion on AI-based automation to drive productivity improvements. “We took a look at the labor and wage calculations that GenAI can impact,” said Keith Weiss, head of Morgan Stanley’s U.S. Software Research, in the report. “And that’s where productivity gains come from.”
“We’ve heard from most of the big spenders in the market that they’re continuing to scale up,” Weiss said. “This is not a bubble. This is investment that’s driven by these areas of economic benefit that our software and internet teams are seeing.”
