Goldman Sachs Warns Ai Bubble Could Burst Datacenter Boom
Datacenter capacity is forecast to surge 50 percent by 2027 driven by AI demand, with the sector’s energy consumption doubling by 2030, according to the latest research from Goldman Sachs. But the financial services biz says it’s watching for signs that AI adoption may fall short of current hype.
The AI boom has created a “frenzied atmosphere” where major tech companies are “living in fear of being disrupted and deploying capital to play as much offense as they’re playing defense,” according to Goldman Sachs Managing Director Eric Sheridan.
Global datacenter capacity currently stands at approximately 62 gigawatts, according to research published by Goldman Sachs, with cloud workloads accounting for 58 percent, traditional workloads 29 percent, and AI just 13 percent – up from virtually nothing in early 2023.
By 2027, the prediction is that AI workloads will represent 28 percent of all capacity, while cloud’s relative share falls to 50 percent, and traditional workloads to 21 percent. This doesn’t imply cloud or traditional workloads are declining, just that AI is growing faster and taking up an increasing part of a larger overall market.
The scale of investment matches research from Omdia showing datacenter spending now rivals mid-sized economies, with Amazon alone investing over $100 billion annually – roughly equivalent to Costa Rica’s entire GDP.
The AI datacenter boom could see global semiconductor revenues double between 2024 and 2030, to reach more than $1 trillion, according to Counterpoint Research.
A key catalyst for growth, it claims, is advanced AI server infrastructure, driven by sustained and potentially rapid growth in demand for forthcoming AI applications. Most of this is likely to come from hyperscalers in both the short and longer term.
Counterpoint highlights “the token economy” with a huge increase in hardware capability needed to drive token generation for agentic AI applications.
AI training requires specialized hardware and this is driving dramatic changes. Where cutting-edge systems featured eight GPU accelerators per server just two years ago, by 2027 leading configurations will pack 576 GPUs into filing-cabinet-sized racks consuming 600 kilowatts – enough to power 500 US homes, Goldman Sachs notes.
As a consequence, datacenter power use is forecast to rise 165 percent globally by 2030, from 1-2 percent of global electricity in 2023 to between 3 and 4 percent by the decade’s end.
Goldman Sachs expects renewables to meet 40 percent of the extra power needed for bit barns, with modest nuclear expansion targeted at AI workloads. However, natural gas generators will supply the remaining 60 percent, adding between 215 and 220 million tons of greenhouse gas emissions by 2030 – equivalent to an extra 0.6 percent of global energy emissions.
Despite the bullish projections, analysts at the financial services biz are “on heightened alert” for any signs of market weakness that could derail forecasts. Risks include failure to monetize AI or innovations that make it cheaper to build and commoditize the models.
The base case for bit barn capacity assumes 17 percent compound annual growth reaching 92 GW by 2027, but scenarios range from 14 percent (if AI interest disappoints) to 20 percent (in a more bullish environment).
These concerns echo broader skepticism. OpenAI CEO Sam Altman recently acknowledged the industry is experiencing an AI bubble, while consulting firm McKinsey warned earlier this year that nobody really knows what future AI service demand will look like. ®
A considerable amount of time and effort goes into maintaining this website, creating backend automation and creating new features and content for you to make actionable intelligence decisions. Everyone that supports the site helps enable new functionality.
If you like the site, please support us on “Patreon” or “Buy Me A Coffee” using the buttons below
To keep up to date follow us on the below channels.
