Goldman identifies power and compute as key AI bottlenecks

Goldman's note directs investors to the physical infrastructure of AI, pointing out that securing high-voltage power is now a bigger hurdle than having capital for chips. That gives an edge to firms that already have grid connections, and indicates that power availability, rather than chip supply, may determine how fast new data centres can be built through 2027. On the software side, spending shifts toward data and cybersecurity suggest a growing divide between foundational platforms and application providers, whose pricing strategies could face pressure as agents replace human users. The focus on returns from core platforms also suggests investors might favour proven compute monetisation over unproven new revenue streams.

---

According to Goldman's analysis of the AI surge, having money allows purchasing GPUs, but power determines who can actually operate them.

Summary:

• According to Goldman, AI infrastructure demand continues to exceed available capacity, and the gap between compute supply and demand is expanding, based on conversations with management teams at technology firms.
• A projected deceleration in model training is being counterbalanced by increased spending on AI safety and pacing monitoring, which sustains high compute demand.
• Major hyperscale operators are locking in power deals for 2028 and later, as grid capacity for 2026 and 2027 is nearly completely taken.
• Capital expenditure is increasing due to component price inflation and demand for additional chips, while revenue per unit of power has climbed by roughly two-thirds over the last year.
• Goldman identifies data and security platforms as some of the most obvious winners from AI investment, and states that application software is increasingly designed for AI agents instead of humans.
• Meta achieves two to three times ROI per megawatt in its primary operations, while Google employs a full-stack strategy that includes custom chips, cloud services, security, and distribution.

Goldman Sachs says AI infrastructure demand keeps outpacing capacity, and the constraint is shifting from chips to power, based on its recent talks with tech company management teams.

The investment bank stated that the supply-demand gap for compute is expanding, due to fresh infrastructure needs and steady corporate spending. Goldman forecasts a slowdown in model training, but notes that firms are increasing outlays on AI safety and pacing monitoring, tasks requiring more computing power, thereby maintaining elevated compute demand overall.

According to Goldman, power is now the primary bottleneck. Grid capacity for 2026 and 2027 is all but fully reserved, the bank noted, prompting leading hyperscalers to aggressively pursue power contracts for 2028 and later. Goldman argued that simply having capital for GPUs is insufficient; access to high-voltage energy infrastructure now determines who can expand. Capital expenditure is rising because of component inflation and the need for more chips. Meanwhile, revenue per unit of power has increased by roughly two-thirds compared with earlier price points over the past year.

Goldman also highlighted a structural change in enterprise software. It thinks data and security platforms are some of the most obvious beneficiaries of AI spending, since AI requires clean, structured, and secure data. CIOs are increasingly directing budgets toward basic data layers and cybersecurity tools, aiming to safeguard data integrity and limit risks from unauthorized "shadow AI" within organizations.

Application software is becoming "headless," meaning it is increasingly created for AI agents instead of human users. Goldman said this enables vendors to integrate more advanced AI capabilities into current platforms and currently fuels industry expansion. However, it warned that this shift could ultimately change how software firms price and monetize their offerings.

Among the firms it spotlighted, Goldman noted that Meta continues to achieve two to three times ROI per megawatt in its main business while seeking new incremental revenue streams. Google, the bank said, leverages a full-stack edge combining custom chips, cloud, security, and a massive distribution network to boost core monetization. In this context, TPU chip sales are just one element of a wider portfolio strategy.

Goldman concluded that AI monetization provides significant optionality. Nonetheless, in its view, core platform monetization remains one of the more reliable and profitable uses of compute, suggesting that established platforms already generating strong returns on their power may continue to be the more dependable winners while capacity stays limited.