Goldman Sachs Warns of AI Infrastructure Risks
Goldman Sachs is cautioning investors that AI infrastructure faces heightened risks in the second half of the year. The bank's analysis suggests that while the sector has seen significant gains, it may be vulnerable to market shifts. The warning implies that reliability and risk management for AI workloads are becoming critical, as infrastructure failures could have an outsized business impact.
The surge in AI infrastructure spending is monumental, with giants like Alphabet, Amazon, Meta, and Microsoft projected to collectively invest around $650 billion in 2026. This represents a nearly 60% increase from the $410 billion spent in 2025, as these companies reallocate capital from programs like share buybacks to fund the buildout of data centers, custom chips, and networking gear. A key risk flagged by Goldman Sachs is the potential for a sharp deceleration in this capital expenditure growth. Analysts warn that once the explosive growth rate peaks, the valuations of infrastructure stocks that have priced in continued acceleration could become "extremely vulnerable." An extreme scenario where spending reverts to 2022 levels could erase an estimated 30% of projected S&P 500 sales growth for 2026. This investment cycle is creating financial arrangements that warrant scrutiny. A pattern of "circular financing" has emerged, where hyperscalers and chipmakers invest in AI startups, who in turn use the capital to purchase GPUs and cloud services from their investors. This creates a web of interdependence that could be artificially inflating demand and valuations, echoing vendor financing models from the dot-com era. Physical constraints are also creating significant bottlenecks. The production of advanced AI accelerators is being capped by shortages in high-bandwidth memory (HBM) and advanced chip packaging technologies like CoWoS. Memory suppliers are reallocating production lines to the more profitable HBM, triggering shortages and price hikes for conventional DRAM and NAND flash memory used in other enterprise hardware. The energy required to power this new infrastructure is staggering. Data centers are projected to consume up to 9% of total U.S. electricity by 2030, roughly doubling their current share, with AI workloads being the primary driver. The International Energy Agency projects global data center electricity consumption could more than double by 2030, reaching levels comparable to the entire nation of Japan. This rapid, concentrated demand is already straining regional power grids. For financial platforms, the reliability of this infrastructure is a critical operational risk. System outages or failures, driven by either hardware faults or cyberattacks targeting the complex AI supply chain, can disrupt critical services. In electronic trading, where algorithms are dependent on real-time data and consistent compute power, such failures could lead to significant financial losses and erode trust in the system. There's also a growing debate around the true economic lifespan of AI hardware. While typically depreciated over 3-6 years, some analysts argue the rapid pace of innovation makes the useful life of an AI GPU closer to 2-3 years. This potential discrepancy could mean that corporate earnings are being inflated by understating depreciation, masking the true cost of the AI arms race.
