Is AI a Bubble? Goldman Sachs Projects $7.6 Trillion AI Infrastructure Over the Next Five Years

AI’s rapid evolution rests on a massive physical foundation: millions of processors, hundreds of thousands of kilometres of cabling, industrial‑grade cooling systems, and power consumption comparable to a mid‑size nation. Most discussions of an AI bubble focus on demand‑side revenue, but Goldman Sachs Global Research releases a supply‑side baseline model that avoids speculative demand forecasts.

Baseline Model and Investment Scale

Using Nvidia’s projected data‑center revenue as a reference, the model estimates AI CapEx of $765 billion in 2026, rising to $1.6 trillion by 2031, with a cumulative $7.6 trillion investment from 2026‑2031. The model treats four core variables as levers that can inflate or deflate the total spend.

Economic Lifespan of Compute Chips

AI accelerators, often numbering in the hundreds of thousands per data centre, have an economic lifespan of 4‑6 years. Physical wear and rapid performance gains drive both wear‑out and economic obsolescence. For example, a $50 k accelerator depreciated over five years costs $10 k per year; shortening the lifespan increases replacement cycles and amplifies total investment. Nvidia’s aggressive release cadence intensifies this tension, making the chip‑life assumption the most influential single variable.

Data‑Centre Cost Surge

AI workloads demand extreme power density, liquid cooling, and robust power‑fault tolerance, pushing per‑megawatt construction costs from roughly $10 million for traditional cloud facilities to $15‑20 million for AI‑optimized centres. Shorter design lifespans (15‑20 years for conventional clouds vs. a few years for AI) accelerate turnover, turning the per‑MW cost into a major driver of overall spend.

Architecture Mix and Construction Delays

The proportion of Nvidia GPUs versus custom ASICs forms a third lever. Inelastic demand scenarios allow cheaper chips to cut total spend, but elastic demand leads to larger models and longer training cycles, offsetting cost savings. Construction delays—grid interconnection, permitting, component lead times—extend timelines and raise coordination costs, though they do not change per‑MW construction costs.

Separating Noise from Real Cycles

Not every headline event reshapes long‑term investment. Shifts in training‑to‑inference ratios, memory capacity growth, or power‑source choices (grid vs. self‑generation) have limited impact on the $7.6 trillion baseline. Short‑term price volatility in memory, optics, or packaging is expected to normalize as capacity expands.

Conclusion

The AI infrastructure investment outlook hinges on underlying assumptions about building standards, chip turnover, and construction pace. While breakthrough technologies could disrupt these assumptions, the current forecast reflects existing technical pathways. Overcoming physical bottlenecks will enable new applications, but the massive capital outlay remains essential to sustain AI’s growth.