The global artificial intelligence build-out is entering a capital-intensive phase that extends well beyond semiconductor production. According to Goldman Sachs, annual AI infrastructure capital expenditure could increase from $765 billion in 2026 to approximately $1.64 trillion by 2031, implying roughly $7.6 trillion of cumulative spending across compute, data centers and power infrastructure over the period.

The scale of the projected investment highlights how AI is becoming a major infrastructure cycle spanning technology, energy and physical assets. However, Goldman Sachs emphasizes that the $7.6 trillion figure is a scenario-based framework rather than a fixed forecast, with the ultimate capital requirement dependent on several structural assumptions.

Compute Remains the Core Driver of AI Capital Spending

Compute represents the largest component of the projected build-out. The Goldman Sachs baseline estimates $494 billion of compute spending in 2026, rising to approximately $1.127 trillion annually by 2031. The model also projects compute spending of $661 billion in 2027, $808 billion in 2028, $934 billion in 2029 and $1.073 trillion in 2030.

This trajectory reflects the growing physical requirements of AI workloads, where specialized accelerators, networking and memory must operate at increasingly large scale. Goldman Sachs notes that the useful economic life of AI silicon is particularly important because accelerators can become economically obsolete faster than conventional long-lived infrastructure. Changes in replacement cycles can therefore materially alter cumulative spending requirements.

The research also highlights a potentially important tension between technological progress and capital efficiency. Faster improvements in performance per dollar can encourage earlier hardware replacement, increasing capital requirements, while older processors may retain economic value for less demanding inference and other workloads.

Data Centers and Power Become Strategic Constraints

Compute cannot expand independently of the physical infrastructure required to operate it. Goldman Sachs estimates data-center spending of $232 billion in 2026, increasing to $436 billion by 2031, while power infrastructure spending rises from $39 billion to $73 billion over the same period.

The increasing complexity of AI workloads is changing the economics of data-center construction. Higher rack densities require more sophisticated cooling, power delivery and networking systems, while the integration of compute, memory and power infrastructure creates increasingly complex facilities. Goldman Sachs estimates that next-generation AI data centers can require substantially greater capital expenditure per megawatt than traditional hyperscale facilities.

Power availability is consequently becoming a central consideration for the AI industry. Interconnection queues, permitting requirements, equipment shortages and limited availability of specialized labor can delay projects even when capital has already been committed. These bottlenecks could affect not only deployment schedules but also the economics of the wider AI infrastructure ecosystem.

The $7.6 Trillion Estimate Depends on Several Critical Assumptions

Goldman Sachs’ analysis makes clear that the headline spending figure should not be treated as a predetermined outcome. The baseline assumes approximately $7.6 trillion of cumulative AI infrastructure capital expenditure between 2026 and 2031, but the final amount could change materially as technology and market conditions evolve.

Four factors are particularly important: the economic useful life of AI silicon, the cost and complexity of next-generation data centers, the mix between GPU and custom-chip architectures, and the extent to which physical bottlenecks elongate the build-out. The research also distinguishes these structural drivers from factors such as training versus inference workloads and memory pricing, which can significantly affect returns and timing without necessarily changing the overall scale of infrastructure required.

For investors, the significance extends across the technology and capital markets. The projected spending creates opportunities for semiconductor manufacturers, data-center operators, networking companies, utilities, power-equipment suppliers and infrastructure financiers, while simultaneously increasing scrutiny of capital efficiency and the eventual economic returns generated by AI investment.

Looking ahead, the key question will be whether AI demand continues to expand quickly enough to justify the infrastructure being built. Investors and companies will need to monitor accelerator replacement cycles, data-center construction costs, power availability, chip architecture changes and evidence of monetization. If infrastructure constraints ease and computing costs decline, lower-cost AI could generate additional demand and new applications; conversely, prolonged delays or weaker-than-expected returns could slow capital deployment. Goldman Sachs’ analysis ultimately suggests that the scale of the AI build-out is substantial, but its final economic footprint will remain highly dependent on how these underlying assumptions evolve.