Goldman Sachs published a report this week projecting $7.6 trillion in cumulative AI infrastructure capital expenditure between 2026 and 2031. The number is extraordinary. It exceeds the annual GDP of every country except the United States and China. It implies a sustained annual infrastructure buildout of roughly $1.2 trillion, growing to $1.6 trillion by 2031. The report is careful, rigorous, and worth reading in full. It is also explicit about something the industry rarely acknowledges openly: the entire projection rests on a small set of assumptions about silicon replacement cycles, data center construction costs, chip architecture mix, and power infrastructure economics that have never been stress-tested at anything approaching the scale the model requires.
Change a few of those assumptions, and the $7.6 trillion figure moves by hundreds of billions of dollars in either direction. That sensitivity is the most important thing the report reveals, and it is receiving far less attention than the headline number.
Demand Is Clear. The Supply-Side Assumptions Are Not
The AI infrastructure investment thesis is broadly understood as a demand-side argument. AI workloads require compute. Compute requires data centers. Data centers require power. Therefore capital investment in all three is justified by the structural demand for AI capability that enterprise adoption will generate over the coming decade. That logic is sound as far as it goes. What it does not address is the supply-side assumption set that determines how much capital actually needs to be deployed to meet a given level of demand.
Goldman Sachs identifies four assumptions that move the capex figure most significantly. The economic useful life of AI silicon. The cost and complexity of next-generation data centers. The chip and architecture mix. And the elongation risk from power, labor, and equipment bottlenecks. Each of these assumptions carries uncertainty that compounds with every year the projection extends. By 2031, the cumulative uncertainty is substantial enough to make the $7.6 trillion figure a central estimate in a wide distribution rather than a reliable forecast.
The Silicon Useful Life Problem
The most consequential assumption in any AI infrastructure investment model is how long a GPU or custom AI accelerator remains economically productive before it must be replaced. A silicon useful life of three years versus five years changes cumulative capex by hundreds of billions of dollars, because it determines the replacement cycle that drives ongoing hardware investment. The Goldman Sachs baseline assumes a useful life that reflects current industry practice, but current industry practice is itself in flux. Nvidia’s shift to an annual release cadence means that each successive GPU generation delivers substantially higher performance per watt than its predecessor. An operator who deployed H100 clusters in 2023 faces pressure to upgrade to Blackwell hardware in 2025 and Vera Rubin hardware in 2026 or 2027, not because the H100 hardware has failed but because the performance gap has grown large enough to affect competitive economics.
The practical question is whether the AI workload economics of 2028 will justify running 2024-era hardware, or whether competitive pressure will force replacement at a cadence faster than the investment models assume. For training workloads at the frontier, where the race to the most capable model drives replacement decisions, useful life may be shorter than three years. For inference workloads serving stable production applications, useful life may be longer than five years. The aggregate replacement cycle depends on the mix of workload types across the installed base, and that mix is shifting as AI adoption moves from training-heavy frontier development toward inference-heavy enterprise deployment.
The assumption that determines hundreds of billions of dollars of projected capex is therefore not a fixed engineering parameter but a dynamic market outcome that will be shaped by competitive dynamics, model efficiency improvements, and enterprise adoption patterns that nobody can forecast with confidence five years out.
The Data Center Cost Trajectory
The Goldman Sachs model assumes $15 million per megawatt for data center construction costs. That figure reflects current market conditions, but current market conditions are themselves a function of supply chain dynamics, labor market conditions, and materials costs that are not stable. Cooling infrastructure costs for liquid-cooled AI facilities are substantially higher than for conventional air-cooled data centers, and the share of AI facilities requiring liquid cooling is rising with every GPU generation. The structural steel, copper, and aluminium costs embedded in data center construction are subject to commodity price volatility and tariff risk that can move project economics significantly. Power infrastructure costs, at $2,500 per kilowatt in the Goldman Sachs baseline, are subject to the same supply chain pressures that are driving transformer lead times to three to five years and switchgear delivery windows beyond 2028.
Data center construction costs have risen faster than general construction inflation for three consecutive years as AI-specific demand has concentrated procurement in a market not sized for this volume. Operators locking in construction contracts today pay a premium over the costs embedded in investment models written 18 months ago. The Goldman Sachs model addresses this through a brownfield adjustment that assumes a growing share of AI capacity will be deployed in existing facilities rather than new builds, reducing average construction costs. That assumption is sensible. However, the pace at which brownfield deployments can accommodate the density requirements of successive GPU generations remains uncertain, because facilities built for NVIDIA H100 specifications may require significant retrofitting to support NVIDIA Vera Rubin or NVIDIA Feynman-era hardware.
What the Sensitivity Analysis Actually Reveals
The most valuable section of the Goldman Sachs report is not the headline projection but the sensitivity analysis that shows how the $7.6 trillion central estimate changes under different assumption sets. Longer silicon lifecycles reduce cumulative capex by hundreds of billions by extending replacement cycles. Greater AI efficiency further lowers required investment by reducing compute intensity per unit of economic output. Persistent supply chain elongation, however, defers rather than reduces capex, altering timing without changing totals and materially impacting returns tied to deployment schedules.
What the sensitivity analysis reveals is that the $7.6 trillion headline is a scenario, not a forecast. It is the output of a model whose inputs are themselves uncertain in ways that matter enormously for the capital allocation decisions that institutional investors, infrastructure developers, and enterprise AI buyers are making right now. As covered in our analysis of the time-to-power crisis as AI’s hidden scaling ceiling, the physical infrastructure constraints on AI data center development interact with each other in ways that compound uncertainty rather than averaging it out. A model that treats power costs, silicon replacement cycles, and data center construction costs as independent variables will underestimate the correlations between them that emerge when all three are under simultaneous stress.
The Assumption the Model Cannot Price
The assumption that no AI infrastructure investment model can adequately price is the possibility that the AI capability trajectory itself changes in ways that alter the compute demand picture. Every projection of AI infrastructure investment is implicitly a projection of AI capability advancement at a pace that justifies continued large-scale hardware deployment. If model efficiency improvements outpace hardware capability improvements at a sustained rate, the compute required per unit of AI economic output declines, and the capex required to serve a given level of AI demand declines with it.
DeepSeek’s demonstration in early 2026 that a highly efficient model could achieve competitive performance at a fraction of the compute cost of leading Western models was a single data point in a trend that has been visible for years. Token costs have fallen by roughly 280 times over the past two years. The efficiency trajectory and the hardware deployment trajectory are moving in opposite directions, and the resolution of that tension will determine whether the $7.6 trillion materialises as projected.
When the Headline Escapes the Model
No one knows the answer. The Goldman Sachs model lays out its assumptions transparently and applies rigorous sensitivity analysis. It does what good analytical work should do: make the structure of uncertainty explicit rather than hide it behind a single confident projection. The issue lies not with the model itself, but with how people use it. Once the headline number enters a press release, an earnings call, or an investor presentation, it quickly escapes its uncertainty bounds. The market then treats $7.6 trillion in cumulative AI infrastructure capex as a shared assumption about the cost of the AI era, while losing sight of the assumptions that place the figure closer to $5 trillion or $10 trillion.
The investors, operators, and enterprise buyers making commitments today deserve to understand not just the central estimate but the distribution around it. The distribution is wide. The tails matter. And the industry has a structural incentive to emphasise the centre and minimise the tails.
