Executive summary: the number is not fixed
Most analysis of AI infrastructure capital treats spend as a demand problem — will revenue materialise to justify the outlay? But the investment figure itself is highly malleable, shaped by a handful of supply-side assumptions that are rarely examined with the same rigour.
- 01 The economic service life of AI accelerator chips — where modest changes in replacement cadence move cumulative expenditure by hundreds of billions of dollars.
- 02 The cost and structural complexity of next-generation data facilities, which are rising rapidly as power density and system integration requirements escalate.
- 03 The mix of chip architectures deployed, whose net effect on total spend hinges on whether AI compute demand is elastic or inelastic.
- 04 Timeline elongation stemming from power, labour, and equipment constraints — which in severe scenarios can feed demand-side uncertainty.
The physical weight behind every prompt
Each AI query appears effortless — text in, text out. But behind every response lies a deeply material infrastructure: millions of specialised processors, vast networks of cabling, industrial cooling systems, and power consumption rivalling that of mid-sized nations.
Current market estimates place total AI infrastructure capital investment over the next five years somewhere between $4 trillion and $8 trillion — deployed across new chips, expanded data-centre campuses, and purpose-built power generation. The standard framing of this debate centres on demand: will AI adoption, monetisation, and productivity returns justify spending at this scale?
The required volume of capital is itself far more uncertain than published estimates suggest — sensitive to a small cluster of infrastructure assumptions that have received comparatively little scrutiny.
But there is an equally pressing supply-side question: how much capital does the build-out actually require? The answer is not fixed. It shifts materially depending on assumptions about how infrastructure is designed, refreshed, and constrained.
Not all assumptions carry equal weight. A small number drive the aggregate scale of required capital, while many others — despite dominating market commentary — influence only timing, margin distribution, or who captures value within the ecosystem.
The four most consequential variables for total capital requirements are: the economic useful life of AI chips; the construction cost and complexity of next-generation data centres; the translation of architectural chip choices into system-level costs; and the elongation of deployment timelines due to physical and institutional constraints.
Starting point: ~$7.6 trillion, 2026–2031
Our reference model anchors to current market consensus on accelerator chip deployment and then derives the implied requirements for data-centre construction and power infrastructure. It does not attempt to forecast end-user demand; it provides a consistent baseline against which supply-side sensitivities can be applied.
Under this baseline, annual AI capital expenditure rises from $765 billion in 2026 to approximately $1.6 trillion in 2031, implying roughly $7.6 trillion in cumulative investment across the period. The breakdown spans compute hardware (the dominant cost), data-centre construction, and power infrastructure.
The core unit of AI infrastructure is the accelerator — a processor built specifically for the parallelised computation that AI workloads require. Today's leading systems integrate dozens of these chips into a single rack, connected by high-speed backplanes and linked across facilities by hundreds of thousands of kilometres of optical cabling. These configurations generate extreme heat, demand industrial-grade liquid cooling, and require data centres with dedicated power delivery and redundancy systems far beyond what conventional cloud facilities were designed for.
The economic service life of AI silicon
Silicon service life is the single most influential variable in determining cumulative AI infrastructure spending. Small shifts in replacement cadence, compounded across hundreds of thousands of devices, produce differences measured in hundreds of billions.
AI accelerators are currently assigned useful lives of four to six years under standard accounting treatment. This reflects a tension at the heart of the technology: on one side, rapid generational performance gains push operators toward earlier replacement; on the other, the expanding range of AI workloads means that older silicon can still deliver economic value for longer.
The cadence of new chip architecture releases — now effectively annual, with each generation delivering step-function rather than incremental improvements — makes the tension sharper. Many analysts argue that prevailing depreciation schedules are too slow to reflect the economic reality of how quickly hardware becomes suboptimal.
An accelerator booked at $50,000 and depreciated over five years carries $10,000 in annual expense. When a successor chip arrives offering dramatically better performance per dollar, the operator continues bearing the cost of an asset that no longer delivers commensurate value.
At scale — across entire data-centre campuses — this dynamic can create a material gap between accounting depreciation and operational reality. One counterweight is the emergence of a tiered deployment model: while leading-edge chips serve cutting-edge training workloads, preceding generations can be redirected to less performance-sensitive tasks such as lower-complexity inference, edge deployment, emerging-market compute, and synthetic data generation. Current rental rates for prior-generation hardware suggest market-implied useful lives of five to six years or more — though it remains unclear whether this reflects genuine multi-tier value or simply the acute scarcity of today's AI capacity market.
| Year | Compute CapEx | 3 years | 4 years | 5 years ✦ | 6 years | 7 years |
|---|---|---|---|---|---|---|
| 2026 | $494bn | $165 | $124 | $99 | $82 | $71 |
| 2027 | $661bn | $385 | $289 | $231 | $196 | $165 |
| 2028 | $808bn | $655 | $491 | $393 | $327 | $281 |
| 2029 | $934bn | $801 | $724 | $580 | $483 | $414 |
| 2030 | $1,073bn | $939 | $869 | $794 | $662 | $567 |
| 2031 | $1,127bn | $1,045 | $986 | $921 | $850 | $728 |
| 2026–2031 total | $5,098bn | $3,989 | $3,482 | $3,017 | $2,597 | $2,226 |
Buildings and power infrastructure are long-lived assets depreciated over 20 and 25 years respectively. Silicon is not. The mismatch between these cycles is what makes accelerator replacement cadence such a consequential — and underappreciated — lever on aggregate AI capital requirements.
Data-centre cost and structural complexity
The cost of constructing AI-grade data infrastructure has risen substantially relative to conventional cloud facilities — and continues to move higher as AI workloads push power density and integration requirements to new levels.
Traditional hyperscale cloud facilities were designed around relatively modest power densities of 5–15 kilowatts per rack. Current AI deployments operate at 130–200 kW per rack, with next-generation configurations projected to exceed 500 kW. This is not a quantitative change — it requires a fundamentally different engineering philosophy.
Compute, memory, networking, cooling, and power delivery are no longer designed and layered independently: they must be co-engineered from the outset, shrinking failure domains and raising the consequences of localised outages. The result is that data centres are increasingly built as tightly integrated systems rather than modular commodity infrastructure.
| Generation | Chip architecture | Rack scale | Power / rack | Facility scale | Cooling |
|---|---|---|---|---|---|
| Gen 1Cloud facility | x86 / ARM CPU | Variable | 5–15 kW | Tens of MW | Air |
| Gen 2Early AI retrofit | GPU (Hopper-class) | 8 GPUs / rack | ~40 kW | Tens of MW, tens of thousands of GPUs | Air |
| Gen 3Transitional AI | GPU (Blackwell-class) | 144 GPUs / rack | 130–200 kW | 100s of MW, 100s of thousands of GPUs | Liquid / air hybrid |
| Gen 4AI factory | GPU (Rubin / next-gen) | 576+ GPUs / rack | 500+ kW | >1 GW, millions of GPUs | Liquid only |
These design shifts translate directly into higher capital costs per megawatt. Whereas conventional hyperscale facilities were often constructed at around $10 million per MW, next-generation AI data centres are increasingly priced at $15–20 million per MW, with further upside risk as density and redundancy requirements continue to escalate.
| Year | $11mn / MW | $13mn / MW | $15mn / MW ✦ | $17mn / MW | $19mn / MW |
|---|---|---|---|---|---|
| 2026 | $170 | $201 | $232 | $263 | $294 |
| 2027 | $220 | $260 | $300 | $340 | $380 |
| 2028 | $259 | $306 | $353 | $400 | $447 |
| 2029 | $288 | $340 | $392 | $445 | $497 |
| 2030 | $318 | $376 | $433 | $491 | $549 |
| 2031 | $320 | $378 | $436 | $494 | $553 |
| Total CapEx | $1,574 | $1,861 | $2,147 | $2,433 | $2,720 |
There is an additional risk embedded in rising design complexity. Cloud data centres built in the 2010s were designed to operate for 15–20 years. But AI system design is evolving so rapidly that facilities commissioned just two years ago may already be architecturally inadequate for the next generation of hardware. When the design requirements of a physical asset can shift materially within years of completion — and when transformative innovations such as novel cooling paradigms or new power delivery architectures could reshape the data-centre model altogether — the traditional durability of this asset class becomes as much a source of risk as an advantage.
Chip architecture mix and system-level cost
Most AI compute today runs on GPU hardware from a single dominant vendor. But a growing share of workloads is expected to migrate toward custom silicon — application-specific integrated circuits designed for particular tasks rather than general-purpose parallelism.
Custom silicon offers real efficiency advantages. These chips typically deliver lower cost per unit of useful compute, and in some configurations better power efficiency, than general-purpose GPUs. The dominant merchant GPU vendor earns gross margins of roughly 75% on data-centre hardware — far above alternative providers — creating a structural economic incentive for large buyers to develop or procure custom alternatives.
Organisations build toward a fixed compute target. Cheaper silicon translates directly into lower total capital requirements. Chip architecture choice becomes a meaningful lever on aggregate spend — redirecting value from hardware vendors toward buyers.
Lower compute costs unlock expanded usage: larger models, longer training runs, broader deployment. Total infrastructure footprint remains broadly similar. Architecture mix reshapes margin distribution rather than aggregate capital requirements.
The resolution of this question is not yet clear, and the answer may differ by workload type and market phase. Our baseline scenario aligns more closely with the elastic case — chip mix reshapes the composition of spend and the distribution of returns rather than the total volume of capital deployed.
Over time, the balance may shift. As AI workloads become more inference-dominated and margin-sensitive, and as the returns to incremental compute scale begin to diminish, lower-cost architectures could begin to constrain total spending rather than simply expand utilisation. That dynamic remains plausible but does not yet define the current phase of the infrastructure build-out.
Build-out elongation from structural bottlenecks
Elongation refers to the widening gap between capital commitment and the actual delivery of operational compute capacity. It does not alter the per-unit cost of AI infrastructure — but it introduces risks that can ultimately affect the scale of the investment itself.
The sources of elongation are varied: power interconnection queues, planning and permitting processes, shortages of specialist engineering labour, and long lead times for critical equipment including high-voltage transformers, switchgear, turbines, and industrial cooling systems. Each of these constraints is capable of extending the gap between capital deployment and productive capacity by months or years.
In the baseline scenario, bottlenecks slow the build-out without reducing its ultimate scale. Projects slip; phases extend; capital is sometimes duplicated through workarounds — with behind-the-meter power generation being the most visible current example. The result is an investment programme that is less efficient and more drawn out than announced roadmaps imply, but not materially smaller in aggregate.
When enough projects stall simultaneously, attention migrates from supply-side mechanics to demand-side doubts — whether revenue will materialise on a timeline that justifies the capital at risk. At that point, elongation becomes a feedback loop.
The more consequential risk is a stress scenario in which persistent bottlenecks begin to undermine confidence in the investment case itself. Supply-side friction can, under sufficient pressure, introduce demand-side doubt — leading to deferred or reduced investment plans. The current environment sits closer to the base case, but the margin is not wide. At the scale of capital now being committed, even modest delays invite real scrutiny of the underlying demand assumptions used to underwrite these investments.
Elongation is broadly negative for all participants in the ecosystem. Credit providers face extended duration exposure; offtakers absorb risk through take-or-pay contracts signed against uncertain timelines; equity-backed platforms must sustain investor confidence through prolonged periods of capital deployment without commensurate cash flow.
What doesn't move the aggregate number
Several dynamics attract significant commentary but do not materially alter the aggregate scale of capital required over the medium term. They matter — often greatly — for returns, volatility, and the distribution of value across the ecosystem. They do not, however, move the headline CapEx figure.
Training vs. inference workload mix
The balance between training and inference primarily affects the timing of economic realisation — not the scale of infrastructure. A faster pivot to inference accelerates revenue from a fixed capital base; a training-heavy phase extends the ROI timeline. The aggregate infrastructure footprint is broadly unchanged in either case.
Memory density and pricing volatility
Per-chip memory continues to grow, reflecting longer context windows and more stateful workloads — but this trajectory is already embedded in current system designs and pricing. Near-term memory price volatility, even when dramatic, reflects supply-demand imbalance rather than permanent structural change, and has limited impact on long-run total infrastructure spend.
Behind-the-meter vs. grid power
Captive power generation does increase absolute power infrastructure spend relative to grid-connected alternatives, since it replaces shared assets with bespoke project-level solutions. But power remains a relatively small share of total AI infrastructure investment. Even a wholesale shift in power sourcing strategy is unlikely to materially alter aggregate ecosystem-wide capital requirements.
There is a broader pattern here: supply chain volatility, component pricing swings, and architectural announcements all generate market noise. Many of these dynamics are genuinely significant for understanding who earns returns and when. But the underlying capital envelope — the total volume of investment required to build the infrastructure AI demands — is far less sensitive to these factors than the amplitude of commentary around them might suggest.
What actually moves the number — and what doesn't
The debate around AI infrastructure capital is typically cast as a question about demand: will adoption, monetisation, and productivity gains ultimately justify this level of spending? That framing is important. But it obscures an equally important question: how much capital does the build-out actually require, and under which conditions does that figure rise or fall?
Our analysis suggests that current estimates — however large — are far more conditional than they appear. The true figure is highly sensitive to silicon service life assumptions, data-centre construction costs, and the speed at which bottlenecks resolve. As these inputs shift, so will the estimates.
The most important wildcard may be discontinuous innovation. Current projections are built on current technologies. A genuine step-change in training or inference efficiency — one that materially reduces compute complexity at scale — could reshape the entire investment landscape. The market's reaction to algorithmic efficiency gains in early 2025 offered a glimpse of how quickly this reappraisal can occur, even when the underlying shift ultimately proved more limited than initial signals suggested.
There is a circularity worth noting. Much of what makes this build-out difficult is the physical, institutional, and economic weight of deploying capital at this scale. But if the ecosystem succeeds in overcoming those constraints — if the infrastructure is built, the bottlenecks are cleared, and compute costs continue to fall — the historical arc of technology suggests the result will not be surplus capacity. It will be new categories of demand and application that were not economically viable at higher cost points. The success of today's build-out may be precisely what ensures it is insufficient for tomorrow's.
For investors and operators, the critical discipline is to identify which assumptions underpin current plans — and to stress-test how resilient those plans are to changes in each.
This report has been prepared by Lualdi Advisors for informational purposes only. It draws on publicly available market data, industry disclosures, and Lualdi Advisors' proprietary scenario modelling. All quantitative figures are intended to illustrate sensitivity to key assumptions and do not constitute forecasts, consensus expectations, or investment projections. Forward-looking statements, illustrative computations, and scenario analyses are presented for analytical purposes only. Past performance is not indicative of future results. This material does not constitute investment, legal, tax, or financial advice, and should not be used as the basis for any investment decision. Lualdi Advisors makes no representations or warranties, express or implied, regarding the accuracy or completeness of the information contained herein. Company and product references are illustrative only.
