In brief
- Hyperscaler capital expenditure is approaching $760 billion and a depreciation wall is looming, prompting investors to demand proof of durable software revenue as the AI buildout slows from earlier exuberance.
- Three structural pressure points are driving a reprice: a mathematical divergence between capital deployed and end user software revenue, a depreciation wall compressing cloud operating margins, and physical bottlenecks in power transmission and generation increasing deployment costs.
- A Goldman Sachs Global Institute study titled Tracking Trillions models cumulative global capital expenditure at roughly $7.6 trillion from 2026 to 2031, with annual AI infrastructure spend rising from $765 billion in 2026 to more than $1.6 trillion by 2031.
- The global AI buildout is entering a more turbulent phase where investors seek measurable returns and balance sheets are scrutinised, shifting from narrative to a disciplined, asset based assessment.
For thirty-six months, global capital markets operated under a suspended state of financial gravity. Between late 2022 and early 2025, any public technology company pledging to accelerate its capital expenditure on artificial intelligence infrastructure was rewarded with surging equity multiples. The prevailing consensus among chief executives was captured by Alphabet's Sundar Pichai and Meta's Mark Zuckerberg: in a foundational technological transition, the risk of under-investing vastly outweighed the risk of building excess capacity.
By late 2026, that era of unconditional underwriting has drawn to an abrupt close.
The global AI buildout has entered a distinctly more turbulent, adversarial phase. While headline spending figures continue to climb to staggering heights, institutional investors are no longer content with visionary roadmaps or demonstrations of agentic reasoning. Instead, equity analysts, sovereign wealth allocators, and corporate boards are interrogating the fundamental balance sheets of the infrastructure boom.
Three structural pressure points have converged to force this repricing: an extraordinary mathematical divergence between capital deployed and end-user software revenue; an approaching "depreciation wall" that threatens to compress cloud operating margins; and severe physical bottlenecks in power transmission and generation that are stretching deployment schedules and driving up per-megawatt construction costs.
The Capital Tsunami: Sizing the $7.6 Trillion Edifice
To understand why investor scrutiny has intensified, one must first grasp the sheer scale of the balance-sheet commitments now underway.
According to tracking data compiled by Statista, the combined annual capital expenditure of the "Big Four" hyperscalers—Amazon, Alphabet, Microsoft, and Meta—is projected to reach between $720 billion and $760 billion in 2026 alone. To place this figure in historical perspective, this single-year allocation exceeds the inflation-adjusted cost of the Apollo space programme, the construction of the US Interstate Highway System, and the Panama Canal combined.
Individual corporate guidance outlines the breadth of the deployment:
Amazon is deploying approximately $200 billion annually across its AWS cloud data centre footprint, custom Trainium silicon, and global logistics networks.
Microsoft has committed to an annual capex run-rate approaching $190 billion to support Azure infrastructure, OpenAI compute agreements, and co-developed cluster campuses.
Alphabet has guided to $175 billion to $185 billion, aggressively expanding its custom Tensor Processing Unit (TPU) fleets and global subsea connectivity.
Meta, despite lacking an external public cloud platform to lease spare capacity, has targeted $115 billion to $135 billion to fuel its open-source foundational models, recommendation architectures, and internal compute clusters.
On a cumulative basis, the numbers become astronomical. In a landmark May 2026 study titled Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out, George Lee and Lucas Greenbaum of the Goldman Sachs Global Institute modelled the baseline physical and capital requirements implied by leading-edge semiconductor deployment.
The baseline model indicates that cumulative global capital expenditure across specialized compute silicon, advanced data centres, and dedicated power infrastructure will total roughly $7.6 trillion between 2026 and 2031. Under this baseline, annual worldwide AI infrastructure spend will expand from $765 billion in 2026 to more than $1.6 trillion annually by 2031.
As the report observes, the central intellectual challenge of this buildout is that it rests upon deeply physical, highly capital-intensive assets that must be provisioned years in advance of the revenue streams intended to justify them.
The Monetisation Asymmetry: From the $600 Billion to the $1.5 Trillion Question
The core grievance animating institutional investors is the glaring disparity between the capital poured into the ground and the incremental revenue generated by software applications.
In mid-2024, venture capitalist David Cahn at Sequoia Capital published a widely circulated critique titled AI's $600B Question. Cahn calculated that given the hardware purchase forecasts of semiconductor suppliers (principally Nvidia), the AI ecosystem would need to generate $600 billion in annual software revenue merely to cover the capital depreciation, operational power expenses, and gross-margin hurdles of the physical infrastructure. At the time, actual generative AI run-rate revenues sat below $20 billion.
By mid-2026, as hardware commitments accelerated rather than decelerated, Cahn updated the calculation to what he termed "AI's $1.5 Trillion Question."
While leading foundation-model developers have achieved impressive nominal revenue growth—OpenAI crossing a $12 billion annualised run-rate, with Anthropic climbing past $4 billion—the entire direct generative AI software market remains under $80 billion globally. The ratio of capital expenditure to direct end-market monetisation currently exceeds 8-to-1.
For equity markets, this asymmetry has created a valuation quandary. During the early phases of the mobile and cloud transitions, infrastructure buildouts were closely tethered to immediate, high-margin consumer and business demand: consumers bought smartphones and immediately subscribed to cellular data plans; enterprises migrated legacy databases to Amazon S3 and Azure, immediately cutting internal server operating expenses.
In generative AI, by contrast, the vast majority of capital expenditure is speculative, deployed in anticipation of fully autonomous "agentic workflows" and transformative enterprise productivity dividends that remain largely unproven at scale.
The Accounting Trap: The Approaching "Depreciation Wall"
While financial media focuses on headline cash spending, equity analysts are increasingly preoccupied with corporate accounting mechanics—specifically, the transition from capital expenditure to depreciation expense.
When a hyperscaler purchases a $40,000 AI accelerator or constructs a $2 billion data centre, the full financial impact does not hit the income statement immediately. Instead, the purchase is recorded on the cash-flow statement as capital expenditure and added to the balance sheet as property, plant, and equipment (PP&E). The cost only flows into the profit-and-loss (P&L) statement over time through depreciation.
Historically, cloud computing providers depreciated general-purpose servers over four to six years, while concrete data centre buildings were depreciated over twenty years. However, AI accelerators present a fundamentally different economic reality:
Severe Technological Obsolescence: Leading merchant chipmakers, led by Nvidia's accelerated annual roadmap, introduce architectural step-changes every twelve months. An accelerator purchased in 2024 offers an order of magnitude lower performance-per-watt than silicon commissioned in 2026.
Physical Degradation: Modern accelerators run continuously at extreme thermal and electrical thresholds, with packaging densities exceeding 3,000 watts per package in high-density rack architectures. Thermal cycling and silicon electromigration limit the viable operating life of these chips.
The D&A Lag: As the record-shattering capex of 2024, 2025, and 2026 is fully energised and brought online, the annual depreciation expense must inevitably surge.
As noted by the Goldman Sachs Global Institute, altering the assumed economic useful life of AI silicon from five years to three years increases annual compute depreciation across the ecosystem by hundreds of billions of dollars. If an operator carries tens of billions of dollars in hardware that has become economically uncompetitive to operate before its book depreciation is complete, the enterprise faces an unenviable choice: either run obsolete, power-hungry chips at uncompetitive inference pricing, or write down the value of the assets through sudden impairment charges.
For companies trading at elevated price-to-earnings (P/E) multiples based on expected margin resilience, this "depreciation wall" poses a direct threat to operating income and free-cash-flow yields over the next twenty-four to thirty-six months.
The Physical Collision: Megawatts, Transformers, and Nuclear Restarts
If accounting schedules represent the financial ceiling of the AI boom, electrical grids represent its hard physical floor. The narrative of AI as a frictionless, ethereal software revolution has collided with industrial reality: modern artificial intelligence is an energy-intensive industrial process.
Whereas traditional cloud data centres operated at power densities of 10 to 15 kilowatts per rack, advanced liquid-cooled AI clusters require 100 to 150 kilowatts per rack. Constructing these facilities has driven average capital costs from roughly $10 million per megawatt (MW) for traditional facilities to between $15 million and $20 million per MW today.
In the primary data centre corridors of the world—such as Northern Virginia, Central Ohio, Silicon Valley, and Dublin—regional transmission grids are effectively exhausted. Regional transmission organisations report interconnection queues stretching between four and seven years, exacerbated by severe global shortages of specialized high-voltage step-up transformers and switchgear, where procurement lead times now routinely exceed four years.
This infrastructure bottleneck has forced hyperscalers into unprecedented utility interventions:
In September 2024, Microsoft signed a historic 20-year Power Purchase Agreement with Constellation Energy to fund the restoration and restart of the 835-megawatt Unit 1 reactor at the Three Mile Island nuclear generating station in Pennsylvania (rebranded as the Crane Clean Energy Center), as reported by The Guardian.
In October 2024, Google executed a corporate master plant agreement with Kairos Power to purchase 500 MW of clean electricity across a fleet of seven small modular reactors (SMRs), with the first units targeted for commissioning by 2030.
Amazon Web Services (AWS) paid $650 million to acquire a 960-megawatt data centre campus directly interconnected with Talen Energy’s Susquehanna nuclear power station, sparking contentious regulatory battles before the US Federal Energy Regulatory Commission (FERC) over whether collocated data centres unfairly shift transmission costs onto public ratepayers.
These physical interventions underscore the structural elongation of the buildout. When billions of dollars of compute capital are committed years before the power plants and transmission lines required to energise them can be built, the cost of carrying unproductive capital rises sharply, eroding project-level internal rates of return (IRR).
The Demand-Side Logjam: Enterprise "Pilot Purgatory"
The ultimate justification for premium valuations is enterprise software transformation: the premise that generative AI will replace routine knowledge work, automate coding, streamline legal discovery, and reinvent enterprise resource planning (ERP).
Yet across corporate IT departments, the reality of adoption has been far slower and more fraught than early prognostications suggested. This phenomenon, widely documented among enterprise software buyers, is known as "pilot purgatory."
While over 80% of Global 2000 enterprises have deployed generative AI proofs-of-concept, fewer than 25% have deployed high-impact applications into broad enterprise-wide production.
The friction is fundamentally architectural: large language models (LLMs) are probabilistic engines deployed into enterprise environments that require 100% deterministic precision. A customer service bot that hallucinates inaccurate refund policies, or a code-generation assistant that introduces subtle cryptographic vulnerabilities into proprietary software repositories, incurs direct operational liabilities.
Moreover, chief information officers (CIOs) are balking at the compounding cost of inference. Unlike traditional software-as-a-service (SaaS), where the marginal cost of serving an additional user or query approaches zero, every generative AI query consumes compute cycles and electrical power. As enterprise pilots scale, software providers attempting to pass through token costs are meeting severe procurement resistance.
Rather than unseating incumbent enterprise software monopolies, generative AI has largely been incorporated as a margin-diluting feature within existing platforms (such as Microsoft Copilot, Salesforce Agentforce, or Google Workspace), where customer willingness to pay significant monthly premiums remains uneven.
Circular Financing and Synthetic Leverage
Adding to institutional investor disquiet is the degree of synthetic leverage and circularity sustaining parts of the AI ecosystem.
A substantial portion of the cash flows celebrated by frontier model developers originates not from unaffiliated enterprise software buyers, but from the hyperscalers themselves. Over the past three years:
Microsoft invested billions of dollars in OpenAI, with a decisive portion of that capital returning immediately to Microsoft in the form of Azure cloud hosting fees.
Amazon and Alphabet invested billions in Anthropic, structured with contractual commitments that require Anthropic to purchase equivalent compute capacity across AWS and Google Cloud.
Semiconductor leaders have participated in funding rounds for specialised cloud startups ("neo-clouds") such as CoreWeave, which subsequently leverage their allocated GPU inventories as collateral to secure multi-billion-dollar debt facilities from private credit lenders—debt that is then used to purchase additional merchant chips.
This closed financial loop creates an optical illusion of self-sustaining top-line growth. In a rising market, circular capital allocation appears synergistic; in a plateauing or decelerating market, it concentrates counterparty risk. If end-market consumer and enterprise demand fails to expand rapidly enough to absorb this capacity, the synthetic leverage unwinds quickly, exposing cloud providers to vacant data centres and leaving credit funds holding rapidly depreciating silicon collateral.
The Valuation Reckoning: The Shift to Fundamental Discipline
The cumulative result of these headwinds is an ongoing reassessment of equity multiples across the technology landscape.
The market is executing a necessary transition:
De-coupling of the "Pick-and-Shovel" Providers from Cloud Off-Takers: Upstream hardware providers with pricing power and near-monopoly margins continue to capture immediate cash flows, while downstream hyperscalers are penalised for soaring capital intensity that compresses free-cash-flow conversion ratios.
The Scrutiny of Return on Invested Capital (ROIC): Investors are moving past simple revenue growth metrics to evaluate ROIC. If a cloud titan increases its invested capital base by $150 billion in a single year, that capital must generate an incremental $15 billion to $20 billion in net operating profit after tax (NOPAT) merely to cover its cost of capital.
The Rise of Alternative, Efficient Architectures: The market is closely watching architectural transitions. As documented by the Goldman Sachs Global Institute, if compute demand becomes more cost-sensitive and shifts from expensive general-purpose GPUs to purpose-built application-specific integrated circuits (ASICs) or smaller, open-weight distilled models, aggregate capital expenditure may contract, compressing the margins of upstream semiconductor suppliers.
This does not signal that artificial intelligence is a mirage. Technological history confirms that every profound general-purpose technology—from the transatlantic telegraph cables of the 1860s and the transcontinental railroads of the 1870s to the fibre-optic telecommunications buildout of the late 1990s—undergoes an identical structural cycle. An initial phase of genuine scientific discovery sparks speculative frenzy, leading to massive over-building and capital misallocation. The subsequent market correction does not kill the technology; rather, it clears away financial excess, forces operational discipline, and establishes the low-cost infrastructure upon which durable, enduring industries are ultimately constructed.
The global AI buildout is not collapsing, but its easy chapter is over. As physical limits bind, depreciation mounts, and corporate treasuries demand measurable returns, the industry has transitioned from the realm of boundless narrative into the unforgiving discipline of basic corporate accounting.
