We must approach the current expansion in artificial intelligence infrastructure not as a sudden rupture, but as the latest stage in a gradual evolution of industrial organisation—one that, like all significant economic adjustments, operates with distinct time horizons, production frictions, and substitution possibilities. The build-out now underway is among the largest capital allocations in modern economic history, and Amazon Web Services (AWS), as a hyperscale cloud provider, is situated at its centre. To grasp the implications for Amazon.com Inc. (AMZN), it is essential to distinguish between the temporary bottlenecks of the present conjuncture and the structural forces that will shape longer-run competitive equilibria.
The Changing Character of Demand: From Training to Inference and Agentic Workloads
The nature of demand for AI compute is itself evolving. We observe a decisive shift from a training-led regime to one dominated by inference, a transition that analysts at Stifel expect to be well advanced by the end of 2026 22. This is not simply a matter of shifting proportions; agentic AI systems—those that execute tasks, run code, and orchestrate multi-step operations—are broadening the spectrum of required compute, drawing in CPU resources alongside the GPU accelerators that captured earlier investment 31,41. That the majority of current AI capital expenditure now flows toward serving infrastructure, such as agents and chatbots, rather than toward training alone 19 suggests a maturing of demand that, once hardware is deployed, lends durability to revenue streams. For AWS, which offers both custom Trainium accelerators and a diverse portfolio of compute instances, this broadening is particularly instructive: it implies that the addressable market is expanding across multiple workload categories, not merely intensifying along a single dimension 41.
The Architecture of Supply Constraints
The supply side, however, presents a more complex picture of adjustment and delay. At present, demands for AI infrastructure comfortably exceed available capacity across every link in the supply chain—market demand for AI currently exceeds the supply of AI infrastructure and services 2,11 (2 sources), and global demand for high-performance computing hardware surpasses available supply 58. The bottlenecks extend well beyond the familiar shortage of leading-edge GPUs; they encompass power, transformer capacity, networking, cooling, land, labour, advanced packaging, and the usable data-centre space that can be brought online within relevant time frames 32,49. These frictions are not temporary irritants but rather structural features of an industry that adjusts only gradually. Persistent supply chain constraints for AI servers are projected to create delivery bottlenecks throughout 2027 44, and the most carefully constructed forecasts indicate that global AI compute capacity constraints will persist through at least that year 28. Such persistent excess demand naturally confers pricing power on those who can supply it, and we see evidence of this in the upward movement of AI compute service pricing—reportedly rising approximately 20% in July 2026, following a 15% increase in January 24.
A particular feature of the current supply architecture merits separate examination: memory has emerged as the critical chokepoint that physically limits the number of data centres that can be constructed globally 24,36. High Bandwidth Memory (HBM), DRAM, and NAND-based SSDs are primary beneficiaries of the AI super cycle 16,18,45,56, yet it is precisely their scarcity that is driving significant cost spikes across RAM and storage 8,20,26. The combination of tight memory chip supply with strong GPU demand is thus inflating operational costs for cloud providers 24, and the memory upcycle is being extended by these very pressures 15. Quantitatively, the magnitude of the shift is striking: compute tile demand has increased twentyfold while memory capacity per compute tile has increased twelvefold 20. For AWS, rising memory costs represent a margin headwind on commodity workloads, but the pricing power observed in the market provides a path to pass through these increases—provided the elasticity of substitution among cloud providers remains sufficiently low.
Capital Intensity, Depreciation, and the Return on Investment
These margins, however, must be set against the sheer capital intensity of the current cycle. Concerns about an AI infrastructure bubble are grounded in the observation that specialized AI chips depreciate rapidly 42. The replacement cycle for AI data centre hardware is estimated at less than three years 37—a striking contrast with the multi-decade lifespans of the underlying power plants and substations 21. The cash-flow implications are tangible: hyperscale cloud providers are experiencing profit growth that does not keep pace with their capital expenditure growth 14, and major AI companies continue to receive approval for new data centre construction even as they report multi-billion-dollar losses 58. Sceptics rightly warn that hundreds of billions invested in data centres may fail to generate adequate returns 45, and enterprise spending on AI itself faces a risk of contraction once initial business model shifts are complete 58. A particularly revealing contrarian signal is the projection by David Linthicum of Deloitte that actual built capacity will reach only about half of what construction-start data currently suggest 21. This implies that not all planned investment should be taken at face value; the market may impose its own rationing.
Energy, Water, and the Natural Limits to Growth
Beyond the financial calculus lie physical constraints that no pricing mechanism can bypass. AI data centres consume vast amounts of electricity and water for cooling and operations 3,4,23,25,46 (6 sources), and data centre power requirements have increasingly outpaced the deployment of renewable energy initiatives 50,54. Emissions tied to data centre construction are themselves rising—Amazon specifically reported a more than 40% year-over-year increase in emissions tied to data centre construction 52 (2 sources)—and the energy intensity of the supply chain for chips and servers further compounds the carbon footprint 52. Grid decarbonisation is failing to keep pace with the AI build-out 50, and some AI facilities may require dedicated nuclear reactor power sources to meet their requirements 57. The cost of utilities rises in consequence and is passed through to consumers 46, while regulatory pressure 35,57 and reputational risks mount as the AI race drives carbon emissions higher for major cloud providers 52. For AMZN, sustainability is not a peripheral concern but a binding constraint on the long-run scalability of the infrastructure thesis.
Geographic Diversification and the Reconfiguration of Supply
One important equilibrating mechanism is the geographic redistribution of capacity. India’s data centre capacity, for example, grew from approximately 350 MW in 2019 to roughly 1.6 GW by 2025, representing a compound annual growth rate of 29% against a global average of 20% 39,43. The country is projected to become a significant hub for AI computing infrastructure 33, and global hyperscalers—including AWS—are expanding accordingly 33,43. This diversification carries implications not only for cost and latency but also for competitive positioning against domestic providers and for navigating geopolitical pressures toward data localisation.
Those geopolitical pressures are themselves shaping market structure. Enterprise customers increasingly require AI data centres to be located within the United States due to data security and compliance concerns 5, and governments are mandating that sensitive training data remain within domestic borders 1. Foreign customers may shift AI consumption to domestic alternatives in China and the European Union 38, and nations are investing in nationalised GPU clusters to achieve compute independence 1. For AMZN, these trends reinforce the value of its existing global infrastructure footprint while simultaneously introducing market access risks that vary by jurisdiction.
Data Gravity and the Evolution of Architecture
A complementary insight concerns data gravity—the principle that compute resources must be placed near expanding datasets 53. This dynamic favours distributed architectures and edge deployments, creating opportunities for AWS’s regional availability zones and Outposts offerings. As AI clusters scale from single racks to multi-rack deployments, bottlenecks are shifting from raw compute capacity to the efficiency of data movement between GPUs, switches, and racks 6, increasing the demand for networking bandwidth and optical interconnects 6—areas where AMZN’s investments in custom silicon and networking provide a differential advantage.
Revenue Signals and the Test of Market Validation
Despite the cautionary signals on the cost side, concrete revenue indicators do offer a degree of validation. Google Cloud Platform reported 800% year-over-year growth in enterprise AI revenue 19 (6 sources), and the platform has recorded revenue growth driven by enterprise AI adoption 19. The broader AI and machine learning industry is projected to reach $350 billion in annual revenue by 2030 10, and global AI infrastructure investment is expected to reach trillions of dollars 12,17. Hyperscalers are expected to host the majority of large language models 34, creating a dependency that is structurally favourable for platforms like AWS.
At the same time, we must consider counter-narratives that could alter the long-run equilibrium. Floating data centres are emerging as a response to land and power constraints 7; SpaceX has begun leasing out excess data centre capacity while using some for internal training 40; and Meta is planning to monetise its own surplus AI compute 30. The possibility that local LLM development on laptops could reduce data centre demand between 2027 and 2030 13, and that open-source AI model proliferation is driving commoditisation that may erode competitive moats 19,29, introduces an element of uncertainty that tempers the most bullish projections.
Strategic Implications for Amazon.com Inc.
For AMZN, the evidence reveals a dual-edged narrative. On one side, AWS is positioned at the centre of a multi-decade structural investment cycle in electrical equipment, data centre infrastructure, and semiconductors 47, with sustained demand drivers that include chip replacement cycles 21, the training-to-inference transition 22, agentic compute requirements 31,41, and the 100 GW of planned global data centre capacity additions between 2026 and 2030—representing a doubling of current capacity 48. The current supply-demand imbalance provides pricing power, as demonstrated by the 20% July compute price increase 24, and the company’s scale, custom silicon, and geographic footprint create barriers to new entry.
On the other side, the same dynamics impose costs that must be managed. Rising memory and component costs compress margins 8,51; capital expenditure growth outpaces profit growth 14; hardware depreciation risks 21,42 raise the possibility that a material fraction of investment may not be recovered; and estimates that only half of planned capacity will materialise 21 call into question the sustainability of current spending levels. The company’s rising emissions profile 52,54 adds regulatory and reputational exposure, while geopolitical fragmentation 1,5,9,38 could constrain cross-border data flows and segment markets.
The interesting question is not whether AWS will generate significant revenue growth—that seems probable under prevailing conditions—but how these contending forces will shape the margin profile and the return on invested capital. The transition from asset-light to higher-capex business models represents a fundamental shift in the tech sector’s return characteristics 27, and investors in AMZN must underwrite this shift with a clear-eyed view of the time horizons over which returns are earned and the adjustment lags that separate expenditure from revenue. Under current conditions, the evidence suggests that AWS’s near-term trajectory remains favourable: constrained supply, rising prices, and broadening workload diversity support revenue growth through at least 2027–2029, when major build-outs are expected to deliver 22,31,55. However, the margin sustainability depends critically on the company’s ability to pass through component cost increases to enterprise customers and to monetise infrastructure before technological obsolescence erodes its value. The path is not predetermined; it will turn on the elasticities of demand, the pace of competitive entry, and the extent to which natural resource limits—energy, water, and the decarbonisation gap 46,50—impose binding constraints. These are precisely the kinds of organic adjustments that, in the Marshallian tradition, we would expect to observe and analyse as the industrial organism matures.