Generative AI is turning the cloud into a 'new heavy industry'

Generative AI has pushed the cloud's capital expenditure (CapEx) to a level akin to 'infrastructure.' The reason isn't just buying more GPUs, but an entire data center stack being forced to be rewritten: computing power (GPUs/accelerators), memory (HBM), networking (high-speed interconnects), power supply (transformers and redundancy), cooling (liquid cooling), and physical construction all expanding simultaneously. The guidance and market tracking for 2026 have already written this 'heavy industrialization' in numbers: Alphabet indicated its CapEx may double to a range of $175-$185 billion by 2026; Amazon is also reported to possibly increase its CapEx to $200 billion by 2026, a significant jump from $131 billion in 2025; Meta provided a CapEx range of $115-$135 billion in its 2026 guidance.

Behind these figures lies the reflection that cloud giants are rushing ahead on the 'supply side': whoever secures data center space and power capacity first will have a better chance of leveraging computational supply capabilities to gain pricing power in the coming years. As a result, the investment market has started using the 'cloud CapEx cycle' to understand macro fund flows. UBS’s view, cited by Reuters, even revised the potential CapEx scale of hyperscale cloud operators upwards to approximately $770 billion, linking to rising bond financing needs, indicating that this round of AI investment is no longer simply tech spending, but a reallocation at the capital market level.

Depreciation Period and 'Compute Utilization'

Since CapEx resembles infrastructure, returns cannot be viewed with a 'quarterly payback' mindset. The ROI of cloud AI infrastructure usually hinges on core variables such as depreciation period × utilization rate × monetization capability per unit of compute. Accounting-wise, the depreciation periods of servers and network equipment naturally impact return timing; in recent years, cloud giants have been adjusting this 'payback timeline.' Industry research indicates that hyperscale cloud companies modify the useful life of servers and network assets—for instance, Meta mentioned in its annual report that some servers and network assets have a useful life of 5.5 years, while Amazon adjusted some server lifespans to 5 years (shorter than the previous year). The implication of these adjustments is straightforward: faster equipment iteration, quicker technological obsolescence, and greater reliance on high utilization rates and rapid commercial scaling for returns.

If we formulate ROI into a practical model, it closely resembles a data center version of the 'single-store model': every $1 of CapEx must be recovered within the depreciation period through (1) rentable hours of GPU/servers (utilization rate), (2) effective revenue per GPU hour (price × actual usage volume), and (3) gross profit after deducting power, maintenance, network, and depreciation costs. When utilization rises from 30% to 60%, the payback speed doesn’t increase linearly but exhibits a leverage effect after crossing the breakeven point; conversely, if demand slopes decline or price competition intensifies, cash flow recovery will extend, prompting markets to reassess the rationality of CapEx.

Applying this logic to enterprises adopting generative AI for ROI expectations reveals a common phenomenon of 'extended cycles.' Deloitte's 2025 survey indicates that most respondents believe typical AI use cases often require 2–4 years to achieve satisfactory ROI, with only 6% of respondents stating they could break even within a year. This aligns with the return rhythm of cloud infrastructure: transitioning from investment to generating sustainable paying demand typically requires a 'penetration period' involving productization, workflow transformation, data governance, and internal adoption—not immediately reflecting in revenue and cash flow upon launch.

From 'Selling Computing Power' to 'Selling Outcomes'

The explosion of CapEx by cloud giants is, in essence, a bet on 'investing in supply first, then finding demand,' but it's also rewriting how the cloud generates revenue. The first phase is the most straightforward: renting out GPU instances and selling inference and training usage; the second phase is more critical: embedding generative AI into high-retention services like databases, analytics, collaboration, and security. This makes AI not just a new revenue stream but an 'accelerator' driving overall cloud service ARPU (Average Revenue Per User) and retention. This is why management increasingly emphasizes 'capacity constraints' and 'demand visibility' when discussing CapEx — because what truly shortens the payback period isn't buying more machines, but turning those machines into sustainable, recurring payment behaviors.

However, the market has become more sensitive in questioning returns because this round of AI infrastructure expansion faces two external constraints: power and capital. On the power side, the electricity demands of AI data centers are pulling grid and utility investments into the same chain, sparking macro-level discussions about inflation and infrastructure burdens. On the financial side, as CapEx scales approach hundreds of billions of dollars, the tension between external financing, free cash flow, and shareholder returns becomes more pronounced. Any 'late arrival of demand' could translate into stock price volatility and valuation markdowns.

Therefore, when assessing the CapEx cycle brought by generative AI, the key is not guessing which company is spending the most, but tracking three signals that can validate returns in advance: whether computing power utilization continues to rise, whether AI-related revenues can transition from one-time pilots to predictable renewals and usage growth, and whether, under the pressures of power and depreciation, cloud businesses can still maintain operating leverage. When all three conditions hold, the market will view this round of spending as 'front-loaded investment'; if any link breaks, CapEx may be redefined as 'overbuilding.' The return on hardware investments for generative AI remains a marathon measured in years — it’s not that they won’t pay off, but the speed of payback will determine how the next round of capital markets prices the AI narrative.

(Chip and Computing Power Series #33)

$Meta Platforms (META.US)$ $NVIDIA (NVDA.US)$ $Advanced Micro Devices (AMD.US)$ $Microsoft (MSFT.US)$ $Tesla (TSLA.US)$ $Amazon (AMZN.US)$

113K Views