The explosion of hardware capital expenditure brought by generative AI

In the past, when discussing cloud capital expenditure (CapEx), the market mostly interpreted it as routine capacity expansion of data centers; after the advent of generative AI, the tone of CapEx has changed—it has become more like an arms race for computing power and electricity. Bridgewater estimates that giants like Alphabet, Amazon, Meta, and Microsoft may collectively invest about $650 billion in AI infrastructure by 2026, up from approximately $410 billion in 2025. ([Reuters][1]) This is not at the level of 'buying a few more servers,' but rather pushing companies from being light-asset cloud service providers to becoming more infrastructure-heavy enterprises: higher depreciation, longer payback periods, more noticeable free cash flow pressure, and more sensitive investor accountability.

Similar trends can be observed at the individual company level. Reuters reported that Alphabet's guidance for capital expenditure in 2026 is between $175 billion and $185 billion, and the market’s initial reaction was not applause but concern over 'spending too fast, returning too slowly.' ([Reuters][2]) As for Amazon, Reuters mentioned that its capital expenditure budget for 2026 could reach $200 billion, significantly higher than $131 billion in 2025. ([Reuters][3]) When these figures are plotted on the same graph, you’ll understand why investors are saying 'AI is the future' while also starting to get nervous: when CapEx becomes a routinely massive expense, it no longer plays a supporting role in the growth story but takes center stage in valuation.

ROI Model: The question is not whether returns will come, but how long they will take and in what form

The return cycle of hardware investments determines whether the narrative of cloud-based AI can endure through economic cycles. If we break ROI down into the simplest model: (revenue per unit of computing power × utilization rate × gross margin), whether it can cover (CapEx + electricity and operation Opex + financing costs) within the depreciation period is key. The problem is that the revenue side of generative AI is often 'small at first, then large,' while the cost side is 'large at first, then even larger': the initial phase requires setting up training clusters and high-speed networks before customers’ paying behavior gradually forms a stable curve.

A more realistic constraint lies in depreciation. Reports studying data center project financing indicate that the industry generally considers the economic depreciation of GPUs to be around 3 to 4 years, but different companies may extend accounting depreciation longer (e.g., some depreciate over 6 years, others over 4 years). ([Center for Public Enterprise][4]) This means the 'payback window' is effectively locked by the speed of hardware updates: if utilization cannot be ramped up within 3–4 years, if the price per unit of computing power is suppressed by competition, or if electricity and cooling costs exceed expectations, the asset may be rendered obsolete by technological iteration before it is recovered on the books. Hence, you see cloud providers simultaneously doing two seemingly contradictory things: accelerating investment on one hand, while urgently steering demand toward 'normalizing inference (inference)' and 'long-term enterprise contracts,' because only stable, predictable inference usage can lock in utilization rates and cash flow curves.

This 'invest first, recover later' characteristic is also reflected on a macro level. Bridgewater estimated in its research that AI CapEx could contribute about 140 basis points (1.4 percentage points) to US GDP growth in 2026 and approximately 150 basis points in 2027. ([Bridgewater][5]) This shows, on one hand, the scale of investment is large enough to impact overall data; on the other hand, it serves as a reminder: GDP being lifted by investment does not mean shareholder returns will immediately follow, as investment returns must pass through four hurdles—commercialization, pricing, competition, and depreciation.

The biggest risk in the return cycle isn’t 'lack of demand,' but rather 'profits being eroded.'

What the market fears most is not that there’s no demand for computing power—demand remains strong at present; what it fears most is that despite strong demand, returns are being fragmented. As the four major cloud providers and more 'computing power intermediaries' expand simultaneously, the unit price of computing power will eventually be driven down by competition, while bottlenecks in the supply of electricity, land, equipment, and talent continue to prop up costs. Thus, the focus of the 'return cycle' shifts to whether cloud providers can use differentiated capabilities to protect their margins—such as proprietary chips, network stacks, hardware-software integration, and enterprise-grade security compliance; whether model companies can use productization and tiered pricing to convert inference demand into sustainable revenue; and whether corporate clients can truly embed generative AI into their workflows, turning payments into fixed budgets rather than trial fees.

This is why the market has recently been bullish on hardware and computing power chains while becoming increasingly sensitive to 'excessive investment in cloud infrastructure.' The surge in CapEx itself is neither a positive nor negative signal—it merely brings the issue to the forefront: the returns from generative AI are not nonexistent, they just take time. And time happens to be the most expensive cost in the capital markets. To ensure this round of investment outlasts the depreciation cycle, the competition among cloud providers will no longer be about who spends the most but who can turn 'computing power' into 'billable, renewable, and explainable business' the fastest.

(Chip and Computing Power Series #35)

$MONTAGE TECH (06809.HK)$ $GIGADEVICE (03986.HK)$ $SMIC (00981.HK)$ $ASMPT (00522.HK)$ $NVIDIA (NVDA.US)$ $Broadcom (AVGO.US)$

41K Views