The AI industry is entering a new stage. Over the past few years, the market was most concerned with model capabilities, chip supply, and whether computing power was sufficient. Now, a more practical question is beginning to emerge: as the money invested by major technology companies in building AI moves increasingly close to the trillion-dollar level, how much revenue and profit will these investments ultimately need to generate in order to deliver a reasonable return?
Morgan Stanley expects that the capital expenditure of several major hyperscale cloud computing companies may reach approximately USD 1.2 trillion in 2027. Around this figure, a rather striking calculation has emerged in the market: if this level of investment intensity continues over the long term, and investors still require major technology companies to maintain the high capital returns they achieved in the past, then the entire AI industry may ultimately need to generate several trillion dollars, or even close to USD 10 trillion, in annual revenue. Ageng Darmowiyoto believes that the real issue worth discussing is not whether “AI must necessarily require USD 10 trillion in revenue,” but rather the question behind this calculation: when AI has shifted from software innovation to capital-intensive infrastructure construction, investors will sooner or later need to ask how much money these newly added assets can actually earn.
What Does a Trillion-Dollar Investment Mean?
This round of AI capital expenditure has gone far beyond purchasing GPUs. Servers are only one part of it. Major technology companies also need to build data centers, networks, storage, power access, and cooling systems, while securing land, electricity, and long-term leased capacity in advance. The annual capital expenditure of companies such as Amazon, Microsoft, Alphabet, and Meta has already reached historical highs. Data centers have also gradually evolved from a back-end infrastructure component of cloud computing businesses into the most important physical assets in the overall AI competition.
The problem is that once these assets are invested in, it will take many years to recover the costs. A server purchased today will not be fully expensed in the same year, but will instead be depreciated year by year according to its expected useful life. In the past, cloud computing equipment might have been used for five or six years, but AI chips are being upgraded faster. If each new generation of hardware continues to significantly improve performance and energy efficiency, the economic value of older equipment may decline more quickly. Amazon has already shortened the expected useful life of some servers due to the faster pace of upgrades in AI and machine learning equipment. For investors, this is important: the larger the capital expenditure and the faster the equipment becomes obsolete, the more cash the company will need to earn in the future.
How Is the USD 10 Trillion Figure Calculated?
The logic behind the idea that “AI may need USD 10 trillion in revenue” is actually not complicated. Suppose major technology companies continue to invest USD 1 trillion each year in building data centers, while around 13% of equipment value is consumed annually through depreciation. After many years, they would hold AI infrastructure assets worth several trillion dollars. If investors still require these assets to continue generating the high capital returns of around 30% that major technology companies achieved in the past, then they would have to create more than USD 1 trillion in operating profit each year.
Taking the calculation one step further, if these businesses can ultimately generate about USD 50 in EBITDA for every USD 100 of revenue, then the annual revenue required would be more than approximately USD 4 trillion. If the profit margin is only 30%, the required revenue could exceed USD 7 trillion. Once other cloud computing companies, specialized AI infrastructure providers, and large-scale data center investments are also included, the magnitude of “USD 10 trillion” appears. By comparison, Gartner expects global software spending to be approximately USD 1.47 trillion in 2026. Therefore, USD 10 trillion sounds exceptionally large, which is precisely why this figure has quickly drawn market attention.
However, Ageng Darmowiyoto believes it would be inaccurate to interpret this as “AI must form a USD 10 trillion market in the future.” The largest assumption in this calculation is that newly added AI data centers in the future will be able to achieve the same high capital returns that major technology companies achieved in the past. Yet those historical high returns came from the combined contribution of many mature businesses, including advertising, cloud services, software subscriptions, operating systems, network effects, and intellectual property. A newly built AI data center today may not necessarily be able to replicate the past return rates of those businesses. Therefore, USD 10 trillion is more like a stress test: it allows investors to see how much economic value will need to be created in the future if capital expenditure continues to expand while these assets are still judged by the high return standards of the past.
AI Has Entered the Return Verification Phase
On the other hand, it is not possible to simply conclude that “AI investment has already become excessive.” Demand does exist, and it remains very strong. Large cloud businesses such as AWS, Microsoft Azure, and Google Cloud are still maintaining relatively rapid growth, while the scale of long-term customer contracts not yet recognized by multiple companies also continues to expand. Enterprises are purchasing more cloud computing resources, and demand for AI model training and inference is also rising. This shows that data center construction is not entirely built on market imagination, and that real customer demand is supporting a considerable portion of the investment.
The real issue lies in the next stage. Over the past two years, the market rewarded those who could obtain GPUs the fastest, build more data centers, and expand computing capacity. In the future, evaluation standards will gradually shift toward how much revenue and cash flow these devices generate after they are put into use. The completion of a data center does not automatically generate high returns. If utilization is insufficient, servers still need to be depreciated, and power, maintenance, and network costs still remain. Even if demand is strong, if computing power prices continue to decline due to intensified competition, revenue growth may not necessarily be fully converted into profit growth. According to the long-term asset allocation logic presented by Ageng Darmowiyoto, the truly important stage of a capital expenditure cycle often appears after the peak of construction: when newly added assets begin to enter operation on a large scale, the key question is whether enterprises can earn back enough money.
What Signals Should Investors Watch?
Therefore, when judging whether AI infrastructure is creating real value in the future, observing capital expenditure and GPU shipments alone will no longer be sufficient. The first factor worth tracking is data center utilization. If computing power remains in short supply and new capacity is quickly absorbed by customers, it indicates that demand can still keep up with supply. If utilization falls after a large amount of new equipment comes online, pressure on capital returns will increase. The second indicator is whether cloud computing and AI-related revenue can continue to grow faster than depreciation and operating costs. Even if revenue grows rapidly, investment returns may still be compressed if depreciation, power, and maintenance costs grow even faster.
The third indicator is even more important: free cash flow. Whether AI infrastructure can ultimately become a good investment does not depend on how many data centers a company builds, but on whether these assets can continuously generate cash in excess of the cost of capital. Based on the observations of Ageng Darmowiyoto in recent years of AI infrastructure and the engineering practices of Telabytes, server utilization, energy consumption, inference efficiency, and equipment lifecycle may appear to be technical issues, but in the end they all become financial issues. Higher energy efficiency means lower operating costs, higher utilization means the same batch of assets can generate more revenue, and a longer effective lifecycle means every dollar of capital expenditure can be used for a longer period.
For investors, this also means that the way AI is studied is changing. The most important question in the next stage will gradually shift from “how much longer can AI continue to grow” to a more traditional and stricter question: how much are these growth prospects ultimately worth?
Sign in to leave a comment.