Distinguishing Revenue, Speed, and Risk Effects Often Missed When Evaluating AI Investments Solely on Cost Reduction: Designing Standards for AI CAPEX Investment Judgments Based on the Alibaba Case

When Alibaba released its earnings for the quarter ending June on August 20, 2026, completely contrasting headlines poured out based on the exact same data. One side read, "Net Profit Plummets 75%," while the other declared, "Cloud Revenue Grows 45%, Highest Growth Rate in 22 Quarters." Both are true. And both came from the exact same cause.

Let us look at the numbers. Quarterly revenue came in at 268.953 billion yuan, up 9% compared to the same period last year. On the other hand, net profit plunged 75% to 10.444 billion yuan, and the operating margin was cut in half from 14% to 6%. However, explaining this 75% decline solely through AI investment is inaccurate. According to the company's announcement, along with the decrease in operating income, a decline in investment asset disposal gains, a decrease in equity investment evaluation gains, a 550 million euro provision related to the European Union Digital Services Act, and a 4.458 billion yuan goodwill impairment acted together. The portion purely explained by the expansion of technology and infrastructure investment is the pressure on operating profit, while the rest consists of one-off and valuation factors of a different nature.

Meanwhile, capital expenditures rose 75% to 67.678 billion yuan from 38.676 billion yuan a year earlier. Cash generated from operating activities actually increased 11% year-on-year to 22.945 billion yuan. The problem lies next. With CAPEX significantly exceeding operating cash inflows, free cash flow recorded a net outflow of 44.67 billion yuan. This is more than double the outflow of 18.815 billion yuan recorded in the same period last year. To sum it up in one sentence: Alibaba is generating cash from operations, but its pace of spending on AI infrastructure is much faster. And management says that is according to plan.

This case is meaningful to South Korean business executives not because of the scale of investment. With quarterly CAPEX alone reaching roughly 14 trillion won in Korean currency, the scale itself is difficult to use as a direct reference. What is truly worth noting is the decision-making structure. Faced with the situation where "money was spent on AI and profits decreased," will management read this as a failure or as an ongoing investment? What will they base that judgment on? Alibaba's earnings announcement is a rare case that publicly exposes that entire decision-making structure to the outside world.

And this question is already sitting on the desks of most companies. McKinsey's "State of AI" survey, released on August 25, 2026, covered 1,719 respondents across 97 countries worldwide, revealing that the proportion of respondents who answered that AI contributed even slightly to company-wide operating profit remained at 37%, virtually unchanged from a year ago. Conversely, the proportion who answered that AI increased individual productivity reached 80%. A state where productivity is clearly felt at the individual level, yet fails to register on the corporate income statement. This is the exact point where most companies currently stand.

Stage 1: Breaking Investments Down into Layers, Not a Single Lump Sum

When discussing the AI investment payback period, the very first point of failure is not the calculation, but the definition. The moment you ask for the payback period regarding a single item called "our company's AI investment," an answer cannot emerge because expenditures of completely different natures are bundled into one basket.

Alibaba exposed this problem itself by reorganizing its business segment structure this quarter. It combined its existing cloud business unit and semiconductor design organization into "AI Cloud & Computing Services," and merged its model research organization and consumer AI app organization into "AI Labs & Applications." And the report cards of the two divisions split completely.

AI Cloud & Computing Services grew revenue by 45% to 48.437 billion yuan, and adjusted EBITA surged 133% to 5.628 billion yuan. Converted into a profit margin, this is 11.6%, up from the 7% level a year ago. In contrast, AI Labs & Applications saw revenue grow by a modest 16% to 3.338 billion yuan, while recording an adjusted EBITA loss of 13.861 billion yuan—more than four times larger than the loss of 3.224 billion yuan a year earlier. The company explained the reasons as the expansion of AI capability investments and rising inference costs for consumer AI apps.

The same company, the same quarter, and the exact same label of "AI investment." Yet one is already generating profits with margins improving nearly twofold, while the other has seen its deficit widen fourfold. If the two had been combined to calculate the "profitability of Alibaba's AI business," neither the recovery signals from the well-performing side nor the warning signals from the cash-leaking side would have been visible.

In actual practice, this layer is generally divided into three. First is the infrastructure layer. These are foundational expenditures that, once laid down, serve multiple purposes, such as servers, GPUs, data pipelines, and internal data refinement. They have long payback periods and are difficult to allocate to individual projects. Second is the application layer. These are expenditures whose effects are captured relatively quickly, such as AI tools attached to specific tasks, customer service automation, and document processing automation. Third is the capability layer. These are expenditures that appear as costs on the accounting books, but whose results manifest in the other two layers, such as workforce redeployment, training, workflow redesign, and governance establishment. If you mix these three and ask, "What is our AI investment ROI?", the person in charge has no choice but to manufacture numbers.

Management Insight

This is akin to putting new factory facility investments and consumable purchases onto a single ledger and asking for the "return on equipment investment." A press machine is used for 10 years while cutting fluid disappears in a month, so what meaning does an average payback period combining the two hold? AI investment is exactly the same. At the next management meeting when reviewing the AI budget, ask the responsible executive to provide three numbers instead of a grand total: how much foundational spending is laid across multiple businesses, how much spending is directly attached to specific tasks, and how much was spent on changing people and processes. An organization that cannot separate these three should be regarded as remaining at the execution stage of managing AI investments, rather than the governance stage. And usually, the moment this question is asked, the person in charge attempts that separation for the very first time.

Stage 2: Decomposing Payback Effects into Cost, Revenue, Speed, and Risk

The second stage involves sorting out the forms in which money returns. Most companies capture the effects of AI investment solely through cost reduction, which heavily distorts the calculation of the payback period.

The McKinsey 2026 survey precisely illustrates this point. The areas where respondents reported the greatest cost reduction effects were supply chain management, service operations, and manufacturing. Conversely, the areas where revenue growth effects were most frequently reported were marketing and sales, followed by product and service development and software engineering. This means the department where costs decrease and the department where revenue increases are entirely different. If a company introduced AI to its marketing organization and only asked, "How much did labor costs decrease?", that investment was likely recorded as having no performance. The account subject where the effect manifests was misidentified from the start.

A study on AI performance released by PwC on April 13, 2026, shows the extent to which this divergence stretches. In this study surveying 1,217 executives of large publicly traded companies across 25 industries, the top 20% of companies captured 74% of the economic value generated by AI. The AI-based financial performance of these top companies reached roughly 7.2 times that of the rest. PwC attributes the cause of this gap not to the volume of tools adopted, but to the difference in aiming points. Leading companies aimed AI at growth and business model redesign rather than labor cost reduction. According to the survey, they were 2.6 times more likely than other companies to answer that AI enhanced their business model reinvention capabilities, and 2 to 3 times more likely to use AI to find new growth opportunities at points where industry boundaries blur. Another striking figure was released alongside by PwC: only one in eight CEOs reported experiencing simultaneous revenue increases and cost reductions.

Two more effects must be added here. The third is the speed effect. This is the value derived from producing identical results more quickly, which is not directly captured on the income statement. This effect is well demonstrated in the Alibaba case. The company stated that through its consumer AI apps, 250 million people experienced AI-based shopping for the first time. This figure is not accounted for as revenue anywhere on the AI division's profit and loss statements. However, it is likely to influence customer acquisition and repeat visit rates in the e-commerce division. Actual contribution is an area that must be verified through separate metrics such as customer acquisition cost, conversion rate, and repurchase rate, and the company did not disclose those figures separately in this announcement. Whether you call it cross-segment synergy or anything else, the structural reality remains unchanged: the place where the investment is made and the place where effects appear (or can appear) are different.

The fourth is the risk effect. This is the value of "things not getting worse," such as error reduction, regulatory compliance capability, and reduced quality variance. This is the most difficult to measure and thus most frequently omitted. However, to be fair, opposing risks must also be captured in this category. This quarter, Alibaba recognized provisions for a 550 million euro fine related to the European Union Digital Services Act alongside a 4.458 billion yuan goodwill impairment. As platform scale expands and the scope of automation broadens, regulatory risk operates not in the direction of shrinking, but expanding.

Management Insight

Looking solely at cost reduction is similar to buying a car and only checking fuel efficiency. Fuel efficiency is easy to calculate and falls neatly into numbers, so people keep looking at it. Yet if the real reason for buying that car was to make three more client visits per day, the fuel efficiency table is of no help to decision-making. When reviewing an AI investment proposal, ask these questions: "To which account are the costs reduced by this investment credited, to which department's performance is the increased revenue attributed, what tasks are accelerated and what more can be done with that speed, and what new risks does this investment create?" It is fine if all four blanks are not filled. The important thing is leaving blank spaces blank. The moment an unverified effect is forcefully shoved into the cost reduction column, that proposal gets passed, and a year later no one pulls those numbers back out.

Stage 3: Agreeing First on the Denominator of the Payback Period, the Useful Life

The payback period is a calculation composed of a numerator and a denominator. Yet in practice, debate concentrates solely on the numerator—how much will be earned—while the denominator is set out of convention. In AI investments, this convention is dangerous.

We can see how this time lag manifests in numbers from Alibaba's earnings. This quarter's CAPEX was 67.678 billion yuan, while depreciation and amortization of property and equipment, impairment, and land use right lease costs reflected in the calculation of adjusted EBITDA amounted to 11.814 billion yuan—up 71% from 6.891 billion yuan a year earlier. However, directly comparing these two figures within the same quarter is not an accurate accounting approach. Depreciation expenses represent numbers expensed by dividing assets purchased over multiple past years, whereas current-quarter CAPEX is a future cost that will be reflected on the income statement over several years ahead. To state it accurately: CAPEX immediately worsens cash flow at the time of expenditure, but is reflected in earnings over the useful life of the asset. The profit decline visible right now is not the full pressure that AI investment will exert on earnings, but merely a portion of it appearing early.

Therefore, over how many years to divide becomes the core variable. This issue became a full-blown debate in global capital markets starting in the latter half of 2025. Investor Michael Burry argued that while major cloud service providers depreciate GPUs over 5 to 6 years, their actual economic lifespan is closer to 2 to 3 years, which would cause the industry's overall depreciation expenses from 2026 to 2028 to be understated by roughly 176 billion dollars. It must be clarified that this is an individual investor's estimate rather than audited official figures. Industry counter-arguments are equally formidable. The logic is that chips phased out from the latest generations continue to generate revenue by shifting to inference or lightweight workloads, so physical lifespan and economic lifespan must not be confused.

What is intriguing is that corporate judgments split under the same technological environment. In early 2025, Amazon shortened the useful life of certain servers from 6 years to 5 years, citing the accelerated pace of technological advancement in the AI and machine learning fields. In contrast, Meta actually lengthened the useful life of most server and network assets during the same period. At a time when NVIDIA releases new architectures practically on an annual basis, it has become evident that useful life is not an objective fact, but a management judgment. Microsoft CEO Satya Nadella's remark that he does not want to lock 4 to 5 years of depreciation onto a single generation of hardware falls along the same vein.

Alibaba's management offered a relatively concrete answer regarding this issue. CEO Eddie Wu stated during the earnings conference call that, based on current average gross margins, AI computing investments will reach a break-even point within three years, which could be pulled forward to two years if profit margins improve. He added the explanation that increasing the deployment of self-designed chips to replace externally purchased chips will significantly improve gross margins and profitability.

What matters in this statement is not the number three years, but the fact that conditions were specified. With the premise of "based on current average gross margins" attached, a drop in profit margins will extend the payback period. It was stated in a form that allows external verification. Most corporate AI investment proposals lack such premises. They contain only the conclusion of a 3-year payback period, without recording what assumptions that 3-year span rests upon.

Management Insight

From the perspective of a mid-sized enterprise executive, the GPU depreciation debate sounds like someone else's problem. Yet the shape of the question is identical: "For how many years will this AI system we are introducing right now remain competitive?" An answer of three years makes payback calculations tight, while an answer of five years makes them generous. And typically, the number where the calculations neatly align is quietly selected. Flip this around. Before calculating the payback period, sit the technology lead and finance lead at the same table and agree on the useful life first. Asking in that room, "Do you think we will still be using this system as-is three years from now?" immediately reveals how far apart the financial statement's five-year span and the frontline's sense are. A payback period calculated without knowing that gap is merely an incorrect number calculated through an accurate method.

Stage 4: Establishing Stop Criteria and Re-evaluation Points Concurrently with Approval

The final stage is a matter of procedure, not calculation. No matter how elaborately a payback period is derived, if no provision is made for when to re-verify that calculation, the numbers remain only on the approval documents.

What Alibaba is doing at this juncture is worth noting. The company announced a 380 billion yuan AI infrastructure investment plan over three years in February 2025, and stated during this earnings conference call that cumulative execution reached 190 billion yuan as of the end of the June 2026 quarter, with progress largely matching plans. At the same time, it specified why spending scale was somewhat high this quarter—citing fluctuations in hardware delivery schedules, expanded CPU procurement driven by surging demand for AI agents, and rising chip component prices—clarifying that quarterly CAPEX is not evenly distributed.

Three things stand out here. There is a planned total amount, the execution rate up to the present, and an explanation of variances against the plan. Once these three are in place, whether external investors or the internal board of directors, they can pose the exact same question again next quarter. Conversely, if there is only a total amount without an execution rate or variance explanation, the investment effectively moves outside of monitoring post-approval.

The McKinsey survey suggests that the presence or absence of this procedure links to performance. Only 6% of high-performing companies that reported generating over 5% of company-wide operating profit through AI stated they possessed procedures to measure the effects of AI initiatives, but these companies were twice as likely to report having such procedures compared to other companies. They were also twice as likely to report that management demonstrates execution resolve regarding AI initiatives. However, because this survey is survey-based, it is more accurate to understand this not as a causal relationship where performance occurred because procedures were in place, but as the characteristic most consistently observed in companies that achieved performance.

There is one more item easily overlooked when designing stop criteria: operating expenses. In the McKinsey survey, one in five respondents answered that they were restricting AI usage due to AI operating expenses, including token costs. This means that what prevents the recovery of AI investments is increasingly turning out to be monthly recurring usage fees rather than initial deployment costs. This portion is easily omitted in CAPEX reviews. At the time of deployment, only license fees and build costs are calculated, and inference costs that mount the more the system is used quietly pass into operating expense categories later on. One of the reasons cited directly by Alibaba for the quadrupling of losses in its AI Labs division was consumer AI app inference costs.

Management Insight

Organizations that establish stop conditions concurrently when approving investments are rare because it appears as though one is talking about failure before even starting. Yet if you invert the sequence, this is actually a device protecting the person in charge. If a sentence exists from the beginning stating, "If this metric fails to reach this level in 6 months, scale back; if there is no improvement even at 12 months, pull the plug," the person in charge loses any reason to hide bad news. Conversely, without stop criteria, the poorer the performance, the later the person in charge reports that fact because poor performance directly translates to personal accountability. Add just one more line here: a condition for re-evaluation if monthly operating expenses exceed a certain threshold. Deployment costs end with a single sign-off, but operating expenses quietly increase every month. And that quiet increase is what most frequently breaks payback period calculations.

Why Seeing the Same Numbers Leads to Different Conclusions

Let us return to Alibaba. Immediately following the earnings release, U.S.-listed shares reportedly fell 4.6% in pre-market trading, with losses later widening to the 6% range. This was driven by non-GAAP diluted earnings per share coming in at 8.52 yuan, falling short of market expectations of 10.53 yuan. On the other hand, some securities firms, including Nomura, picked cloud growth failing to reach its peak yet as the single most important signal, emphasizing that this growth is manifesting alongside margin improvements without sacrificing profitability.

The reason different conclusions emerged from looking at the same data is not because the information differed. The criteria for what constitutes a signal of recovery simply varied. Viewed on the basis of quarterly net profit, this quarter is a clear retreat. Viewed on the basis of segment margin trends and progress against execution plans, it is an ongoing investment. Which side is right will only become known after several quarters pass.

What South Korean business executives should take away from this case is not Alibaba's numbers, but this structure. Before asking about the AI investment payback period, four things must be sorted out first. Have you divided what was spent into layers? Have you divided what returns into costs, revenue, speed, and risk? Have finance and operations jointly agreed on how many years to divide it across? Have you determined at the time of approval when to review it again and under what conditions to stop?

Payback period numbers derived without these four elements sorted out generally fall into one of two categories: numbers manufactured to get an investment passed, or numbers manufactured to block an investment. Neither can be used as a basis for management judgment. Considering the McKinsey survey finding that the proportion of respondents planning to increase AI investments reaches 60%, this question will return repeatedly over the next few years. Whether to newly manufacture answers each time, or to establish decision criteria once and continuously refine those criteria—that is what must be decided before calculating the payback period.