---
title: "Microsoft's AI Capital Expenditure Enters the Realization Phase: How Azure and Copilot Determine ROIC"
type: "Topics"
locale: "en"
url: "https://longbridge.com/en/topics/42849449.md"
description: "The AI infrastructure race has entered its second phase. The market no longer looks solely at the number of GPUs, data center scale, and cloud revenue growth rates, but instead begins to ask three more direct questions: How much revenue can new computing power generate? How much profit can be accumulated per megawatt of computing power? When will capital expenditures convert back into free cash flow? This is precisely the core of the current valuation divergence for Microsoft. Azure proves whether AI demand is real, Copilot determines whether Microsoft can transform underlying computing power into application-layer revenue, while gross margin and free cash flow are responsible for testing whether this business model possesses a sufficiently high return on capital..."
datetime: "2026-07-22T08:53:09.000Z"
locales:
  - [en](https://longbridge.com/en/topics/42849449.md)
  - [zh-CN](https://longbridge.com/zh-CN/topics/42849449.md)
  - [zh-HK](https://longbridge.com/zh-HK/topics/42849449.md)
author: "[潘驴邓晓闲缺一](https://longbridge.com/en/profiles/27015735.md)"
---

# Microsoft's AI Capital Expenditure Enters the Realization Phase: How Azure and Copilot Determine ROIC

The AI infrastructure race has entered its second phase. The market no longer looks solely at the number of GPUs, data center scale, and cloud revenue growth rates, but begins to ask three more direct questions: How much revenue can new computing power generate, how much profit can be accumulated per megawatt of computing power, and when will capital expenditures convert back into free cash flow.

This is the core of the current valuation divergence for Microsoft. Azure proves whether AI demand is real, Copilot determines whether Microsoft can transform underlying computing power into application-layer revenue, and gross margin and free cash flow are responsible for testing whether this business model has a sufficiently high return on capital.

In the following text, FY refers to Microsoft's fiscal year ending in June, and E stands for institutional forecasts.

## **I. Capital Expenditures Enter High-Pressure Zone; Microsoft Must Prove the Unit Economics of AI Investments**

Morgan Stanley estimates that Microsoft's total capital expenditures for FY26 to FY29 will be $145.4 billion, $250 billion, $312.6 billion, and $375.1 billion, respectively; operating cash flow during the same period will increase from $172.5 billion to $298.5 billion, but free cash flow will drop rapidly from $62.5 billion to $14.6 billion, $6.3 billion, and $7.9 billion. Institutional models also assume that Microsoft will issue approximately $45 billion in additional debt in FY28 and reduce the scale of long-term share buybacks to preserve balance sheet space needed for AI investments.

**Financial Indicators**

**FY26E**

**FY27E**

**FY28E**

**FY29E**

Operating Revenue

$329.5 billion

$389.3 billion

$474.1 billion

$587.5 billion

Total Capital Expenditures

$145.4 billion

$250 billion

$312.6 billion

$375.1 billion

Free Cash Flow

$62.5 billion

$14.6 billion

$6.3 billion

$7.9 billion

Gross Margin

67.7%

65.7%

64.4%

63.4%

Operating Profit Margin

46.6%

46.5%

46.7%

47.2%

EPS

$17.35

$19.62

$23.86

$29.83

According to the above model, the compound annual growth rate (CAGR) of revenue from FY26 to FY29 is approximately 21%, and the CAGR of EPS is approximately 20%, but the free cash flow margin will drop from 19.0% in FY26 to 1.3% in FY28. It should be noted that "total capital expenditures" by institutions may include items such as finance leases, which are not exactly the same caliber as the cash capital expenditures deducted from free cash flow, so one cannot simply recalculate FCF by subtracting total Capex from operating cash flow. However, the trend is unambiguous: over the next two to three years, Microsoft's income statement will remain strong, but its cash recovery ability will clearly yield to infrastructure construction.

Therefore, the indicator most prone to misjudgment currently is forward PE. Calculated based on the closing price of $402.29 on July 20, 2026, Microsoft's PE corresponding to FY28 EPS of $23.86 is approximately 16.9 times, which appears lower than most large software companies; however, rough calculation based on the market capitalization of approximately $29.95 trillion on the report date and predicted free cash flow of $6.3 billion for FY28 shows an FCF yield of only about 0.2%.

Microsoft is not a "low-valuation software stock" in the traditional sense. The current valuation implies a special stage where capital expenditures are front-loaded and profit recognition is faster than cash recovery.

Whether a forward PE of 16 to 17 times is cheap depends on whether AI assets can deliver returns higher than the cost of capital in the future, rather than whether EPS itself can grow.

## **II. Accelerating Azure Growth Only Proves Demand, Not Return on Investment**

The main constraint on Azure in the recent stage is not insufficient orders, but insufficient available computing power. Microsoft management expects Azure's growth calculated at fixed exchange rates to reach 39% to 40% in the fourth quarter of FY26, and judges that the growth rate in the second half of 2026 will accelerate slightly compared to the first half. After new data centers and GPUs go online, Microsoft will not only be able to meet current demand but also absorb workloads previously delayed due to capacity shortages.

Morgan Stanley estimates that Azure and other cloud service revenues will increase from $105.6 billion in FY26 to $150.3 billion in FY27, $214.9 billion in FY28, and $305.9 billion in FY29, with corresponding growth rates of 40%, 42%, 43%, and 42%. This means that after the revenue base exceeds $100 billion, Azure may still maintain growth exceeding 40%.

However, Azure's growth rate is not a sufficient condition for AI investment returns. New revenue can come from three completely different types of businesses: The first type is GPU computing, training, and inference capacity, where the core competitive variables are hardware supply, utilization rate, and unit computing power price; the second type is database, storage, security, development tools, and AI platform services, with revenue deeply tied to customer workloads; the third type is first-party applications such as Microsoft 365, GitHub, Dynamics, and Security, where Microsoft can simultaneously obtain software subscription, package upgrade, and AI call revenues.

The capital efficiency differences among these three types of revenue are significant. Institutions estimate that in FY28, Microsoft's Azure AI will occupy 8.71 GW of capacity, generating approximately $92.7 billion in revenue, with unit MW revenue of approximately $10.65 million; first-party applications will only occupy 4.25 GW of capacity but can generate approximately $128 billion in revenue, with unit MW revenue of approximately $30.1 million, nearly 2.8 times that of Azure AI. The unit MW revenue for M365 Copilot is expected to be approximately $20.14 million, also significantly higher than the Azure AI infrastructure layer.

Therefore, Microsoft's advantage does not lie simply in owning more data centers. What truly affects return on capital is the structure of computing power allocation: how much capacity remains at the GPU leasing layer, how much capacity can drive Azure platform services, and how much capacity ultimately enters M365, GitHub, and enterprise workflows.

If new computing power is mainly used to sell basic computing resources, Azure revenue can grow, but gross margin and return on capital may not improve; if AI workloads can continuously drive database, security, development tools, and first-party software revenues, the commercial value of the same GPU will become significantly higher.

## **III. Revenue per Megawatt Determines Whether Microsoft is a Cloud Platform or a Computing Power Leasing Company**

Based on the estimated 8.71 GW Azure AI capacity for FY28, Morgan Stanley constructed three monetization scenarios. Although the hardware scale is the same across the three scenarios, the differences in revenue and profit are enormous.

**FY28E**

**Commoditized Computing Power**

**Azure Platform Model**

**Full-Stack AI Platform**

Azure AI Revenue

$86.2 billion

$117.9 billion

$198.3 billion

Revenue/MW

$9.9 million

$13.54 million

$22.77 million

Gross Margin

25%

30%

35%

EBIT

$12.1 billion

$22.4 billion

$47.6 billion

EBIT/MW

$1.39 million

$2.57 million

$5.46 million

In the first scenario, Azure AI approaches commoditized Neocloud: customers mainly purchase GPUs, training, and inference resources, with low platform service attachment rates. Microsoft bears data center, chip, power, and depreciation costs, but pricing power is suppressed by the expansion of computing power supply. In FY28, revenue per MW is only $9.9 million, with an operating profit margin of 14%.

The second scenario replicates the path of traditional Azure: IaaS serves as the entry point, subsequently driving consumption of storage, databases, security, analytics, and development platforms. Revenue per MW rises to $13.54 million, and the operating profit margin increases to 19%. Compared to commoditized computing power, Microsoft can generate approximately $31.7 billion in additional revenue and $10.3 billion in EBIT without increasing capacity.

The third scenario requires Microsoft to simultaneously control infrastructure, AI platforms, and application workflows. Azure AI not only provides computing power but also drives Copilot, Agent, data services, development tools, and automation products. Revenue per MW reaches $22.77 million, and EBIT/MW reaches $5.46 million. Compared to the first scenario, the revenue difference in FY28 is approximately $112.1 billion, and the EBIT difference is approximately $35.5 billion.

What affects Microsoft's market capitalization is not the 8.71 GW itself, but whether these capacities can migrate from infrastructure revenue of approximately $10 million per MW to platform revenue above $13.5 million.

The cautious pricing in the current market is closer to the first scenario; the institutional baseline forecast lies between the first and second scenarios. The main expectation gap generated thereby is not that Azure demand exceeds expectations, but whether Azure AI can replicate the profit path of traditional Azure migrating from IaaS to PaaS.

This judgment requires observing three operational indicators: the attachment rate of Azure AI platform services, unit MW revenue, and Azure AI gross margin. A single-quarter Azure growth exceeding 40% alone cannot prove that Microsoft has passed the commoditized computing power stage.

## **IV. Copilot is the Application-Layer Outlet for Increasing Unit Computing Power Output**

The importance of Copilot does not lie in whether it becomes a multi-billion-dollar product, but in whether it can change the revenue structure of Microsoft's AI. Azure is responsible for selling computing power to customers, while Copilot directly enters Office, Teams, GitHub, and enterprise workflows, converting computing costs into subscription, package upgrade, and consumption revenues.

Microsoft is forming a three-tier charging structure. The first tier is Copilot paid seats, with revenue mainly driven by the number of users; the second tier is the M365 E7 upgrade, bundling E5, Copilot, and Agent capabilities to increase per-user revenue on the M365 install base; the third tier is Agent calls, task orchestration, and workflow automation, charged based on actual usage, making revenue more closely correspond to inference costs.

**Copilot Three-Tier Monetization Path**

Paid Seats → E7 Package Upgrade → Agent and Workflow Consumption Based on Usage

Institutions estimate that M365 Copilot revenue will increase from $4.43 billion in FY26 to $22.48 billion in FY29, with paid seats increasing from 26 million to 114.9 million. More noteworthy is that in the model, FY29 revenue still grows by 51%, while the end-of-period seat growth rate has dropped to 11.7%. Although annual revenue and end-of-period seats cannot be directly used to calculate ARPU, this gap clearly indicates that institutional forecasts no longer rely solely on new seats, but implicitly include average seat growth, E7 upgrades, and contributions from pay-per-use consumption.

CIO surveys also support rising enterprise adoption rates. 88% of respondents expect to use M365 Copilot in the next 12 months, with the current proportion of enterprises primarily adopting E7 at about 7%, rising to an expected 21% for the next year. However, this survey only covers 60 US and European CIOs, suitable as a procurement direction indicator, but cannot replace actual paid seats, deployment penetration rates, usage frequency, and renewal data.

Whether the Copilot business model can succeed hinges not on "whether enterprises have purchased it," but on conversion rates at four levels: the proportion of pilot customers converting to formal deployment, the speed of expanding formal deployments to all employees, the proportion of E5 customers migrating to E7, and whether Agent calls generate stable consumption revenue.

If Copilot remains stuck in fixed seat charging for the long term, increased usage frequency may lead to inference costs growing faster than revenue, putting pressure on gross margins; if Microsoft can link prices to usage through package quotas, overage calls, and consumption commitments, the higher the inference demand, the larger the revenue and profit pool. This determines whether Copilot is merely a relatively costly Office feature or a high-value outlet for Microsoft's AI capital expenditures.

## **V. Declining Gross Margins Do Not Equal Worsening ROIC, But Unit Costs Must Continue to Decrease**

Microsoft's overall gross margin is expected to drop from 67.7% in FY26 to 63.4% in FY29, a cumulative decline of 4.3 percentage points; meanwhile, the operating profit margin will rise from 46.6% to 47.2%, and EPS will increase from $17.35 to $29.83. Institutions assume that revenue scale and expense discipline can offset gross margin pressure, allowing operating profits to continue growing rapidly.

The decline in gross margin mainly comes from three aspects: the increasing proportion of AI infrastructure in revenue; newly launched assets starting to depreciate before utilization rates mature; and the computing costs of fixed-price products like Copilot being higher than traditional SaaS.

Institutional models predict that Azure AI gross margins will rise from approximately 10.3% in FY25 to 28.5% in FY26, 42.5% in FY28, and 47.5% in FY29, still lower than the approximately 64% to 65% gross margins of Core Azure during the same period.

For this profit margin path to hold, Microsoft needs to simultaneously improve asset turnover, software efficiency, and hardware costs.

At the infrastructure delivery end, Microsoft stated that the device-to-online cycle in the third quarter of FY26 was shortened by nearly 20%, the Fairwater data center went into operation six weeks early, and added over 1 GW of capacity. Shorter construction and debugging cycles can reduce the time gap between assets starting depreciation and generating revenue. Meanwhile, inference throughput increased by approximately 40%, meaning the same GPU can process more billable Tokens.

At the chip end, Maia 200 improves output per dollar of Tokens by over 30% compared to Microsoft's existing latest chips, and Cobalt server CPUs have been deployed to nearly half of the data center regions. Self-developed chips not only reduce concessions to third-party hardware manufacturers' profit margins but also allow Microsoft to collaboratively optimize networks, storage, virtualization, models, and inference software stacks.

Traditional cloud computing provides a reference. Microsoft's capital expenditures expanded more than fourfold from FY12 to FY17, cloud revenue grew approximately tenfold from FY14 to FY18, cloud business capital intensity dropped from over 200% in FY14 to approximately 50% in FY18, and further dropped below 30% in FY22.

However, AI cannot mechanically replicate the traditional cloud computing cycle. GPU depreciation is faster，单机 costs are higher, power constraints are stronger, and model efficiency iterations may rapidly change hardware demand. The steady-state gross margin of AI infrastructure is likely lower than traditional Azure. Therefore, judging whether the decline in gross margins is reasonable cannot rely solely on historical analogies, but must verify whether Azure AI unit costs decrease according to the model.

The most important risks include utilization rates below expectations, premature obsolescence of assets due to GPU iterations, rising power and supply chain costs, declining AI computing power prices, and usage growth of fixed-price products outpacing revenue.

## **VI. Valuation Divergence Primarily Comes from Business Quality, Not EPS Differences**

Morgan Stanley provided three valuation scenarios: optimistic scenario FY28 EPS $27.39, 29x PE, corresponding to $795; baseline scenario FY28 EPS $23.86, 25x PE, corresponding to $600; pessimistic scenario FY28 EPS $21.64, 12x PE, corresponding to $250.

**Scenario**

**FY28 EPS**

**PE**

**Target Price**

Optimistic

$27.39

29x

$795

Baseline

$23.86

25x

$600

Pessimistic

$21.64

12x

$250

Notably, the FY28 EPS difference between the optimistic and pessimistic scenarios is only about 27%, yet the target prices differ by 218%. In other words, most of the valuation difference does not come from near-term earnings forecasts, but from what valuation multiple the market is willing to assign to Microsoft.

If Azure AI is viewed as a capital-intensive, fiercely price-competitive computing power business, and Copilot fails to form stable ARPU, Microsoft's software valuation framework will be disrupted, and the market may price it using multiples close to those of infrastructure companies. Conversely, if Azure AI's revenue per MW continues to rise, and Copilot drives M365 package upgrades and consumption revenue, the market will view it as a platform software company in the AI era, giving forward PE conditions to recover above 25 times.

The institutional baseline model's predictions for FY28 revenue, operating profit, and EPS are 4.1%, 4.6%, and 3.2% higher than consensus expectations, respectively, but gross margins are approximately 96 basis points lower than consensus expectations; by FY29, revenue and operating profit predictions are 6.8% and 7.8% higher than consensus expectations, respectively. This shows that its bullish logic does not rely on short-term gross margin improvements, but on the sustained growth of Azure and Copilot and the decline in operating expense ratios.

The main problem with the baseline scenario lies here too: a 25x FY28 PE requires the market to believe that FY28 is not the earnings peak, that capital expenditure growth rates will decline, that Azure AI gross margins will continue to improve, and that Copilot revenue can still maintain high growth. If the investment cycle continues to extend and free cash flow remains close to lows, a 25x PE lacks sufficient cash flow support.

## **Buy-Side Conclusion: Medium-Term Logic is Bullish, but Azure Growth Rate is Not a Sufficient Buy Signal**

The core advantage of Microsoft's AI investment lies in simultaneously controlling infrastructure, cloud platforms, enterprise software, and user workflows. Compared to pure computing power suppliers, Microsoft has the ability to charge multiple times for a single underlying computing resource towards Azure PaaS, M365, GitHub, Security, and Agent. This is the fundamental reason why its unit computing power revenue and long-term ROIC may be higher than Neocloud.

But we are still in the proof stage. Azure's growth exceeding 40% only confirms strong demand and capacity release, unable to independently confirm return on capital. What truly determines whether valuation can migrate from 16 to 17x FY28 PE to 25x is whether Azure AI's revenue per MW can move from approximately $10 million towards the $13.5 million platform scenario, whether Azure AI gross margins can enter the above 40% range, and whether Copilot revenue can continue to grow faster than paid seat growth.

**Five Verification Indicators for the Next Two to Four Quarters**

1\. Azure fixed exchange rate growth rate and speed of new capacity conversion;  
2\. Azure AI Revenue/MW;  
3\. Gap between Copilot paid seat growth and revenue growth;  
4\. Contribution of E7 and Agent consumption revenue;  
5\. Marginal changes in cash capital expenditures, free cash flow, and financing needs.

A more reasonable judgment currently is: Microsoft possesses the optimal asset portfolio for full-stack AI monetization, with medium-term fundamentals leaning bullish, but the cash flow side has not yet provided an unconditional safety margin. If Microsoft can only prove "AI demand is very strong," but cannot prove "each unit of computing power is becoming more profitable," a 16 to 17x FY28 PE does not constitute absolute undervaluation; if Azure platform attachment rates, Copilot ARPU, and AI gross margins are realized simultaneously, the $600 baseline scenario will possess sufficient earnings and valuation support.

Data and viewpoints source: Microsoft public disclosures, Morgan Stanley Research, Visible Alpha, etc. Predictive data in the article are all institutional estimates and do not represent company guidance.

This article is for research exchange only and does not constitute any investment advice.

### Related Stocks

- [MSFT.US](https://longbridge.com/en/quote/MSFT.US.md)
- [MSFL.US](https://longbridge.com/en/quote/MSFL.US.md)
- [MSFO.US](https://longbridge.com/en/quote/MSFO.US.md)
- [MSFD.US](https://longbridge.com/en/quote/MSFD.US.md)
- [MSFU.US](https://longbridge.com/en/quote/MSFU.US.md)
- [MSFX.US](https://longbridge.com/en/quote/MSFX.US.md)
- [MSFY.US](https://longbridge.com/en/quote/MSFY.US.md)

## Comments (3)

- **投资小达人 · 2026-07-23T06:14:10.000Z · 👍 1**: What 🪜 is good for placing an order to buy Microsoft?
  - **坐等个股起飞** (2026-07-23T06:14:51.000Z): Recommend 🚀 or clash meta, etc.
- **顺势&择时 · 2026-07-22T19:37:27.000Z**: This is written quite clearly. Friends who have been tortured by $Microsoft(MSFT.US), don't know what happened, and can only curse the stock should take a look.At least you'll know what you're buying, what you're betting on, roughly how long the potential return period is, and whether a six-month ca