中文 · English
EQUITY RESEARCH · MSFT November 16, 2026

Microsoft (MSFT) FY27 Q1 Earnings Review: Azure Up 48%, Fair Value $447–517

Both businesses are accelerating. I remain bullish.
Transurf Research · November 16, 2026
Microsoft Azure M365 OpenAI D&A Valuation
Both businesses are accelerating while the disclosures ahead grow harder to see
By: Transurf Research Published: November 16, 2026
+48%
Azure growth, year over year
3 points above the company's guidance of about 45%
49-50%
Azure guidance, next quarter
Constant currency
$10.0B
OpenAI revenue, one quarter
10.84% of quarterly revenue
$447-517
Fair range
Price $521.91 (October 30 close)

Overview

▼

Microsoft's quarter came with one new number and four missing sets of numbers.

Azure grew 48%, guidance for next quarter is 49–50%, and a batch of leased compute slipped to the first half of next year. M365 commercial cloud revenue grew 18% in constant currency, up from 15% a year ago. Price drove the acceleration. Seat growth actually slowed.

The new number is OpenAI. In one quarter, Microsoft recognized $10 billion of revenue from it, about 11% of the total. Of the four missing sets, the most important are commercial bookings and RPO excluding OpenAI. April and July still gave growth rates. October gave only adjectives. Think of a restaurant that reports what its biggest table ordered and, this quarter, stops reporting what the other tables ordered.

Nadella says he wants a customer base that is broad and diversified. This quarter gives no way to measure how far he has got, and no evidence that it has worsened.

Returns are the same story. In four earnings calls, an analyst asked every time, and not once got a number. So I ran the numbers myself. In the base case, annual D&A by FY29 is close to three times FY26. The model also holds a cost whose source I cannot pin down. It affects whether the stock is expensive now, so I break it out separately.

Both businesses are accelerating. I remain bullish.

Ticker: Microsoft (MSFT) · Period: FY27 Q1 (Microsoft's fiscal year runs from July 2026 to June 2027), reported after the close on October 28, 2026 · Written: November 1, 2026 · Reference price: $521.91 (October 30 close)
Key conclusions

Microsoft runs two main businesses: the office software companies use every day, and the cloud they use to run AI (Azure). This quarter, third-party compute arrived late, yet Azure still grew 48%, with 49–50% guided for next quarter. M365 commercial cloud revenue is also accelerating. Foundry, Microsoft's platform for companies to build AI applications, more than tripled its revenue year over year. On the other side, the company no longer gives numbers on who is buying besides OpenAI. Nor has it said how much this compute will earn. I ran the numbers myself. At 26–30x FY29 EPS, the fair range is $447–517, and today's $522 sits only about 1% above the top. The model holds a cost of about $9B whose source I cannot yet explain. Without it, $522 sits inside the fair range. Both businesses are accelerating, and I remain bullish.

1. Azure: compute arrived late, growth still hit 48%, and the next-quarter guide is not aggressive

▼

Third-party compute arrived late, yet Azure still grew 48%. Next quarter's 49–50% needs only about $2.9B more than this quarter, less than the $3.4B and $3.7B added in the past two

Azure did not wait for the late machines this quarter. Azure is the Microsoft business that rents out compute and AI services. Its revenue this quarter was $33.2B, up 48% year over year and 3 points above the company's guidance of about 45%. It is 36% of total revenue, the largest piece.

CFO Amy Hood said on the call that part of the third-party compute (machines leased from other companies) expected to arrive this quarter was delayed. In other words, the extra 3 points did not come from those machines. The next section breaks down where they came from.

On profit, company-wide operating margin was 50%, slightly above a year ago. Jefferies' Brent Thill asked whether 50% revenue growth and a 50% margin could hold at the same time. Hood's answer was that growth itself is the answer, and that efficiency brings stability and quality, not only savings. She gave no number.

Azure guidance for next quarter is 49–50% in constant currency. On this quarter's revenue, the midpoint needs about $2.9B more than this quarter. The previous two quarters added $3.4B and $3.7B. The required increase is smaller than in either of those quarters, so the guidance is not aggressive. It also shows that part of the rising growth rate comes from the base: the same quarter last year added only about $1.7B.

Two things the company cannot call with confidence will decide whether next quarter's guidance holds: the token price, and whether third-party compute arrives on time. A token is the smallest unit of text an AI model processes. Charging per token is like charging per word. Hood said the company still needs to find more efficiency next quarter, and it is counting on that. She also noted that the second half compares against last year's acceleration, a higher base.

2. Where the beat came from: Foundry, machine utilization and token prices

▼
01FoundryRevenue more than tripled YoYRepeatable02Fuller use of machinesIdle CPU capacity monetizedCopilot throughput >50% QoQRepeatable03Higher token pricesThe company cannot call itIn-quarter prices still unclearUncertainNote: 90% of Copilot requests useautomatic routing; Copilot Fusion costsup to 77% less, pushing prices downMy model assumes AI prices fall 5% to 10% a year FoundryRepeatableRevenue more than tripled YoYFuller machine useRepeatableIdle CPU capacity monetizedCopilot throughput >50% QoQHigher token pricesUncertainThe company cannot call itIn-quarter prices still unclearNote: 90% of Copilot requests auto-routed;Copilot Fusion costs up to 77% lessMy model assumes AI prices fall 5% to 10% a year

The beat came from Foundry, fuller use of existing machines and higher token prices. The company cannot call the price, and its own routing keeps pushing it down

The beat this quarter came from three places: Foundry, fuller use of machines, and token prices. The first two can repeat. The third is uncertain.

Hood said Foundry and better CPU monetization drove Azure growth. Foundry is Microsoft's platform for companies to build AI applications. On it, customers call models, build agents and pay by usage. Its revenue more than tripled year over year.

The second source is fuller use of existing machines. On the CPU side, engineering teams found idle capacity and monetized it. For Copilot workloads, throughput rose more than 50% quarter over quarter. In effect, the company earns more from its old machines before it buys new ones.

The third source is the token price. Hood said token prices this quarter were higher than expected, so the same usage brought in more revenue. The company cannot call this number. In the same answer, she added that token prices within the quarter still carry some unknowns.

The company's own strategy is pushing the price down. Copilot Fusion combines several models, and the company says it reaches frontier-level quality at up to 77% lower cost. Automatic routing handles 90% of Copilot requests and picks a model on quality, speed and cost. The more cheap models it uses, the lower the average price. This is my inference. The company has not said it.

So on this line, Foundry and efficiency can repeat, and the price is uncertain. My model assumes AI prices fall 5% to 10% a year. It does not assume they keep rising.

3. M365: seat growth slows and pricing drives the acceleration

▼

Paid seat growth fell from 7% to 5%, yet M365 commercial cloud growth in constant currency rose from 15% to 18%. GitHub Copilot drove most of the beat

M365 is accelerating, but the seat count is not what is accelerating.

M365 cloud (commercial plus consumer) brought in $27.8B this quarter, 30% of total revenue. Within it, commercial cloud, which counts only business customers, grew 18% in constant currency, against 15% a year ago. All growth rates here use the figures the company recast in September, when it restated its history under the new segment structure.

Paid seats grew only 5% this quarter, against 7% in the July quarter. Both figures use the recast September definition. A paid seat is one employee paying a monthly fee. Slower seat growth means new seats are selling more slowly.

So where does the 18% come from? Year-over-year growth rests on two things: price and seats. Price gains come mainly from premium bundles such as M365 Copilot, E5 and E7, which cost more than the base suite. E7 already has several million paid seats. The upside versus the company's guidance came mainly from GitHub Copilot adoption and usage, partly offset by weaker E3 (the base suite for large enterprises) and standalone offerings.

In September, the company changed its classification and moved GitHub into M365 commercial cloud. So GitHub Copilot revenue now sits in this line, and seat growth also uses the new classification.

Guidance for M365 commercial cloud next quarter is about 19% in constant currency.

One line on Windows: OEM and Devices revenue fell 22% year over year. Guidance for next quarter is a decline of about 35%, and channel inventory is still high.

The company did not update the total Copilot paid seat count this quarter. Bank of America's Tal Liani asked directly whether there was a Copilot number. Hood answered that the company would not update it this quarter. The last figure was more than 30 million seats at the end of June. In January, M365 commercial seats were above 450 million, up 6% year over year. Divide 30 million by that and Copilot penetration is about 7%. The two figures come from different dates, so read the result only as an order of magnitude.

4. AI agents: usage is rising, but only two revenue signals are visible

▼
Agents are doing a lot of work, but most of the revenue ledger is still blank

Outside coding, the company expects a slower pace, yet the revenue under agents is already accelerating: Foundry went from more than doubling to more than tripling, and Copilot revenue growth is speeding up

Agent revenue is accelerating, but only part of it is visible.

An AI agent is an AI program that breaks a task down and completes it in several steps, without a person directing each step. A chatbot and an agent differ the way a GPS and a hired driver do. The GPS tells you the way. The driver drives you there.

Two accelerations are visible. In July, Foundry revenue had "more than doubled." In October, it had "more than tripled." Copilot revenue growth sped up quarter over quarter. That statement covers M365, GitHub and Security together, so it cannot be read as the growth rate of M365 Copilot alone.

There are also three usage figures. More than 50 million agents are registered in Agent 365 (Microsoft's platform for managing enterprise agents), up more than 30% quarter over quarter. Agents open more than a third of pull requests on GitHub. Copilot customers on consumption billing grew more than 75% quarter over quarter. The first two are a count and a share, and the third is a customer count. None of them is revenue.

Discount the registration count. The 50 million is the number registered in a directory. It is not the number of paid agents, nor the number that actually work each month. Treating it as 50 million employees overstates it.

Foundry is not pure agent revenue either. It is a platform for building AI applications, and it holds more than agents. The company has never broken out the split. On the call, the very next point was that Azure OpenAI benefited from OpenAI's new models. So Foundry's growth cannot all count as agents.

Three things remain unmeasured: standalone agent revenue, the number of paid agents, and absolute daily users. The company gave only growth rates. Copilot daily users rose nearly 50% quarter over quarter and more than 5x year over year. In April, the company put the AI business's annualized revenue above $37B, up 123% year over year. The two calls since then have not updated it.

Management itself says agents will move more slowly from coding into other knowledge work. Hood also said the company "still sees encouraging signs" and "remains optimistic." CEO Satya Nadella said the breadth of enterprise AI, meaning how many customers and use cases it reaches, is "a decade-long thing," not a one-quarter thing.

Usage is rising, but revenue is visible in only two places, and knowledge-work pricing has no numbers yet. That raises the next question: besides OpenAI, who else is buying?

5. OpenAI is 11% of revenue, and four demand metrics are no longer disclosed

▼
Commercial bookingsexcluding OpenAIRPOexcluding OpenAICopilot paid seatsGW added per quarterApr+7%growth+26%growthNot given~1 GWper quarterJul+18%growth+25%growthOver 30M seatsend of June~1 GWper quarterOctAdjectives onlyAdjectives onlyNo updateNot disclosedNew this quarter: OpenAI at $10.0B for the quarter, 10.84% of revenue. AprJulOctCommercial bookings excluding OpenAI+7%growth+18%growthAdjectivesonlyRPO excluding OpenAI+26%growth+25%growthAdjectivesonlyCopilot paid seatsNot given30M+ seatsend of JuneNo updateGigawatts added per quarter~1 GWper quarter~1 GWper quarterNot disclosedNew this quarter:OpenAI $10.0B in the quarter, 10.84% of revenue.

OpenAI brought in $10.0B in one quarter, about 11% of revenue. On who else is buying, four sets of numbers stopped in October

This quarter added one big-customer figure and dropped four sets of numbers.

The 10-Q's related-party note covers counterparties in which Microsoft holds equity. It states that revenue recognized from OpenAI this quarter, including revenue share, was $10.0B, or 10.84% of quarterly revenue. $10.0B is one quarter, on the order of $40B a year. Microsoft is also an OpenAI shareholder. It committed $13.0B, has invested $12.0B, and has $1.0B left to invest.

This quarter shows nothing about how large demand outside OpenAI is. Think of a restaurant that reports what its biggest table ordered and, this quarter, stops reporting what the other tables ordered. Microsoft disclosed a customer above 10% of revenue and, at the same time, removed four sets of numbers.

The first two sets, commercial bookings excluding OpenAI and RPO excluding OpenAI, measure demand outside OpenAI directly. The other two are Copilot paid seats, which measure Microsoft's own product, and quarterly gigawatts added, which measure capacity. A gigawatt is a unit of power, one billion watts. Here it measures how much electricity data centers can draw. In April, the first two still came with growth rates (bookings +7%, RPO +26%), and so did July (+18% and +25%). October gave only adjectives. The last two were not updated in October.

The wording also changed. Last fiscal year's 10-K said "no single customer or country other than the United States accounted for more than 10% of revenue." The September 10-Q changed it to "no single country other than the United States accounted for more than 10%," and the word "customer" is gone. The sentence names no one. The name is in the related-party note. The filing itself discloses it, and I do not think this is deliberate concealment. But readers can no longer read concentration from that sentence.

RPO is large, but new contracts are shrinking. It stands at $677B, up 73% year over year, and this quarter includes a large Azure contract. Weighted by dollar value, the average remaining term is 2.3 years, and about 30% will be recognized in the next 12 months. Commercial bookings fell 12% year over year, against +112% a year earlier, a very large base. This figure is total contract value, not profit, and it says nothing about gross margin.

What if OpenAI's revenue disappeared? At a 20% tax rate and 7.446 billion shares, losing $10B of revenue in one quarter with no cost savings cuts EPS by about $1.07. Over a full year that is about $4.30 (on unrounded figures), about a fifth of base-case FY28 EPS of $20.85. The base case is the scenario I think most likely. If the company could cut costs equal to half of that revenue, the annual hit would be about $2.15. This is a sensitivity, not a forecast.

Three analysts touched on the topic this quarter, and none got a number. Bernstein's Mark Moerdler asked about the mix of enterprise and AI-native customers and got a string of customer stories. UBS's Karl Keirstead mentioned the intensity of OpenAI and got no answer. Goldman's Gabriela Borges asked about the 11% and got a pitch on model choice.

Nadella stated his position on the same call. In his view, a few large customers buying lots of compute to carry a few quarters of growth is not a business. He wants breadth: many, diverse customers. When large customers come to buy, he will sell some if the price is right. If it is not, he will patiently build breadth. This describes what the company wants, not how much breadth it has achieved.

So this quarter gives no way to measure breadth, and no evidence that it has worsened. Who is buying decides how much this compute will earn. First, look at how large a bill the company has signed for it.

6. Capex and leases: $364.6B on top of the $170B

▼

The company's 2026 capex guidance moved from $190B to about $170B, while leases not yet commenced grew $35.5B in one quarter to $364.6B

This bill has two columns. One is the capex the company shows you. The other is leases not yet on the balance sheet. Read the two columns together.

The two types of lease follow different accounting. A finance lease is in substance an installment purchase of equipment. The equipment counts as the company's own, and the payments count as debt service. An operating lease is in substance a rental, and the rent counts as an operating cost. Finance leases count toward capex, and operating leases do not.

Capex guidance for calendar 2026 was $190B in April, about $175B in July and about $170B in October. The two cuts had different causes. In July, the company moved part of its future data center leases from finance leases to operating leases. In October, third-party compute deliveries slipped. In July, the company said that apart from the lease classification, its 2026 investment plan was unchanged. In October, it said the delayed portion moved to the second half of this fiscal year (January to June 2027).

Actual capex for the quarter ended September was $45.2B, and guidance for the quarter ending December is above $50B. With $31.9B for March, $41B for June, $45.2B for September and $50B for December, the year totals at least $168.1B. That is consistent with "about $170B."

The really large column is in the notes. Leases not yet commenced were $364.6B at the end of September, against $329.1B at the end of June, up $35.5B in one quarter. These leases will commence between FY27 and FY34 (Microsoft's fiscal year starts in July), with terms of 1 to 20 years, and some are conditional.

Put the four companies side by side, and Microsoft holds the least cash and the highest free cash flow. The table uses each company's latest quarterly report. Cash, debt, net cash and leases are as of September 30. Free cash flow and capex cover the trailing 12 months. Free cash flow is operating cash flow minus capex. Net cash is cash minus debt.

ItemMicrosoftAlphabetAmazonMeta
Cash and short-term investments$64.5B$265.1B$114.2B$82.4B
Debt (carrying value)$35.5B$126.1B$180.0B$83.7B
Net cash+$29.0B+$139.0B-$65.7B-$1.2B
Net cash after finance leases-$41.9B+$135.3B-$80.2BNot calculable
Trailing 12-month free cash flow$60.0B$19.7B-$42.4B$25.1B
Leases not yet commenced$364.6B$130.5B$164.9B$355.7B
Trailing 12-month cash capex as % of revenue39%34%27%43%

Differences in the table use unrounded figures, so the last digit may be off by 0.1. Meta does not disclose finance lease liabilities separately, so its net cash after finance leases cannot be calculated. The four companies' methods are not fully the same. Use this table for direction, not for a precise ranking.

Is it enough? Nobody outside can tell, but it is not underinvestment

Whether it is enough cannot be calculated from outside, but it is not underinvestment: $170B is less than Alphabet and Amazon, yet trailing cash capex is 39% of revenue, above their 34% and 27%

Outsiders cannot calculate whether it is enough. But the conclusion that Microsoft is underinvesting does not hold either. Hood said demand still exceeds supply. That is the company's claim, and it means capacity is short now. How much would be enough, outsiders cannot calculate.

Here is why. From this quarter on, the company no longer discloses gigawatts added each quarter. It disclosed about 1 gigawatt in each of the previous three quarters, and nothing this quarter. Without knowing how much compute went in, there is no denominator. Without a denominator, there is nothing to divide capex by.

Spending less does not mean less compute. When a lease moves from finance to operating, the money is still paid, but it no longer counts as capex. Leasing someone else's compute directly means paying a service fee, not buying equipment. In both cases, the reported number falls while the compute does not. The third-party compute delay is different. Both the compute and the bill arrive late, so it is a timing issue.

Microsoft leases compute from neoclouds such as Nebius, IREN and CoreWeave, companies that specialize in renting out AI compute. Their public filings show total contract value. That is the total at signing, not what remained owed at the end of September, so it cannot tell us how much Microsoft still owes.

Spending less has another cause: different denominators. Alphabet's capex also carries Search, YouTube and model training. Amazon's also carries retail and logistics. 2026 guidance is $195–205B for Alphabet, about $220B for Amazon, $135–145B for Meta and about $170B for Microsoft. Comparing these totals directly does not compare like with like.

Switch to a share of each company's own revenue, and the ranking changes. Over the trailing 12 months, cash capex as a share of revenue was 39% for Microsoft, 34% for Alphabet, 27% for Amazon and 43% for Meta. Microsoft spends fewer dollars than Alphabet and Amazon, but a higher share of its own revenue than either. This ratio shows only that Microsoft is not underinvesting. It cannot show that the amount is enough.

To answer whether it is enough, the company would need to resume disclosing installed compute, utilization and revenue per unit of compute. None of the three is available now.

Balance sheet: net debt once finance leases count

Including finance leases, net cash is -$42B, not +$29B. But the $25B new bond is still AAA, and trailing free cash flow of about $60B beats Alphabet, Amazon and Meta

The balance sheet can carry this bill, but Microsoft is no longer a net cash company.

Cash and short-term investments at the end of September were $64.5B, down $12.3B from $76.8B at the end of June. Debt is $35.5B at carrying value and $41.3B at face value. Finance lease liabilities are $71.0B. Cash minus debt is +$29.0B. Subtract finance leases too, and it is about -$42B.

The $27.5B of operating leases is not in the -$42B. Including it gives -$69.4B. My model already counts operating lease rent as a cost every year, so subtracting the liability again would double-count it.

Now the new bond. On October 29, Microsoft priced $25B of new bonds, with settlement expected on November 4. So the bond is in neither September's debt nor September's cash. The ratings are Aaa from Moody's and AAA from S&P, the top rating at each agency, both with stable outlooks. The stated use is "general corporate purposes," which can cover operations, capex, buybacks, acquisitions and debt repayment. It is not earmarked for AI.

Interest is a small burden. Before the issue, trailing 12-month interest expense was $3.2B, 1.6% of operating cash flow of $195.9B.

The money goes mainly to capex and shareholder returns. Trailing 12-month free cash flow was $60.0B, the highest of the four (Alphabet $19.7B, Meta $25.1B, Amazon -$42.4B). Over the same period, dividends were $27.0B and buybacks $21.0B, about $48B to shareholders in total, or 80% of free cash flow. The company has not said why it is borrowing now, and I will not speak for it.

On "hiding." These obligations sit in the notes, and the notes are part of the public filing. They are not in the capex headline because of the accounting rules. A finance lease enters capex when it commences. An operating lease does not, and a lease not yet commenced enters neither. Leases not yet commenced at the four companies: Microsoft $364.6B, Meta $355.7B, Amazon $164.9B, Alphabet $130.5B. All of these figures are in the notes. They are just not in the headline.

A guarantee means you cover the other party's debt if it cannot pay. Meta guarantees two joint ventures that co-build its data centers. I found no similar arrangement in Microsoft's notes, but that does not mean none exists. This bears on whether there are other off-balance-sheet obligations.

So this bill is affordable. The question moves to returns: how much will this compute earn?

7. Valuation: my own model puts fair value at $447–517

▼
A certain depreciation cost on one side of the scale, a self-estimated return on the other

No return figures in four earnings calls. In the base case, FY29 D&A is about $109B, nearly three times FY26's $38.5B, and the fair range is $447–517

No return figures, so I built my own

No one gives a return figure, so I have to calculate it myself. The fair range I get is $447–517.

On each of four earnings calls, an analyst asked about returns, and not once did the company give a number. In January it was Morgan Stanley's Weiss and Bernstein's Moerdler. In April it was Moerdler, in July Borges, and in October Morgan Stanley's Adam Wood. This quarter, Wood asked whether returns come more from higher utilization and lower unit cost, or from charging at more layers. Nadella's answer was "all of the above."

So I ran the numbers myself. In the base case, capex for FY27 to FY29, including new finance leases, is $220B, $245B and $260B. Of that, 70–75% is short-lived equipment, depreciated over six years. By FY29, D&A is about $108.6B, 2.8 times FY26's $38.5B.

This figure is model output, not company guidance. To refute it, you would need to show that cumulative investment is smaller, or that the same assets earn more.

Fair range and the exit multiple

The fair range is the base case discounted to today at the required return, the minimum annual return an investor should demand from this stock. I use 8.98%: the 10-year Treasury yield of 4.98% plus 4 points of risk premium. Every price in the table is a present value, meaning future value discounted to today at the required return. The midpoint is the center of the range.

ZonePriceWhat this zone means
Extremely expensiveAbove $728.78Above the top of the bull case. Even the best case does not earn the required return
Expensive$516.68 to $728.78The base case returns less than required. Only the bull case holds up
Fair$447.22–516.68, midpoint $481.95The base case earns roughly the 8.98% required return
Undervalued$143.68 to $447.22Below the bottom of the base case. The base case returns more than required
Deeply undervaluedBelow $143.68The present value of the bear case. For this number to hold, the company must raise funding and keep costs down

By this table, $521.91 sits just above the top of the fair range, in the "expensive" zone, only about $5 or 1% higher. The "expensive" zone is wide, and $522 sits at its very bottom. I remain bullish because both businesses are accelerating and the premium is small.

All of these numbers rest on one assumption: the exit multiple at the end of FY29, meaning what the market will pay then for each $1 of EPS. The base and bull cases both use 26–30x, applied to FY29 EPS before financing costs: about $21.84 in the base case and about $30.55 in the bull case. The stock now trades at about 25.1x next-12-month earnings. So my assumed multiple is 4% to 20% above today's, and the two EPS measures are not fully the same. This is my assumption, and I have no historical data to verify it. At 24–28x, the top of the fair range falls to $481.95, and $522 is about 8% above it.

Bull and bear cases

The bull case runs like this. Capex for FY27 to FY29 is $230B, $275B and $290B. Compute arrives mostly on time. AI machine utilization rises to 96% by FY28. AI prices barely fall, and Copilot seats and usage keep growing fast. The result is FY29 revenue of about $642B and EPS after financing costs of about $30. The base case gives about $535B and about $21. $521.91 is about 17% below the bottom of the bull case, $631.63. That gap depends entirely on my bull-case assumptions, which have not happened yet, so I do not count it as a reason to be bullish.

The bear case does not need office software to collapse. The problem is the AI investment. Compute arrives slowly. AI utilization falls to 75% in FY28 and 70% in FY29, prices fall another 15% and 12%, and capex still runs $210B, $200B and $185B. FY29 revenue is only about $408B, D&A about $96B, operating margin about 30% and EPS about $13. On my method, cash flow in FY27 and FY28 is -$32.5B and -$4.1B. Cash flow recovers only after the company cuts capex to 22% of revenue. Discounted to today, that is $143.68. To get there, the company must raise funding and must be able to cut commitments it has already signed.

The $9B cost I cannot source

One part of this model I cannot explain, and it also affects whether $522 is expensive. The model carries a cost reserve of about $9B that I added myself. I backed it out so that the full-year operating margin matches the company's guidance of "down less than 1 point." I cannot find what the cost actually is. The candidates are severance, higher memory and storage prices, third-party GPU service fees and operating lease rent, and none of them has a figure. Without it, the base-case fair range becomes $470.52–543.17, midpoint $506.84, and $521.91 falls inside the fair range. Putting $522 in the fair range needs two things at once: removing this cost, and using a 26–30x exit multiple. Remove only the cost and use 24–28x, and the top of the range is $506.84, so $522 is still somewhat expensive.

★ Expert Insight · What you didn't ask but should know

Blind spot 1: The $9B cost may come from July's lease reclassification

Starting in July, part of the future data center leases moved from finance leases to operating leases. A finance lease's cost splits into depreciation and interest, and the interest sits below operating income. An operating lease's rent goes entirely into operating costs and does not enter D&A. The same machines, once reclassified, lower capex and also lower operating margin. As the $364.6B of leases not yet commenced start one by one, this cost will grow further (Inference).

Impact: Reaching the fair range by removing the $9B may not hold up.

Two cash-flow views and the gap to Wall Street

There are two ways to count cash flow, and both figures here are my estimates. On the company's usual method, operating cash flow minus cash capex is positive. I estimate +$20.6B for FY27 and +$37.6B for FY28. The company has never given these two figures. On my method, the number is negative: -$38.6B and -$4.8B in the base case for those two years, turning positive at $14.9B only in FY29. The difference comes mainly from two points. New finance leases need little cash in their first year, but I count all of it as capex that year. I also count stock issued to employees as a cost. So leases are funding part of the investment in these two years.

Revenue explains only part of the gap with Wall Street. Take Wells Fargo. On October 29, it put FY28 EPS at $23.86, against my base case of $20.85. My revenue forecast is lower than theirs. At the same margin, that accounts for $1.35 per share. I cannot attribute the remaining $1.66 to specific items, because public summaries do not include its full financials. The earnings reports will show who is right.

Looking back at April

A look back at the April piece (in Chinese). On April 30 the stock was $407. I was bullish and put a better entry price at $370–385. I got the driver right: Azure growth went from 39% to 48%. I got the framework wrong. I focused on valuation returning to that range, but the stock touched $349.20 on June 25, recovered quickly, and closed at $521.91 on October 30. The price stayed in that zone for only a few weeks, and the change in growth was the more important thing to write about. The $370–385 came mainly from chart levels and the 22.6x valuation at the time. This time the range comes from a bottom-up cash model, so the two numbers do not use the same yardstick. I also got one thing wrong in April. The $190B capex guidance referred to calendar 2026, and I took it as FY26. From that, I inferred June-quarter capex of more than $70B. The actual figure was $41B. Three new things have emerged since April: some leases moved to operating leases, OpenAI passed 10% of revenue, and four sets of numbers disappeared.

8. Conclusion: still bullish, with four things to watch next quarter

▼

Azure and M365 are both accelerating, and $522 is only about 1% above the top of the fair range. I remain bullish

Both businesses are accelerating and the price is close to fair. I remain bullish.

The stock closed at $531.01 on October 28 before earnings and at $521.91 two trading days later (both Yahoo Finance closes), down 1.7%. Azure beat, and the stock fell slightly anyway. My reading is that the market is watching two things outsiders cannot see.

The current price is about 25.1x next-12-month earnings. That price embeds two assumptions: that others besides OpenAI are buying, and that returns on this compute will cover depreciation. The first cannot be tested this quarter. On the second, the company gave no number in four earnings calls.

The second assumption is also the most important item in my model. $447–517 is my base-case calculation. If this compute earns less than I expect, the range has to move down.

Watchlist for next quarter

Next, watch for the missing data. The next earnings report, in January, is the first checkpoint.

  1. First, will the four sets of numbers come back: commercial bookings and RPO excluding OpenAI, Copilot paid seats, and quarterly gigawatts added? If the first two return, demand outside OpenAI becomes measurable. If Copilot seats return, sales of Microsoft's own product become measurable. If gigawatts return, the question of whether compute is enough gets its denominator.
  2. Second, the next earnings report and 10-Q should show how much of the contract base is locked in. Of the current $677B of RPO, the company has not said how much has price and volume locked for the full contract term. The average remaining term of 2.3 years is the time left. The machines depreciate over six years, so they must keep earning after the contracts end.
  3. Third, cloud gross margin once capacity lands in the second half of this fiscal year. It was 66% this quarter, and guidance for next quarter is about 65%. As new capacity comes in, whether gross margin holds 65% or keeps falling is the most direct read on returns. My base case already assumes company-wide gross margin (a different figure from cloud gross margin) falls from about 65% in FY27 to about 55% in FY29. It also assumes operating margin falls from about 46% to about 38%.
  4. Fourth, the token price. This quarter it helped revenue, yet the company's routing strategy is pushing it down. I will watch how Hood talks about pricing next quarter.

Revenue is accelerating, the bill is affordable, and the price is close to fair. I remain bullish.

This article is research analysis by Transurf Research. It is not a recommendation to buy or sell any security. The figures are estimates within an analytical framework, not price forecasts. Investing involves risk, and you may lose your principal.