Microsoft Shifts Gears: Azure +40% for Fifth Straight Quarter, $190B Capex, and the OpenAI Restructuring That's Repricing the Stock

OurAlpha Scorecard

Composite 7.1/10 — a legitimate core long position within a Mag 7 portfolio, but not to be valued under the old "exclusive OpenAI proxy" framework.

I. Q3 FY26 Earnings: Another Beat, but Capex Rattles the Market

Microsoft Q3 FY26 (quarter ended 3/31/2026) results^{[Microsoft IR]}:

Q4 FY26 Guidance^{[Investing.com Earnings Call Transcript]}:

• Total Revenue $86.7B–$87.8B (+13–15%)
• Intelligent Cloud $37.95–38.25B (+27–28%)
• Productivity & BP $37.0–37.3B (+12–13%)
• More Personal Computing $11.75–12.25B
• Azure CC growth guidance maintained at +39–40% (essentially "hold the current pace")

Segment Breakdown:

Intelligent Cloud cost of revenue +47% (CC +44%)^{[Alphastreet]} — this is the capex/depreciation pass-through in action. Revenue +30%, cost of revenue +47% — over time, Intelligent Cloud gross margins face structural compression. This was the number sell-side couldn't stop talking about the week of May 16.

OurAlpha Take: Revenue, EPS, and Azure all beat across the board, yet the market's verdict was a −4–5% close on 4/30^{[24/7 Wall St.]} — driven entirely by capex shock (see Section 6). This is a significant signal: the market is repricing MSFT — shifting its valuation framework from "AI beneficiary" to "AI-heavy capital infrastructure company."

II. Azure +40% and the Shifting AI Disclosure Framework

Q3 FY26 results: Azure and other cloud services growth of +40% (+39% CC)^{[Microsoft IR]}. Note that this reflects Microsoft's reported segment — Azure alone is not broken out; it is bundled with other cloud services.

Recent quarterly growth trajectory (CC basis, per public filings):

By any recent baseline, this is a strong growth rate for a hyperscaler. Hood reiterated on the earnings call that compute capacity remains tight through the full year 2026^{[Futurum Analysis]}.

A notable shift in disclosure framing: in prior quarters, Microsoft more consistently disclosed "AI services contributed X points to Azure growth" as a percentage-point breakout. In Q3 FY26, the narrative centerpiece shifted to the $37B AI run-rate. This is not necessarily a negative signal, but it reduces investor visibility into the AI vs. non-AI split within Azure growth.

A note on the $37B AI run-rate figure: this is a management-defined annualized revenue run-rate for AI, which per management commentary broadly encompasses Azure AI customers, frontier model company workloads, and Microsoft's first-party AI products and services — but Microsoft has not provided a fully audited segment breakdown^{[Microsoft Official X]}. It should not be treated as a standalone segment revenue figure; investors should be cautious about comparability and the gap between run-rate headline growth and underlying margin quality.

Azure supply/demand dynamics:

• Still capacity-constrained — Hood again flagged compute tightness through all of calendar 2026^{[Futurum Analysis]}
• ~1 GW of net new data center capacity added in Q3, with management guiding to a doubling of global physical footprint within two years^{[Global Datacenter Hub]}
• $11.1B in data center lease spend in Q1 FY26 alone^[DCD] — an OpEx figure on top of capex, illustrating that Azure is pursuing capacity through both owned build-out and third-party leases in parallel

OurAlpha Take: The hard numbers underpinning the Azure thesis remain intact — +40% / CC +39% with demand outstripping supply is a textbook demand-pull signal. The disclosure reframing looks more like Microsoft consolidating its messaging around the $37B run-rate as the primary investor narrative anchor. Near-term, however, the AI vs. non-AI boundary within Azure becomes fuzzier, and valuation frameworks will need to be recalibrated accordingly.

III. Copilot: The Real Structure Behind 20 Million Paid Seats

Key disclosed metrics^{[Microsoft IR][Microsoft Official X]}:

• Microsoft 365 Copilot paid seats exceed 20 million
• Accenture surpasses 740,000 seats (largest single deployment)
• Customers with 50,000+ seats grew 4x YoY
• Several large enterprises (Bayer / Johnson & Johnson / Mercedes-Benz / Roche, among others) have deployed at significant scale

GitHub Copilot (last official update: Q2 FY26 / January 2026)^{[Motley Fool]}:

• ~4.7 million paid subscribers (+75% YoY)
• Transitioning to token-based billing effective June 1, 2026^{[GitHub Blog]} — a shift to usage-based monetization, tying future ARR directly to consumption

A note on Copilot ARR disclosure: Microsoft does not break out Copilot ARR separately — it rolls into the $37B AI run-rate figure. Back-of-the-envelope math at $30/user/month list price puts 20 million seats at $7.2B annualized list-value; however, enterprise volume discounts, bundling, phased enablement, and actual seat utilization rates likely put realized ARR meaningfully below list-value.

What Copilot's numbers are actually signaling:

1. Enterprise deployments are scaling up: 4x growth in 50,000+ seat customers, plus Accenture's 740K — enterprise adoption is entering "org-wide" territory
2. 20 million → 50 million is the next order of magnitude: Whether MSFT reaches 40 million by mid-FY27 and when it hits the 50 million pricing inflection are key variables for the dual-engine of ARPU expansion and seat growth
3. Agent 365 GA launched May 1 ($15/user/month)^{[Microsoft Security Blog]} — the next SKU layer above Copilot, enabling enterprise users to orchestrate AI agents; AWS Bedrock / Google Cloud registry integration in preview — MSFT is deliberately positioning Agent 365 as a cross-cloud product
4. M365 E7 SKU at $99/user/month^[SAMexpert] — the new top-tier SKU launched in May, and the linchpin for ARPU upside

OurAlpha take: Copilot is the most conviction-carrying application-layer asset in MSFT's current AI cycle. The question isn't whether it grows — it's that the lack of disaggregation within the AI run-rate limits the market's ability to independently value Copilot. Once investors want to value it as a standalone SaaS business (à la ServiceNow on P/S), more disclosure becomes necessary. That's the pressure management will face over the next 2–3 earnings cycles.

IV. The OpenAI Relationship, Overhauled: From "Exclusive Distribution" to "Primary Cloud + Azure-First + Non-Exclusive IP + Capped Revenue Share"

This is the most significant structural shift in the MSFT story for 2026. Two key inflection points:

2025/10/28 (OpenAI Restructuring)^[CNBC]:

• OpenAI split into OpenAI Group PBC (for-profit) + OpenAI Foundation (nonprofit)
• Based on a $500B OpenAI valuation, MSFT received approximately 27% economic interest in OpenAI Group PBC (implying roughly $135B at that valuation — note: this is an economic interest implied by the valuation, not a directly consolidable equity asset in the conventional sense, and should not be read as a cash equivalent)^{[Tom's Hardware]}
• Foundation holds 26%; employees + investors hold 47%

2026/4/27 (Second Amendment)^{[OpenAI Official]}:

• IP license locked through 2032, and converted from exclusive to non-exclusive
• The new agreement eliminates or substantially weakens the AGI-trigger uncertainty. The prior deal raised market concerns that an AGI declaration by OpenAI could materially alter Microsoft's access rights, revenue share, and collaboration boundaries; the new agreement fixes Microsoft's IP license to OpenAI models and products through 2032 — though that license is now non-exclusive^{[The Decoder]}
• Azure's exclusive distribution rights are gone — OpenAI may now deliver products through any cloud provider; Microsoft nonetheless remains OpenAI's primary cloud partner — per Microsoft's official statement, "OpenAI products will ship first on Azure, unless Microsoft cannot and chooses not to support the necessary capabilities"^{[Directions on Microsoft]}

The revenue-share direction requires careful distinction (this is the most commonly misread section):

• Microsoft's official statement: Microsoft will no longer pay a revenue share to OpenAI — i.e., Microsoft stops paying OpenAI a revenue share^{[OpenAI Official]}
• Per Reuters: OpenAI's 20% revenue share payable to Microsoft remains in place through 2030, but with an undisclosed cap — once the cap is hit, the obligation falls away^[Reuters]

The catalyst: the 2026/2 AWS–OpenAI mega-deal^{[Nerd Level Tech]}:

• AWS committed up to $50B in OpenAI investment
• AWS became the exclusive third-party cloud for OpenAI's "Frontier" enterprise platform
• This is the first concrete manifestation of MSFT–OpenAI exclusivity unwinding

MSFT's countermove — absorbing the Stargate assets OpenAI walked away from:

• Stargate Norway: MSFT stepped in after OpenAI/Oracle exited; MSFT is leasing 30,000 NVIDIA Vera Rubin chips from Nscale on a 5-year contract, targeting 100,000 GPUs total; builds on MSFT's prior $6.2B commitment^[Bloomberg]
• Stargate Texas: MSFT took over after Oracle/OpenAI dropped out^[Bloomberg]
• According to media reports, OpenAI has significantly scaled back and reprioritized its long-term infrastructure spending plans (early market estimates ran to $1T+ cumulatively through 2030; the latest reported figure has been revised down to ~$600B)^[CNBC] — this is the backdrop for MSFT picking up portions of the Stargate footprint

4/2: MSFT unveiled three in-house foundation models^[TechCrunch] — a clear signal that MSFT is formally transitioning from "OpenAI reseller" to "independent model provider."

OurAlpha's take: This restructuring is not a one-sided negative for Microsoft — it does compress the exclusive-distribution premium, but it simultaneously reduces the tail risk from OpenAI being a single-contract dependency:

1. **A 27% economic interest (implying ~$135B at the $500B valuation)** offers better long-term IRR than "perpetual exclusivity" — if OpenAI's valuation doubles, the implied value of MSFT's economic interest doubles with it; that said, this is not a directly consolidable operating asset
2. Removing AGI-trigger uncertainty: under the old deal, an AGI declaration posed tail risk to Microsoft's access rights, revenue share, and collaboration scope; the new agreement locks the IP license through 2032, substantially reducing that risk (though the license is now non-exclusive)
3. Azure-first priority on new model launches: per Microsoft's official statement, OpenAI products still ship on Azure first, unless Microsoft is unable or unwilling to support them — this preserves a "main-artery AI" position, but one that requires Microsoft to actively exercise it
4. Revenue-share dynamics reset: Microsoft no longer pays OpenAI; OpenAI still pays Microsoft 20% (capped through 2030) — a "settlement path" both sides accepted

Downside risks:

• OpenAI can now compete directly against MSFT for the same enterprise customers on AWS — both AWS and MSFT sell OpenAI, so enterprises may simply go where pricing is lower
• Azure-first is still preferential, not exclusive: if Azure capacity falls short, OpenAI can contractually route workloads to AWS — the real value of that priority depends on Microsoft's capacity investment and commercial commitment
• The exclusive-distribution premium is gone: the simple "OpenAI = Azure" valuation narrative is over, and the market needs a new anchor

五、Maia 200：MSFT 的 ASIC 反击战

Maia 200^{[Microsoft Blog]} is MSFT's second-generation in-house AI inference chip, officially launched on January 26, 2026:

MSFT self-reported performance (vendor presentation figures, not independently benchmarked)^{[Microsoft EMEA Source]}:

• 30% better performance-per-dollar vs. Maia 100 — a generation-over-generation comparison within MSFT's own fleet
• FP4 compute roughly 3x AWS Trainium 3 (MSFT self-reported)^{[Tom's Hardware]}
• FP8 compute exceeds Google TPU v7 (Ironwood) (MSFT self-reported)

Caveat: The cross-vendor comparisons above are MSFT's own presentation data with no independent MLPerf-style benchmarks, and the figures span different precision levels, system configurations, and single-chip contexts. These numbers illustrate Microsoft's ASIC ambitions; they do not establish that Maia 200 comprehensively outperforms Trainium or TPU.

Why Maia 200 matters for the industry:

1. The ASIC counterstrike: MSFT is one of the most NVDA-dependent buyers among the Mag 7 (see the ASIC siege analysis in the NVDA deep dive). Maia 200 is the centerpiece of MSFT's effort to reduce that dependency.
2. The inference battleground: Maia 200 is explicitly positioned as an inference accelerator — it does not challenge NVDA's training dominance, instead targeting inference, the natural home turf for ASICs.
3. A vertical integration proof point: If the next generation of large-model inference actually ramps on Maia 200, it signals that MSFT has closed its first real loop across model, silicon, and data center.
4. Phoenix is just the start: Iowa + Phoenix is the opening move; Europe and Asia-Pacific deployments are expected post-2027 — which would meaningfully erode NVDA's GPU procurement share within Azure data centers over time.

OurAlpha's take: Maia 200 is not about "replacing NVDA" — it's about giving MSFT leverage at the negotiating table with NVDA. Near-term, roughly two-thirds of 2026 capex still goes to GPU procurement. The medium-term question is whether the Maia line can capture a meaningful share of Azure inference workloads and genuinely reduce unit inference costs — that is the key metric for assessing MSFT's capital efficiency. The specific "20–30%" share figure cited for Maia's slice of the Azure inference pool is a scenario assumption (Microsoft has not disclosed inference pool breakdowns), but the directional implication is clear: NVDA's pricing power with one of its largest customers faces structural compression — one concrete realization of the customer-concentration risk for NVDA.

VI. Capex Guidance Repricing: How Calendar 2026 ~$190B Blew Up the Valuation

Calendar 2026 capex guidance of ~$190B (+61% YoY vs. 2025)^[CNBC]——the figure Hood put out on the 4/29 earnings call.

A critical definitional note: $190B is on a calendar year 2026 basis, not Microsoft's fiscal year——MSFT's fiscal year ends June 30, so $190B actually spans FY26 H2 (Jan–Jun 2026) + FY27 H1 (Jul–Dec 2026). All comparisons in this article use the "calendar 2026" convention.

Disclosed figures:

• Q3 FY26 capex: $31.9B
• Q4 FY26 capex guidance: >$40B
• vs. Street consensus of ~$154.6B → ~$35B above expectations^{[Global Datacenter Hub]}

Capex increase breakdown (per management):

• ~$25B from memory/component price inflation——same capacity, higher cost (HBM3e/HBM4 scarcity, TSMC 3nm pricing, etc.)
• Remaining ~$10B from actual capacity expansion
• Roughly 2/3 of total capex going into GPU/CPU silicon, with 1/3 into shell, land, power, and cooling

Why MSFT sold off 4–5% on Q3 earnings day^{[24/7 Wall St.]}:

1. Looming depreciation drag (rough estimate): Amortized over a 5–7 year useful life, $190B in AI infrastructure investment will create sustained depreciation pressure through FY27–FY30. That said, you can't simply divide $190B by 6 and call it annual incremental D&A——capex is deployed on a rolling basis, asset classes vary (land, shell, equipment, GPUs, networking, leases), and both deployment cadence and depreciable lives differ across categories.
2. The market is starting to demand AI ROIC accountability: In 2024–2025, investors were willing to take capex on faith. In 2026, the question has shifted to "what revenue does every $1 of capex generate?" Microsoft's answer——"Azure +40% and a $37B AI revenue run rate"——hasn't been paired with an explicit ROIC figure from management.
3. Capital return compression: Q3 shareholder returns (dividends + buybacks) came in at $10.2B^{[Microsoft Source]}——annualizing to roughly $40B, a fraction of the $190B calendar 2026 capex.

A non-standard but intuitive ratio——"capital returns / capex" (note: this is not a standard Wall Street metric; it's an OurAlpha construct used purely to illustrate the directional shift in capital allocation):

• FY25: capital returns ~$36B, capex ~$80B → ratio ~45%
• FY26: capital returns ~$40B (est.), calendar 2026 capex ~$190B → ratio drops to ~21%

This is the flip side of MSFT's AI cycle——the front side is the Azure +40% story; the back side is cash flow being reinvested rather than returned to shareholders. That shift will prompt some dividend-oriented and GARP investors to reassess their positions.

OurAlpha take: ~$190B in calendar 2026 capex is not, in itself, the problem——MSFT's cash generation can absorb it. The real question is the pace of ROIC realization:

• Bull case: Azure sustains +40% growth, Copilot accelerates toward 50 million seats, and the Maia series meaningfully cuts inference unit costs——FY28 EPS growth stays robust.
• Bear case: Azure growth decelerates on tough comps, OpenAI's multi-cloud strategy dilutes Microsoft's share of AI workloads, and depreciation erodes EPS——FY28 EPS growth could slip to the low-to-mid teens.

Three upcoming events will each reset the market's expectations for calendar 2026 capex returns: NVDA earnings on 5/20, Q4 FY26 results in late July, and Q1 FY27 results in late October.

VII. Valuation & What's Priced In: $569 Consensus, 22x Forward P/E, $3.13T Market Cap

Current data (as of market close 5/15/2026)^{[Stock Analysis]}:

Where MSFT sits on the Mag 7 valuation spectrum:

• On StockAnalysis's current methodology, MSFT's forward P/E is roughly 22x — below the high-20s range the market typically assigned Microsoft in recent years, and below some of the higher-growth Mag 7 names (note that forward EPS definitions vary across data platforms, so comparisons require care)
• +35% upside to consensus target + Strong Buy rating — the Street is meaningfully more bullish than the current price implies

What the consensus already prices in:

• Azure growth sustaining +35–40% through FY27
• Copilot seats continuing to double
• The post-restructuring 27% economic interest in OpenAI + Azure-first rights holding the line on model exclusivity
• Maia 200 ramping on schedule, giving Microsoft pricing leverage over NVDA starting in 2027
• $190B in capex translating into ROIC expansion by FY28–FY29

Downside risks not yet priced in (any one of these could derate the multiple by 10%+):

• Azure growth hitting the inflection point on an "acceleration → plateau → deceleration" curve (five consecutive quarters of re-acceleration makes a slowdown increasingly likely)
• Copilot per-seat ARPU coming in below expectations after enterprise volume discounts
• OpenAI competing on price directly against Azure on AWS and Google Cloud, compressing enterprise OpenAI ASPs
• Maia 200 volume ramp or yield falling short, leaving NVDA pricing power intact
• US or EU antitrust scrutiny targeting the Microsoft–OpenAI relationship or Copilot bundling
• $190B capex guidance revised further upward in 2027, breaching the market's ROIC tolerance threshold

OurAlpha's take: A ~22x forward P/E for a company with Azure CC +39%, a 27% economic interest in OpenAI, and 20 million Copilot seats is not obviously stretched. The real question isn't the multiple itself — it's the EPS trajectory:

• If FY27 EPS continues climbing at the FY26 pace, the $569 consensus target is achievable
• If depreciation and component costs continue weighing on EPS growth, the multiple compresses back toward 18–20x

OurAlpha's simplified back-of-envelope (not a sell-side model): 22x × $25 FY28 EPS (bull case consensus) ≈ $550; if FY28 EPS only reaches $22 (depreciation-drag scenario), you'd need a 26x multiple to support $570 — which requires the market to assign MSFT an "AI operating system" premium. This is OurAlpha's own simplified reverse-engineering, not a standardized sell-side price target methodology.

Eight. The Next 3 Quarters: 5 Key Metrics That Could Reprice the Stock

5 actionable metrics:

1. Q4 FY26 Earnings (late July): Can Azure & other cloud services CC growth hold at +39%——guidance is +39–40%; anything below +38% will be read as "the acceleration is over"
2. Copilot seat disclosure cadence——which quarter the 30M / 40M milestone lands, plus whether management provides an ARPU / ARR breakdown — the key inputs for pricing the Copilot business
3. Quality of Azure-first fulfillment after OpenAI's restructuring: Does OpenAI's next flagship model still launch on Azure first? Any simultaneous launch on AWS Bedrock / Google Cloud would materially dilute the Azure-first value proposition
4. Maia's share of Azure inference workloads and per-unit inference cost trajectory——post Iowa + Phoenix, the ramp pace in 2H 2026 / 1H 2027 is the key watch item; whether MSFT publicly discloses that share is itself a signal
5. Calendar 2027 capex guidance——the number the market is most sensitive to. Another large upward revision will pressure valuation again; a meaningful reduction will be read as "the AI cycle is cooling" and also hits valuation. Only "modest upside + accompanying ROIC signals" will be viewed as a positive

九、反向风险：什么情况会让这个判断翻车

• Valuation re-rating risk: a rotation from "AI beneficiary" to "AI heavy-capex infrastructure company" — the latter naturally commands lower multiples, amplifying volatility during the transition
• OpenAI relationship further loosening: the 4/27 revision is only an interim state; if OpenAI completes a fully independent IPO in the 2028–2030 window, MSFT's "model supply advantage" enters a wind-down path
• Copilot ARPU shortfall: meaningful gap between the list price on 20 million seats and actual post-discount ARPU leaves substantial room for disappointment
• Maia 200 volume ramp bottleneck: TSMC 3nm capacity and HBM3e supply remain tight; any yield or delivery issues will delay the window to renegotiate with NVDA
• Broad hyperscaler capex pullback: if any one of the Big Five signals a definitive spend tightening in 2027, the entire AI sector faces a multiple reset
• Antitrust risk: the Microsoft–OpenAI relationship plus Copilot bundling remain under active scrutiny from U.S. and EU regulators; any escalation becomes a near-term short catalyst
• More Personal Computing sustained weakness: Q3 −1% — Windows OEM and Xbox Hardware softness, though a small slice of revenue (16%), signals that the consumer AI PC upgrade cycle has yet to materialize
• AI safety regulation taking effect: on 5/5 MSFT, Google, and xAI agreed to submit unreleased models for U.S. government pre-screening^[CNN] — a short-term PR positive, but a long-term constraint on model release cadence

X. Betting on Three Things: What You're Really Buying When You Buy MSFT

Buying MSFT is no longer a play on "exclusive OpenAI distribution." At its core, buying MSFT today is a bet on three things:

1. Azure remains the primary gateway to the AI OS layer — even as OpenAI moves toward multi-cloud, Azure's enterprise stickiness, Copilot bundling, and Microsoft 365 installed base keep it positioned as the de facto OS for the AI application tier
2. ARPU expansion across Copilot + Agent 365 + M365 E7 — per-user/per-seat revenue growing from ~$30/month to ~$80/month over the next three years (inclusive of Agent 365 + E7 + Copilot), the cleanest growth engine in the story
3. A second-leg re-rating on MSFT's 27% economic interest in OpenAI PBC — if the $500B → $1T+ valuation path materializes, that stake's implied value rises from $135B to $270B, representing pure "non-operating" value unlock (note: this is an economic interest, not a directly consolidatable operating asset)

If you believe all three → $421 is a reasonable-to-cheap entry; if you're starting to doubt the first (Azure as primary gateway) → trim your position; if you're starting to doubt the second (Copilot ARPU) → MSFT multiples compress back to 18–20x.

MSFT is the most complete story in the Mag 7 — simultaneously cloud, model, application, Agent, silicon, OS, hardware, and gaming. Q3 FY26 earnings told the market two things: fundamentals are still accelerating, but the valuation framework is shifting.

The scorecard: ARPU + ROIC.

Additional Notes on Data Definitions

• Share price and valuation: $421.92 / forward P/E 22.12x as of the U.S. close on 2026-05-15; figures change in real time
• AI run-rate of $37B is management's definition, encompassing Azure AI + OpenAI revenue share + Copilot + GitHub/Dynamics AI; no public methodology document exists
• Standalone Copilot ARR has no audited segment disclosure — the $4.0–5.5B estimate in this article is based on sell-side assumptions about enterprise volume discounts
• The $135B implied value of MSFT's 27% OpenAI stake corresponds to a $500B OpenAI valuation; MSFT's 10-Q accounts for this using the equity/cost method, not at $135B on the face of the balance sheet
• $190B capex is a calendar 2026 figure (covering FY26 H2 + FY27 H1) and does not equal FY26 full-year capex
• Maia 200 self-reported benchmarks (vs. Trainium 3 / TPU v7) are from MSFT's own presentations; no independent third-party benchmarks exist
• All "OurAlpha perspective / opinion / scenario estimates" sections in this article represent editorial judgment, not investment advice

This content is for informational purposes only and does not constitute investment advice, trading advice, or any guarantee of returns.