--- title: "MSFT (Trans): CY26 capex revised to approx. $175 bn" type: "Topics" locale: "en" url: "https://longbridge.com/en/topics/43054173.md" description: "Below is Dolphin Research's compiled $Microsoft(MSFT.US) FY26 Q4 earnings call Trans. I. Core financial highlights recap.1. Shareholder returns: returned $10.2 bn to shareholders this quarter via dividends and buybacks. For the full fiscal year, total shareholder returns exceeded $43.0 bn.2. FY27 Q1 guide: total revenue of $89.85–90.95 bn (+16–17% YoY). COGS of $29.6–29.8 bn." datetime: "2026-07-29T23:25:08.000Z" locales: - [en](https://longbridge.com/en/topics/43054173.md) - [zh-CN](https://longbridge.com/zh-CN/topics/43054173.md) - [zh-HK](https://longbridge.com/zh-HK/topics/43054173.md) author: "[Dolphin Research](https://longbridge.com/en/news/dolphin.md)" --- # MSFT (Trans): CY26 capex revised to approx. $175 bn **Below is Dolphin Research's compiled Trans of** $Microsoft(MSFT.US) **FY26 Q4 earnings call** **I. Key takeaways** 1\. **Shareholder returns**: Returned $10.2bn to shareholders this quarter via dividends and buybacks. For the full fiscal year, total returns exceeded $43bn. 2\. **FY27 Q1 guidance**: Total revenue of $89.85bn–$90.95bn (+16%–17% YoY). COGS of $29.6bn–$29.8bn (+23%–24% YoY); opex of $16.8bn–$16.9bn (+7%–8% YoY); OP margin roughly flat YoY.By segment, Productivity & Biz. Processes revenue of $36.7bn–$37.0bn (+11%–12% YoY), Intelligent Cloud revenue of $40.95bn–$41.25bn (+33%–34% YoY), More Personal Computing revenue of $12.2bn–$12.7bn. Azure revenue to grow ~45% YoY at cc; M365 Commercial cloud adj. cc growth ~16% (reported 15%).Windows OEM and Devices revenue to decline a bit over 20% YoY; Xbox content & services down mid-single digits, hardware down YoY. Ex-OpenAI investment, other income/expense approx. -$0.1bn; ETR ~20%. Capex expected to exceed $50bn (incl. lease reclass impacts from useful-life changes). 3\. **FY27 full-year outlook**: Double-digit growth in revenue and OP at the company level. Opex to grow mid-to-high single digits, OPM down by less than 1ppt YoY, FCF positive, ETR ~20%; capex to rise YoY.M365 commercial products and server products to decline mid-single digits for the year due to a tough comp from last year’s transactional deals tied to product launches. Windows OEM and Devices to fall close to 20% for the year, on weaker PC demand, component cost inflation lifting end prices, a high base from Windows 10 end-of-support, and elevated channel inventory.Assuming FX stays at current levels, FX drags FY revenue growth by less than 1ppt, with no material impact on COGS and opex growth. 4\. **Key financial metrics**: GPM was 67%, down YoY on mix shift to Azure and AI infra investment and usage, partially offset by efficiency gains in Azure and M365 Commercial. OPM edged up YoY to 45%; total headcount fell 2% YoY.**Capex was $41.0bn** (incl. component price increases signaled in guidance), about two-thirds to short-cycle assets (mainly CPUs and GPUs), the rest to long-cycle. Finance leases totaled $5.6bn, mainly for large datacenter sites; cash paid for PP&E was $35.8bn.Operating CF was $55.4bn (+30% YoY); FCF was $19.6bn. Commercial bookings ex-OpenAI rose 18% YoY; including Azure commitments related to OpenAI, up 10% YoY (cc +11%). **Commercial RPO rose 84% YoY to $678.0bn, +25% ex-OpenAI, with a weighted avg. remaining term of 2.3 years**; about 30% will be recognized as revenue in the next 12 months (+37% YoY), and the beyond-12-month portion grew 112% YoY. 5\. **Key policy changes**: From FY27, **extend the estimated useful lives of datacenters and office buildings to 25 years from 15** to reflect operating history and expected asset use; this is embedded in guidance. The change only alters the timing of future depreciation, with a very small benefit to FY27 OP.**The larger impact is on the capex metric** — more future datacenter leases will move from finance leases to operating leases (finance leases count in capex, operating do not). As a result, the CY2026 capex outlook is revised to approx. $175.0bn, but excluding useful-life effects, the actual investment plan for 2026 is unchanged. 6\. **One-offs this quarter**: vs. Apr guidance, several one-time items contributed a net +$0.27 to diluted EPS, including a $3.2bn gain from the Anthropic investment and lower-than-expected voluntary retirement costs, partly offset by severance and Xbox impairment.**Ex these items, revenue, OP and EPS all beat**. EPS was $4.74; ex-OpenAI investment effects, EPS rose 23% YoY; ex-OpenAI, other income/expense was $2.8bn, mainly the Anthropic gain. **II. Detailed call notes** **2.1 Management highlights** **1\. AI platform and infrastructure** a. Added 31 datacenters across five continents this quarter, 88 for the fiscal year, to meet accelerating demand. Ramp-to-live is faster too, cutting the time from GPU racking to power-on by nearly 50% in the largest regions over the past fiscal year.**Added 1GW of capacity this quarter and remain on track to roughly double total capacity in two years.** b. Driving more output from existing infra via cross-layer optimization across silicon, systems, and software. For example, Copilot-type workloads saw 4x throughput since the start of the year. c. AI sovereignty is a growing customer focus. Announced a partnership with Mistral to bring its models to Microsoft Sovereign Cloud, enabling runs in public cloud, customer-controlled, and fully disconnected environments. d. Modernizing the fleet in parallel with in-house silicon and the latest from NVIDIA and AMD. Maia 200 capacity continues to scale, delivering 30% better performance per dollar vs. the newest fleet hardware, and now serves both OpenAI and MAI models.**Microsoft will be among the first cloud providers to deploy next-gen at-scale AI infra based on AMD Helios and NVIDIA Vera Rubin.** e. CPUs are as critical as GPUs when running agents. Cobalt VMs now support both first-party and external workloads, serving Adobe, ARM, Elastic, OpenAI, Sprinklr, TomTom, among others.**Expect Cobalt 200 racks to be deployed in 25+ datacenters globally by month-end.** **2\. Model layer and Foundry** a. Model choice is foundational. Microsoft offers the broadest model catalog in the cloud — 11,000+ models spanning OpenAI, Anthropic, Mistral, xAI, and in-house MAI; YTD, customers building with multi-vendor models are up 5x.Levi Strauss & Co uses both OpenAI and Anthropic on Foundry, consolidating 1,000+ domain-specific agents into a single enterprise AI platform. b. In-house model dev is accelerating, with 10+ new models this quarter across vision, speech, transcription, coding, and security, including the first reasoning model, MAI Thinking 1, all designed for low-cost inference in enterprise scenarios. Co-designed with in-house silicon, MAI models on Maia 200 deliver 40% better performance per watt. c. More importantly, Microsoft is building a new model system that decouples harness, context, memory, and action space from any single model family. This shifts the frontier of the cost-output curve, adding business continuity and resilience because each model is swappable. d. This system is already proven in first-party products. Millions of developers use MAI Code 1 Flash in GitHub Copilot with higher code acceptance and a 10% lower median token count, while still invoking frontier capabilities from OpenAI and Anthropic.In Excel, MAI Code 1 Flash matches GPT-5.6 on common tasks at materially lower cost; in security, MAI Cyber 1 Flash with multi-agent security harness outperforms much larger models at half the cost. More broadly, Dynamics 365 cut GPU costs by 89% with MAI Voice 2 Flash, and PowerPoint cut up to 84% with MAI Image 2.5.This system is available to all enterprises as part of Foundry. e. Foundry is being built as a full app-and-agent stack, offering IQ layers, tools, persistent state and memory, secure sandboxes, rubrics and evals, and even self-improvement loops for agents. Foundry now serves 100k customers with revenue more than doubling YoY.Telefonica standardized on Foundry as its platform base, with initial agents for critical network ops. The number of Foundry customers running at a 1 trillion tokens annualized rate is up 4x YoY. f. Agent 365 provides a control plane extending existing governance, identity, security, and management frameworks to customer-built agents. In just two months, nearly 40mn agents have been registered across tens of thousands of enterprises. **3\. Enterprise data and context** a. Data assets are evolving from supporting human-operated apps to serving agents. Customers are accelerating adoption of AI-optimized databases like Cosmos DB and SQL to give agents fast, secure access to real-time data and context for memory and retrieval.PostgreSQL revenue grew 55% YoY, an accelerating trend for three straight quarters; PostgreSQL customers using Foundry rose 80%. New Horizon DB is a fully managed PostgreSQL service on Azure with 3x the throughput of self-managed deployments. b. **On analytics, Fabric now has 40,000+ paying customers, up 60%+ YoY; over 17,000 customers now use both Foundry and Fabric, also up 60%.** This quarter brought Rayfin — an agent-first SDK offering backend-as-a-service for apps on Fabric, already used by 2,500+ customers. c. Building an IQ layer atop data assets to blend data and model capabilities and deliver the right context at the right time. Tens of thousands of customers, including nearly 90% of the Fortune 500, are using Foundry, Fabric, and Work IQ to inject enterprise context into their agents.Web IQ launched this quarter to give agents real-world intel from across the web and is already used by leading AI assistants, including ChatGPT. **4\. Knowledge work and M365 Copilot** a. Paid M365 Copilot seats now exceed 30mn, with net adds more than doubling QoQ. Copilot is moving rapidly from chat to co-work to autopilot; co-work entered GA last month to execute multi-step tasks on enterprise data under security and compliance, and autopilots launched this quarter as autonomous, long-running, enterprise-compliant agents, including an always-on personal agent powered by OpenClob.These Copilot experiences, including code, are being integrated into a super app spanning consumer and commercial scenarios. b. Quality and performance continue to improve. User satisfaction doubled over the past three quarters to a record high, and latency fell another 25% this quarter.These gains, plus product innovation, drove record engagement — conversations per user nearly doubled YoY, and weekly actives are now on par with Outlook and Teams. Time from deployment to high usage (approx. 80% MAU within the employee base) fell from months to days over the past year. c. Customers with 50,000+ seats grew 7x+ YoY; enterprises deploying Copilot to most knowledge workers rose nearly 75% QoQ. NHS England is rolling out Copilot to 505,000 clinical and support staff, the largest of its kind in healthcare, after pilots showed an average 43 minutes saved per day.KPMG expanded to 276,000+ professionals globally; HSBC committed to 200,000 seats; AstraZeneca, Boeing, Infosys, Coke Inc, P&G, Stellantis, Tata Consultancy Services, UPMC, and Wells Fargo each bought 60,000+ seats. d. The new E7 suite (bundling Copilot, E5, Entra, and Agent 365) is seeing strong traction. Within two months of launch, hundreds of enterprises purchased millions of seats.EY deployed E7 to 400,000 employees this quarter, the largest single win to date. e. **The biz. model is evolving from per-seat to per-seat plus usage, expanding TAM.** Earlier this month, usage-based billing was added for co-work; thousands of customers are already paying and actively using it. **5\. Biz Apps** a. Dynamics 365 is being re-architected for agent-first, opening 650,000+ MCP actions across sales, finance, supply chain, HR, and customer service. Agents can access biz. context and execute actions under the same data models, rules, permissions, safeguards, and audit trails as app users. b. The model is also shifting from seats to seats plus usage. Customer service leads this transition, with usage-based credit consumption up 4x QoQ; clients like Northern Trust use these tools to drive proactive intelligent service. **6\. Developers and GitHub** a. GitHub Copilot users reached 50mn. Usage-based billing was introduced this quarter; commercial and enterprise seats continue to grow, and consumption revenue also rose significantly under the new model.Copilot revenue accelerated by over 60% QoQ. b. GitHub overall users reached 225mn, with 90%+ of the Fortune 500 choosing GitHub for AI-driven development. The agent era is being built on GitHub — one in three pull requests on the platform now involves agents. **7\. Security** a. Microsoft is both protecting customers’ AI deployments and using AI to strengthen their security posture. Purview has audited 50bn+ Copilot interactions to date to meet compliance, up nearly 360% YoY. b. Launched Project Perception this week, a full multimodal agentized security system where coordinated agents simulate attacks, investigate threats, and drive remediation. After private preview, Perception will be offered on a usage-based basis. **8\. Healthcare and science** a. In healthcare, more than 100mn patient encounter summaries are on track to be automated this calendar year, including 28mn this quarter, up 2x YoY. Mass General Brigham rolled out Dragon Copilot to 4,000+ clinicians; a study shows Ambient AI reduced burnout by 21%. b. In science, Microsoft Discovery is now broadly available, providing a complete platform to build and govern agent workflows for scientific and engineering use cases. Early customers include BHP, GSK, and Pacific Northwest National Laboratory. **9\. Delivery organization: Microsoft Frontier Co** a. The most comprehensive and valuable data sits within each customer’s tenant, creating a major opportunity to turn workflows, domain knowledge, and accumulated judgment into AI systems that learn and improve with use. Microsoft launched Frontier Co this month, the industry’s largest outcome-oriented engineering org, deploying 6,000 industry and engineering experts to co-design, co-innovate, and continuously improve AI systems with customers. b. This model completed 330+ projects across 164 customers over the past year. Examples include building an agent with Novo Nordisk to analyze clinical data under strict compliance, and embedding AI into LSEG Workspace to help finance pros quickly answer complex questions across structured and unstructured financial content. **10\. Devices and consumer** a. Xbox: making necessary choices across content, platform, and operations to reset for long-term growth. Microsoft holds top-tier IP and world-class studios and expects the biz. to return to growth in FY27. b. Windows: investing to ensure quality and core functionality while making it the best host for secure edge AI. Microsoft sees a major opportunity for Windows as the substrate for unmetered intelligence, combining strong on-device compute with enterprise-grade security. c. Search and ads: Bing and Edge have gained share for five straight years. LinkedIn engagement remains strong, with membership growing double digits for a fifth year; recruiters at 20,000+ companies use its AI solutions to shorten hiring cycles and improve matching, with seats up 140% QoQ. **2.2 Q&A** **Q: How big can adoption of open and custom models get over the next 1–2 years, given initial enterprise hesitation around open models, and how does Microsoft benefit given its exposure to frontier labs?** A: Our premise is that enterprises should control their destiny — building their own human capital and token capital. If an enterprise is a learning machine, it needs its own learning machine.Models are inputs, not extractions of enterprise knowledge. Every company will assess which vendors truly drive outcomes and help create knowledge, and that is already clearer by the day.Based on that trajectory, **our platform architecture is explicitly designed to separate the harness from the model**. The harness externalizes your memory and context, making any given model swappable at any time. Use frontier models where they shine and low-cost models where they suffice; you can even train your own if you prefer, because outputs, traces, and context remain in your hands.That is the design pattern we promote and use ourselves — in Copilot, GitHub Copilot, and Security Copilot. Within this framework, you’ll see a mix of open-weight and closed-weight models. The recent Hugging Face incident reminds us not to depend on any single model; you may need multiple models to remediate issues caused by one, so the design space is large.When we say ‘frontier,’ each enterprise should own its own frontier, with choice, cost control, and capability. Complementing that, Azure is architected to deliver the right model for the right task efficiently, regardless of model family or whether customers run their own models. **Q: Azure growth accelerated from 43% to ~45%. What drove it — are we still capacity-constrained, or executing better under constraints?** A: **Constraints persist**. Demand continues to exceed available supply, which is visible even in spot-market pricing of relevant assets.We focused first on efficiency — squeezing more output from the fleet across CPUs and GPUs. Engineering did strong work this quarter to unlock that, and given the demand-supply imbalance, efficiency gains monetize quickly in-quarter.We also streamlined lead times from CPU/GPU arrival to production, with further improvements over the past 90 days. At our hyperscale, such quickly monetizable gains supported the step-up in growth, consistent with our Q1 commentary. **Q: If overbuild and excess capacity emerge in datacenters and chips, how does Microsoft protect itself; conversely, how do you avoid steep price hikes or margin damage amid hardware and component inflation?** A: The two are related. Today, demand is well above available supply; over a longer horizon, note that **a large share of spend, especially in capex, has shifted toward short-cycle assets — CPUs and GPUs with shorter lead times**.If demand changes, you slow that largest component, which also drives COGS. Land and datacenter construction are more flexible and a smaller part of the cost structure, with timing levers, especially build pace; you can also stagger planned GPU/CPU ramps.Hyperscalers have long managed such demand shifts. We also benefit from a highly diversified book across regions, customer tiers, and industries, plus a large first-party apps business that consumes capacity.On pricing, the environment affects all vendors. Our job is to keep driving efficiency to deliver value; cloud still offers strong ROI vs. on-prem server purchases, where price inflation hits customers harder. As we add capacity, much is sold via new contracts that can reflect pricing; long term, pricing must work for customers and for us, and we will stay focused there. **Q: M365 Copilot surpassed 30mn paid seats with clear acceleration. How are pilots scaling to broad deployments, and what will drive monetization from here — new seats, mix shift to higher-value SKUs like E7, or consumption?** A: The product form factor evolved meaningfully within the quarter — chat, co-work, autopilot, and code — all converging into a flagship super app for different roles. Time from license purchase to active use has collapsed from months to days, and engagement is now on par with Outlook and Teams.Second, the enterprise ‘wiring’ matters. E7 brings Agent 365 governance so SecOps and FinOps are integrated, and co-work connects to CRM and ERP as skills and plugins, pulling enterprise workflows into the super app, which compounds usage.Third, the model is now seats plus usage. ARPU uplift from E5 plus Copilot is visible, with E7 to add more; E7’s strongest value is Agent 365 with SecOps and FinOps. Everyone will need observability and control over token spend across workflows, which E7 provides, expanding TAM via usage and consumption as IT leans in. **Q: With new frontier models and the launch of Project Perception, how is the security landscape changing, and what does this mean for trust in Microsoft?** A: The physics of security products and SecOps are changing. **We are applying the same ‘intelligence-first, model-forward’ approach used in knowledge work and coding**.Perception ensures you have a red team agent continually probing, a blue team agent triaging, and a green team driving fixes — creating a continuously running agentized defense. It fuses identity (Entra), Defender, network, and app-security signals into context to generate effective protection.A multi-model approach is critical for cost and resilience. In Cybergym with MDASH data, MAI Cyber 1 Flash achieved comparable performance at 50% lower cost by handling 90% of tasks, escalating 10% to frontier models. If any model is unavailable, operations continue, which is essential across coding, security, and knowledge work. **Q: Linking capex and monetization, how do current ROI thresholds compare to a year ago, and what incremental levers, including in-house silicon, can drive monetization?** A: My math hasn’t changed materially. What matters is confidence in TAM expansion and the margin levers we control across products and infrastructure.We still see opportunities for best price-performance in silicon, including first-party. Model diversification itself is a margin lever — delivering optimal outcomes more efficiently, in token use and cost structure.All of this raises confidence in ROIC on invested capital. Our resource pool spans knowledge work, coding, security, and the agent layer (Agent 365), plus Azure-side model, silicon, and component efficiency and hyperscale operating leverage.We will keep pushing these levers with steady, disciplined execution and deliver the gains to customers. **Risk disclosure and disclaimer:**[**Dolphin Research Disclaimer & General Disclosure**](https://support.longbridge.global/topics/misc/dolphin-disclaimer) ### Related Stocks - [MSFT.US](https://longbridge.com/en/quote/MSFT.US.md) - [002245.CN](https://longbridge.com/en/quote/002245.CN.md) - [OpenAI.NA](https://longbridge.com/en/quote/OpenAI.NA.md)