Microsoft (Minutes): Maintains Capex guidance unchanged, supply bottlenecks remain in 4Q
Below is the $Microsoft(MSFT.US) Minutes of Q3 FY2025 Earnings Call. For earnings analysis, see Azure Rises Again: Microsoft Regains AI Leadership
- Key Earnings Highlights:
II. Detailed Earnings Call Content
2.1 Core Management Statements:
1. This quarter, driven by continued strong performance of Microsoft Cloud, revenue exceeded $42 billion, up 22% in constant currency. Cloud and AI are core enablers for every enterprise to expand output, reduce costs, and accelerate growth.
2. On infrastructure, we continue expanding data center capacity. This quarter alone, we launched data centers in 10 countries across 4 continents. Thanks to Scaling Law, model performance doubles every 6 months. We optimize every layer from data center design, hardware/chips, system software to model efficiency, aiming to reduce costs while improving performance. New GPU delivery cycles shortened by nearly 20%, AI performance improved ~30% at same power, and per-token costs more than halved.
3. For cloud migration, we see accelerating demand across industries—from Abercrombie in France to Coca-Cola and ServiceNow expanding Azure footprints. With broader regional availability than other hyperscalers, we remain the preferred cloud platform for critical VMware, SAP and Oracle workloads.
4. In data & analytics, we deeply integrated AI platform with data stack. PostgreSQL usage accelerated for third straight quarter, now adopted by nearly 60% of Fortune 500. Cosmos DB revenue growth re-accelerated, remaining the top global distributed NoSQL database for clients like CarMax, DocuSign, NTT DATA and OpenAI.
5. Analytics consumption accelerated. Microsoft Fabric now has 21,000+ paying customers, up 80% YoY. Fabric unifies data warehousing, data science, real-time analytics and Power BI into an end-to-end solution. Real-time intelligence is Fabric’s fastest-growing workload, with 40% adoption just 5 months post-launch.
6. AI platform & tools. Foundry, as an AI app factory agent, is now used by 70,000+ enterprises like Atomic Work, Epic, Fujitsu, Gainsight, H&R Block and LG Electronics to design, customize and manage AI apps/agents. We processed over 100 trillion tokens this quarter (5x YoY), hitting a record 50 trillion last month alone. The new Agent Service, launched 4 months ago, already has 10,000+ organizations building/deploying agents.
7. Developer tools. GitHub Copilot users surpassed 15 million (4x YoY). Digital natives like Twilio and enterprises like Cisco, HPE, Skyscanner and Target continue adopting Copilot to AI-power developers end-to-end.
8. Microsoft 365 Copilot, designed for human-AI collaboration, now serves hundreds of thousands of global cross-industry clients (3x YoY). Overall deal sizes keep growing. This quarter saw record customers adding more seats. Last week’s major update integrated Agent, Notebooks, Search and Creation into a new work framework. The new Researcher and Analyst deep-reasoning agents can analyze vast web/enterprise data, delivering expert insights on-demand within Copilot.
9. In business apps, Dynamics 365 gained share as Avaya, Brunswick and Softcat switched from legacy vendors. Example: Verizon chose Dynamics 365 Sales to boost sales team productivity.
10. For Windows, Copilot+ PCs are faster with better battery than peers, offering richer AI apps from Adobe, Canva and Zoom. Last week, we rolled out exclusive AI experiences like Recall, Click to Do and Windows Search to all Copilot+ PCs. With Windows 10 support ending, commercial adoption grew ~75% YoY for Windows 11.
11. Security. Last month, we launched Security Copilot Agent with partners to help defenders autonomously handle high-volume security/IT tasks using 84 trillion daily threat signals. We also added Defender, Entra and Purview features to secure/govern AI deployments. Security clients reached 1.4 million, with 900,000+ (e.g., EY Global, ManpowerGroup, TriNet, Regions Bank) using 4+ workloads (up 21% YoY). Entra monthly active users exceeded 900 million.
12. In ads, we’re transforming search/browsing/content discovery with AI assistants. Bing’s Copilot Search reshapes results with AI-curated overviews and embedded chats. Copilot Vision and Edge enable real-time browsing responses. Copilot Discover personalizes MSN via user interactions. The updated Copilot app boosts daily engagement via multimodal convos/search/shopping/travel. Overall, we gained Bing/Edge share, surpassing $20B ad revenue past 12 months.
2.2 Financial Commentary & Next-Quarter Guidance:
1. Capex: Expect Q4 sequential growth. Our Jan H2 FY25 guidance stands. Note cloud infra buildouts/financing leases may cause quarterly volatility.
2. FY26 capex outlook: Committed to demand-driven investments. Reiterating prior comments—expect growth decelerating vs FY25, with higher short-term asset mix (more directly revenue-linked). These investments and execution focus (delivering near-term client value) will maintain our cloud/AI leadership.
3. Commercial remaining performance obligation rose to $315B, up 34%/33% (reported/constant currency). ~40% will convert to revenue in 12 months (up 17%). The remainder (beyond 12 months) grew 47%. This quarter’s annuity mix was 98%.
4. OpEx grew 2% (3% constant currency), below expectations due to cost discipline and some investment shifts to Q4. We continue streamlining via reduced management layers/headcount for agility.
5. Microsoft 365 Commercial Cloud revenue grew 12%, again driven by E5 and Copilot ARPU expansion. Paid commercial seats exceeded 430 million (up 7% YoY).
6. Azure + other cloud services revenue grew 33% (35% constant currency), with AI services contributing 16 points. Enterprise client growth accelerated, with some large-scale deployment improvements. On Azure AI, we onboarded capacity faster than expected.
7. Search/news ad revenue ex-TAC grew 21% (23% constant currency), well above expectations on partner traffic, better rates, and Bing/Edge share gains.
8. Q4 outlook: Below is segment guidance (detailed growth rates omitted—focus on incremental info not in tables).
Through April, demand signals remained stable across commercial, LinkedIn, gaming and search. Our outlook assumes these trends continue in Q4. Changes could impact results. For Windows OEM, we expect Q3’s high inventory levels to decline in Q4. We widened guidance ranges for some More Personal Computing segments reflecting this uncertainty.
On commercial bookings: We expect solid growth despite tougher YoY comps, though the base keeps expanding. Note larger/longer Azure contracts’ timing is less predictable, potentially increasing quarterly booking volatility.
Microsoft 365 Commercial Cloud revenue should grow ~14% constant currency, stable sequentially. We expect continued ARPU growth from E5/Copilot, with seat growth slowing given installed base size.
For Azure, we guide 34%-35% constant currency growth, driven by strong demand across services. In non-AI, focused execution should sustain healthy growth. For AI services, while we’re onboarding capacity as planned, demand is slightly outpacing. Some AI capacity constraints may emerge post-June.
Search/news ads ex-TAC should grow high teens YoY, even against tough comps. We expect continued traffic and revenue-per-search gains, with Bing/Edge share growth. Total search/news ad growth likely mid-teens.
COGS guided at $23.6B-$23.8B (up 19%-20% constant currency); OpEx $18.0B-$18.1B (~5% growth). Thus, even with AI investments, we expect FY25 operating margins to slightly expand YoY.
2.3 Q&A
Q: Media reports mentioned data center commitment changes, even rumors of canceled projects. But AI demand seems insatiable. Can you discuss strategic adjustments?
A: We’ve always adapted construction/leasing pace over 10-15 years—just more scrutinized now. Key is aligning builds/leases to future workload growth, considering demand, workload types and geographies. You don’t want regional imbalances or inflexibility when demand shifts, especially as training methods and compute needs evolve.
While Moore’s Law helps, it’s a composite S-curve with chip advances, system software, model architectures and app efficiencies intertwined. We ensure builds incorporate the latest insights.
I’m pleased with our pace. We’ll face power shortages—we need specific locations with capacity to lease/build at desired speeds. Lead times range from 2-7 years. We aim to balance demand with Satya’s priorities. We hoped to achieve equilibrium by Q4-end. As seen this quarter, demand rose, so supply will remain tight through year-end, but we’re encouraged.
Q: On accelerating cloud migration—can you elaborate?
A: Three concurrent trends interact: 1) Classic migrations (SQL/Windows Server) remain steady, potentially re-accelerating due to cloud efficiency. 2) Strong data growth—Azure Postgres/Cosmos DB, Fabric analytics, plus Databricks/Snowflake on Azure. 3) Healthy core compute consumption by cloud-native players pre-AI, sustaining through the quarter.
Note proportions: e.g., ChatGPT uses AI accelerators plus Cosmos DB, Postgres, core compute/storage. Any AI workload has fixed ratios between AI and non-AI resources.
Four interrelated aspects. One certainty: our largest business is infrastructure. The next platform shift builds on existing foundations—not a rebuild. Azure grows robustly while new platforms depend on it. We’ll stay disciplined.
Q: How would today’s Microsoft fare in a recession vs. past downturns? Would revenue impacts be shallower?
A: In any turbulence, we’d focus on helping clients through cloud efficiency, our SaaS-to-infra stack breadth, and software’s adaptability to inflation/growth pressures. We’d seek shared benefits while delivering client value.
Q: AI drove 16 points of Azure growth—can you break this down beyond capacity additions? Which workloads (e.g., ChatGPT inference vs. enterprise adoption)? Could June quarter’s AI contribution exceed 16 points?
A: Azure’s beat mainly came from non-AI. AI outperformed due to earlier capacity releases, but don’t expect much upside vs prior guidance (given Q4 bottlenecks). The positive was pre-supplying some clients, but non-AI drove the surprise.
Q: Past comments suggested higher capital efficiency could decouple capex growth from Azure acceleration. Updates?
A: Historically, cloud transitions saw capex surges filling CPU capacity, with software/hardware efficiencies emerging later. For AI, this is happening faster, compounded by model diversity and efficiency gains. What’s different is velocity—both efficiency gains and buildout speed. Teams are compressing deployment timelines. Current AI margins are better than server-to-cloud transitions at this stage.
Q: “DeepSeek moments” suggest software efficiencies may extend GPU lifespans. How does this affect AI experimentation? Is GPU lifespan reassessment premature?
A: OpenAI’s inference optimizations (test-time compute + pretraining) demonstrate validated efficiency leaps. Each Moore’s Law iteration can yield 10x gains via software—seen in model architectures, data/compute efficiency etc.
This mirrors virtualization’s evolution (servers → client-server → cloud → AI). The more software drives efficiency, the more demand follows.
On asset depreciation: we’d need longer histories before adjusting assumptions. We maximize asset lives, but this hinges more on software than hardware.
Q: Azure’s non-AI beat—what’s different vs. past optimization headwinds?
A: Broad-based outperformance, but mainly accelerated growth in our largest client cohort (enterprises) and improved large-scale deployment execution (after Q2 challenges). Regional trends were stable—good migration work, sales/partner execution, and data workloads. No singular factor beyond improved execution, though large-scale deployments still need work.
Q: How does Azure AI’s growth trajectory interplay with non-AI?
A: Distinctions blur as digital natives build AI/non-AI workloads in the same cloud. Over the past 2.5-3 years, capex acceleration transparently converted to revenue/products. Workloads share GPUs, CPUs, storage, networking. As AI workloads expand, this integration tightens.
My focus: Azure (both AI/non-AI) executed well in Q3—field/partner teams, backlog conversion, workload value-add. Less about bifurcation, more about holistic execution.
<End>
Disclosures & Disclaimers:Dolphin Research Disclosures
