CompaniesMicrosoft

Microsoft

Thesis

  • The verdict: entrenched enterprise data and workflow platform.

  • The business: a software annuity with a capital-intensive utility bolted on.

  • Moat: distribution and integration; operating leverage

    • Enterprise distribution.
      • Contractual and administrative lock-in (productivity, identity & security, developer tools, databases).
      • Global footprint and data locality.
    • Integration at data/application/workflow level.
      • OpenAI IP.
      • Power and grid interconnection.
      • Chip and fab allocation.

  • Earnings drivers: Cloud growth, enterprise software upsell.

    • Azure consumption. M365 upsell. Search advertising, LinkedIn and Dynamics.
      • OpenAI-like margin with OpenAI IP through 2032, even with premium-priced compute.
    • Price and mix growth on an installed seat base.
    • Operating leverage from a sales force as a shrinking share of revenue.
    • $678B backlog.

  • Challenges

    • Need to pay a premium to neoclouds/SpaceX for compute after the "Great Pause" 2024-25.
    • Chip/system level integration and cost savings is inferior to Amazon/Google.

  • What to watch: margin, cash flow, backlog

    • Cloud growth margin.
    • Free cash flow conversion.
    • Data center leases signed but not yet commenced.

  • AI strategy:

    • Leverage OpenAI IP to scale the high margins until 2032.
    • Massive catch-up on Compute after the 2024-25 "Great Pause", including renting at a premium from neoclouds/SpaceX.
    • Vertical integration and cross-sell to maintain high ROIC. Chips (ASIC), systems, IaaS (bare-metal GPU cluster), PaaS (token factory, Azure Foundry), LLMs, applications
    • Enterprise distribution: relationships, data and workflow lock-in, multiple customer touch points - identity, security, cloud, applications, AI, workflow, collaboration.
    • Global footprint. Leverage its broad geographical footprint and bring AI closer to enterprise customers ("fungible fleet" strategy), catering to their needs of high security, data locality laws, co-processing with non-AI workloads.
      • In contrast with frontier labs who care more about capacity than latency / locality to the user for post-training and inference workloads, and can place datacenters anywhere possible and serve global traffic.
      • The cost to pay: datacenter site selection becomes more complex and constrained.
    • Application: Office 365 CoPilot: OpenAI IP and Office user data.
    • Commoditize models. Let enterprises have their own model for their own data, without sharing with the frontier labs. The real test is the performance trade-off vs being on the cutting edge.
    • AI hardware platform (Project Solana).
    • Developer lock-in.

  • AI weakness and risks:

    • Model capability. Coding moat (GitHub CoPilot) under siege -> ramping up their model supermarket ecosystem bet (Agent HQ) and internal model development (MAI models).
    • Opportunity cost: Microsoft needs to balance Azure — both for its enterprise customers and OpenAI — and its software business.
    • Startup customers. Not a significant player for managed clusters or on-demand VM, given its significant gaps in ease of use, monitoring, reliability and health checks.

Reports

Reference

Built with LogoFlowershow