Microsoft's Efficiency Story Has a Dark Subtext
Microsoft's efficiency gains are real — the question is who keeps them, and the answer is almost certainly not the workers or the firms that flow through Microsoft's pipes.
Microsoft just proved that AI-driven efficiency works. The proof should worry you.
The Stratechery read on Microsoft’s latest earnings is that three things lined up cleanly: a clear strategy, falling costs, and actual customer applications that generate revenue. Ben Thompson calls that combination compelling. He also notes the reason it happened is scarier. He’s right, but he undersells the scariness — so let me try.
The standard corporate AI story runs like this: you invest heavily in the infrastructure layer, revenue lags by a few quarters, then the productivity gains compound and the margins expand. Investors price the lag charitably because the eventual payoff looks enormous. That story is still directionally correct. Microsoft’s numbers just showed it playing out at the top of the stack — Copilot is selling, Azure AI revenue is growing, and operating leverage is returning. The machine is working.
Here’s what the machine is doing to get there. Headcount is declining not through discrete layoff events but through a slower, less visible process: roles that used to require a human are now handled by systems that don’t show up on a payroll line. Microsoft doesn’t have to fire people loudly. It just doesn’t backfill. The cost structure improves quarter over quarter, the efficiency ratios look great on a slide, and the labor market absorbs the attrition so gradually that nobody calls it a crisis. Ben Thompson’s word for this dynamic is “scarier.” The right word is “structural.”
This is meaningfully different from the automation cycles that came before. When spreadsheets replaced bookkeepers, the displaced workers moved into roles that spreadsheets created — financial analysts, FP&A teams, a whole profession of people who use the tools. When ERP systems replaced data-entry clerks, the replacement roles were at least in the same building. The current cycle is compressing at multiple levels simultaneously: the tool that does the task and the manager who used to supervise the person doing the task are both getting squeezed. Microsoft doesn’t just need fewer junior knowledge workers. It may be discovering it needs fewer senior ones too.
Everyone says this is temporary — that AI creates new categories of work faster than it destroys old ones. The opposite is closer to true on the relevant time horizon. The new categories exist. They’re just a lot smaller. One prompt engineer doesn’t replace ten writers; one AI-augmented finance team doesn’t replace the three teams it consolidated. The productivity gains are real. They accumulate at the firm level. They don’t cleanly distribute to workers.
Goldman’s credit team flagged something adjacent this week: software borrowers specifically face growing maturity stress, and private credit is starting to show dispersion. That’s the other side of the Microsoft efficiency story. The firms that aren’t Microsoft — that bought SaaS subscriptions for years, that built headcount on the assumption that software plus labor was the only formula — are now learning that the company selling them the AI is also optimizing away the labor that used to sit next to the software. The cost curve for enterprise software goes up (more AI seats, more tokens consumed) while the staffing models get leaner. That’s margin compression from both sides for any company that isn’t at the center of the stack.
Microsoft sits at the center. The Copilot bundle, Teams, Azure, GitHub — these are the pipes through which the efficiency gains flow. The firms flowing through those pipes get the efficiency, but Microsoft captures an increasing share of the value. This is platform power operating at its most elegant: you sell someone the tool that makes their organization more productive, and in doing so you make yourself more indispensable to every organization that uses you. The DMA-forced iPhone-to-Windows clipboard feature is a footnote, but it’s the same dynamic at a smaller scale — Apple opens a door in the EU because it has to, and Microsoft walks through it.
The stakes are simple: we’re watching the first real-world, large-scale demonstration that AI-native cost structures beat non-AI-native ones, and Microsoft is the benchmark. Every major enterprise will now be measured against that benchmark. The ones who can’t close the gap will face pressure — from boards, from activists, from credit markets — to get there faster than the technology actually permits.
The efficiency is real. The question is who keeps it.