Microsoft’s AI Cost-to-Outcome Strategy: Cheaper Inference, Swappable Models and Usage Pricing

What the letter actually says​

Nadella writes that inference is making marginal cost an important factor in software economics as usage scales, so customers will increasingly measure the cost of tokens against the outcomes they produce. Microsoft calls this the "cost-to-outcome" frontier and says it is focused on optimizing it.

Inference is the computing a trained model does when it answers a prompt or completes a task. Traditional software costs little to serve once it is built. AI is different: every request burns compute, so heavier use means a bigger bill. That is why the cost of a result matters now in a way it rarely did for older software.

The letter, dated October 21, 2026, was reported as an annual shareholder letter. Its ecosystem pitch is that Microsoft is building a "frontier ecosystem" that lets others build more value on and through it than Microsoft captures inside. Nadella says this is how Microsoft creates broad economic surplus and delivers durable returns for investors.

The technical mechanism: separating the harness from the model​

The most practical idea in the letter is architectural. Nadella describes a multimodel system in which the harness, meaning context, memory and action space, is separate from the model. He says this lets Microsoft use the right model for the right job.

On the July earnings call, Nadella said much the same thing to analysts. In the call transcript, he says the harness should stay separate from the model so that memory and context sit outside it and any given model can be swapped. He also says firms can use frontier models, lower-cost models, or train their own.

For IT teams, this has some practical consequences:

• Model choice becomes a cost lever. Not every task needs the most expensive model.
• Memory and context stay with the customer. The harness, not the model, holds them, which limits lock-in to any one model vendor. That limit applies to the model, not necessarily to the platform.
• Resilience improves in theory. Interchangeable models mean one provider's problem need not stop your workflows. Benzinga's report of the letter says Nadella argues this strengthens business continuity because models become interchangeable.

The numbers behind the pitch​

Microsoft's own figures, from the earnings call transcript on its investor site, show the scale:

• Revenue for the fiscal year was above $331 billion, with growth accelerating to 18%. Operating income grew 21% to more than $155 billion.
• Microsoft Cloud revenue passed $214 billion for the year. The letter coverage puts it at $214.4 billion, up 27%. The call also said nearly 90% of that came from customers outside frontier model companies.
• Net paid Copilot seat additions more than doubled sequentially, and paid Microsoft 365 Copilot seats now exceed 30 million.
• Azure growth was reported at 43% for the fourth quarter on the call. The 41% full-year figure and the $100 billion milestone come from the letter coverage.
• Superpowerdaily's account of the letter adds that GitHub Copilot had 50 million users, and Microsoft Foundry passed 100,000 customers with more than 11,000 models in the catalog.

These are company-reported numbers. They show adoption and sales. They do not show whether a given customer's AI project paid off.

Infrastructure: the physical cost of cheaper AI​

Cheaper tokens need capacity, and capacity needs buildings and chips. On the call, Microsoft said it added 31 datacenters across five continents in the quarter, bringing the fiscal-year total to 88. It also said it cut dock-to-live times for new GPUs in its largest regions by nearly 50% over the year. It added another gigawatt of capacity in the quarter and expects to roughly double overall capacity in two years.

The call also gave the efficiency details behind "cheaper":

• Microsoft said it is seeing 40% better performance per watt when running its own MAI models on its Maia 200 silicon.
• It said it announced more than a dozen new models, including its first reasoning model, MAI Thinking 1, designed around cost-efficient inference.
• The CFO said efficiency gains in the CPU and GPU fleets were monetized quickly in the quarter, because customer demand continues to exceed available capacity.

Costs are not all falling, though. The CFO said capital expenditures were $41 billion for the quarter and mentioned higher component pricing. She also said Microsoft is extending the estimated useful life of its datacenters and office buildings from 15 to 25 years from the start of fiscal 2027. The company said that changes the timing of future depreciation and shifts more datacenter leases from finance to operating leases, which brings its calendar 2026 capex expectation to about $175 billion. That is a gap between "cheaper to run" and "cheap to build".

Pricing models are changing too​

Cost-to-outcome is not only an engineering story. It also shows up in how Microsoft charges:

• Dynamics 365: Microsoft said it is moving from seats to a seats-plus-consumption model, and is exposing over 650,000 MCP actions across business functions for agents.
• GitHub Copilot: the CFO cited a June business-model change that aligns pricing with usage, and said consumption was stronger than expected afterward. She also said the change improved margins through the quarter.
• Microsoft 365: the company referred to adding usage-based billing products alongside per-seat licensing in July. The CFO also described E7, with its Agent 365 component, as offering observability and manageability of token spend.

Admins should read that as a signal. Budgeting for AI looks less like buying licences and more like managing a metered utility, so FinOps habits matter.

A reality check​

The letter is a shareholder document, so it is a strategic framing, not independent evidence. A few cautions are worth keeping in mind:

• Adoption is not outcome. Seat counts and cloud revenue measure sales. They do not measure time saved, accuracy or revenue created.
• Customer evidence is selective. The call cited NHS England rolling out Copilot to 505,000 clinicians and staff after a trial that showed an average saving of 43 minutes per day. That is a vendor-cited example, not a general result.
• "Cheaper" competes with "more." Lower unit costs can still produce larger bills if usage and agent activity expand faster. That is general industry reasoning, not something Microsoft states.
• Ecosystem is also strategy for Microsoft. A broader platform means more workloads on Azure. That does not make the offer bad for customers, but it is worth noticing who benefits.

What IT leaders can do with this​

Treat the cost-to-outcome idea as a test you run yourself:

1. Define the workflow. Pick a specific task and a baseline for time, error rate or cost.
2. Measure real usage. Track token or consumption costs at production volumes, not pilot volumes.
3. Compare models. Try a cheaper or smaller model against a frontier one for the same task.
4. Include the hidden costs. Count integration, human review, security and compliance.
5. Keep context portable. Where possible, hold memory and data outside any single model, which is the architecture Microsoft itself describes.
6. Watch the billing model. Check whether seat-plus-usage pricing changes your forecast as your agents do more work.

Bottom line​

Nadella's letter ties together three things: cheaper inference, a swappable model architecture, and a platform others build on. The supporting figures are large, and they come from Microsoft. Whether the economics work for your organisation depends on measuring outcomes in your own workflows.