Thirty-one U.S. patents were issued to Microsoft on March 31, 2026, per the week's grant records. Read as a set, the most enforceable subject matter is not foundation-model design. It is the operational layer: detecting and diagnosing failures, holding power inside a budget, keeping latency within tenant expectations, and putting guardrails around how model data is analyzed. This is the work of running AI services, and it is where this week's coverage sits.

The clearest example is US12591480B2, "Automatic identification of root cause and mitigation steps for incidents generated in an incident management system." The grant describes clustering related incident records, ranking them, and prompting an AI model to return a root cause and mitigation steps. The mechanics are stated plainly.

A prompt is generated to an artificial intelligence (AI) model based on the ranked, related incidents and the AI model returns a response that identifies a root cause and mitigation steps corresponding to the ranked incidents.— Automatic identification of root cause and mitigation steps for incidents, US12591480B2

This is AIOps subject matter — using an AI model to operate infrastructure. It pairs with US12592875B2, which covers detecting cloud-service latency issues by building per-tenant baseline behavior and flagging deviations, then aggregating unhealthy tenants to judge whether a service is degraded. Both describe how a hyperscaler keeps a large, multi-tenant service healthy, and both are now issued claims.

Power as a first-class constraint

Two grants treat power as something to be allocated and capped, the recurring pressure point in AI data centers. US12591448B2, "Server-side control of power based on service-level agreement," describes throttling accelerators in an overprovisioned rack according to SLA-defined workload priorities, distributing power to individual components and throttling in a defined order until consumption falls under a policy limit. US12591292B2 covers configurable local frequency throttling to manage a processor's power demand, dynamically reducing clock frequency in response to a throttle event. Coverage that ties power distribution to contractual service tiers describes exactly the trade-off operators face when accelerator demand outruns available power in a rack.

Productizing the model layer

A third group attaches to the model-and-language layer as a product. US12591707B2, "Privacy preserving insights and distillation of large language model backed experiences," describes storing user prompts in a privacy-protecting datastore from which they cannot be read directly, then deriving clustered theme summaries that move to a separate, queryable store. US12591752B2 covers zero-shot domain transfer for a text-to-text model, training on unlabeled in-domain text plus labeled general-domain tasks to reach specialized domains such as radiology or legal text where labeled data is scarce. And US12591738B2 covers autocorrect candidate selection for a writing assistant, deciding which NLP suggestions to apply automatically versus surface for approval.

The breadth of the batch extends beyond operations. US12591797B2 covers a "geometrically enhanced Clifford quantum computer" method for reducing qubit count, a reminder that the same assignee files across quantum, storage, and graphics in any given week. But the center of gravity on March 31 is operational: keeping AI services up, inside power budgets, within latency expectations, and behind privacy walls.

For a business reader, the distribution is the takeaway, with the usual caveat that one week of grants is a narrow sample and a single batch is not a portfolio. The coverage indicates that what Microsoft locked in this week is concentrated in the discipline of operating AI and cloud at scale — the layer that sits between a trained model and a paying tenant. That is the layer where a hyperscaler's cost, reliability, and contractual commitments are actually decided, and the March 31 claims map onto it directly.