The loudest part of the generative-AI story is capability — bigger models, new benchmarks, smarter demos. The quieter part, the one that shows up in a cloud provider's margins, is what it costs to actually run those models for millions of requests. In the week ending 18 May 2026, Amazon (AMZN) had a run of patents issue that sits almost entirely on that second axis. A granted claim is enforceable coverage, not a hope, so this cluster is a map of positions Amazon has locked in around the operational economics of large language models (LLMs): the latency, the tool-routing, the guardrails, and the training efficiency that determine whether running a model at scale is profitable.

The clearest example is US12626695B1, a grant on cache techniques for LLM processing. Its abstract states the problem and the fix in plain terms:

Within a dialog session, a portion of the LLM prompt may be the same across dialog turns, and instead of recomputing the attention/encodings for such portions, the cached encodings can be used by the LLM during processing.— Cache techniques for large language model processing, US12626695B1

That is a direct attack on inference cost. In a multi-turn conversation, the system prompt and earlier context get re-encoded on every turn unless you cache them; caching the encodings means the expensive attention computation is not repeated. Multiply that across an Alexa-scale or AWS-customer-scale request volume and the saved compute is real money. The grant is coverage on a specific technique for not paying twice for the same work.

A cluster around the cost and reliability of generation

The rest of the set rounds out the same theme from different directions. US12626698B1 covers component shortlisting — identifying which APIs or tools are relevant for an LLM prompt and including only a ranked subset in the prompt. That is both a cost lever (shorter prompts, fewer tokens) and a capability lever for tool-using agents, where stuffing every available tool into context is wasteful and degrades accuracy. On the reliability side, US12626692B2 covers moderating the responses of a generative language model by matching user input to a policy and constraining the output, and US12626691B1 covers mitigating hallucination using contrastive decoding — generating logits for a plain prompt, a context-augmented prompt, and an adversarial prompt, then combining them. Output moderation and hallucination control are exactly the features an enterprise buyer demands before it will put a model in production, which makes them commercial coverage, not just research.

There is a coherent commercial logic to why a cloud provider, specifically, would amass coverage here rather than chasing a flagship model. A model lab competes on capability; a cloud provider competes on the cost of delivering capability at scale, because its customers are billed for tokens and compute and will migrate to whoever serves the same quality cheaper. Every technique in this cluster either shaves the compute per request or removes a reason an enterprise would hesitate to deploy — which is to say each one maps onto a line a cloud business can actually charge for or protect. US12626158B1, an automated contribution-analysis grant that picks the dimensions explaining a metric for natural-language analytics questions, extends the same logic into the data-product layer, where AI is sold as an answer rather than a model.

Two further grants reach into the training and analytics layers. US12626137B1 covers gradient exchange across processing nodes arranged as a hyper-rectangle, using scatter-reduce and all-gather operations — the kind of distributed-training communication scheme that determines how efficiently a large model trains across many accelerators. And US12626128B2 covers continual machine learning in a provider network, with user-configurable retraining and hyperparameter tuning offered as a managed service. That last one is telling: it is a patent on selling AI operations as a cloud feature, which is the shape of how Amazon monetizes the rest of this stack.

What the coverage maps, and the limits

Taken together, the week's grants describe Amazon claiming ground on the layer between a model and a paying customer: serve it cheaply (caching, shortlisting), serve it safely (moderation, hallucination control), train it efficiently (gradient exchange), and sell the upkeep (continual learning as a service). For a company whose AI business is fundamentally about renting infrastructure rather than selling a flagship model, that is coverage aligned with where its revenue actually comes from. Issued claims here are positions a competing cloud or platform working the same inference-optimization and guardrail techniques would have to navigate around or license.

The limits are the usual ones for a grant cluster. Enforceable coverage is not the same as a deployed feature or a disclosed revenue line — these patents describe methods, and the records do not say how widely Amazon uses each one or what it earns from them. Many of these techniques are also actively researched across the industry, so coverage on a particular implementation does not foreclose alternative approaches to the same goal. And patent counts in any single week swing with examiner timing, so the cluster is best read as a snapshot of a sustained investment in inference economics rather than a sudden strategic turn. What the week shows as fact is consistent and specific: in the same seven days, Amazon had grants issue across prompt caching, tool shortlisting, output moderation, hallucination mitigation, distributed-training communication, and managed retraining — a coordinated footprint on the cost-and-reliability plumbing of generative AI, which is precisely the part of the model business a cloud provider gets paid for.