Nvidia · Filed Jan 30, 2025 · Published Jul 30, 2026 · verified — real USPTO data

Nvidia Patents a Two-Stage System That Lets AI Chatbots Run Faster on Less Hardware

Every time a large language model generates text, it fills up memory with a growing table of notes about what it has already read. Nvidia's new patent attacks that memory problem with a two-stage filtering system that keeps only the notes that matter.

Nvidia Patent: Smarter KV Cache Memory for AI Models — figure from US 2026/0220041 A1
Figure from the official USPTO publication.
See all 8 drawings from this filing ↓
Publication number US 2026/0220041 A1
Applicant NVIDIA Corporation
Filing date Jan 30, 2025
Publication date Jul 30, 2026
Inventors Payman Behnam, Yaosheng Fu, Ritchie Zhao, Po-An Tsai, Zhiding Yu
CPC classification 706/15
Grant likelihood Medium
Examiner CENTRAL, DOCKET (Art Unit OPAP)
Status Docketed New Case - Ready for Examination (Mar 1, 2025)
Document 26 claims

What Nvidia's AI memory-trimming system actually does

Imagine you're reading a very long book and taking notes as you go. By chapter 50, your notepad is completely full, and flipping back through all those notes just to write the next sentence slows you down enormously. That's roughly what happens inside an AI language model when it's generating a long response: it maintains a running record (called a KV cache) of everything it has processed, and that record eats up memory fast.

Nvidia's patent describes a two-step approach to keeping that notepad manageable. First, a rough-cut step throws out notes that are unlikely to be useful again. Then a finer-detail step quickly scans what's left and picks only the most relevant entries to pay attention to for each new word generated.

The goal is to let AI models handle longer documents and conversations without needing as much expensive memory, and to do it without retraining the model from scratch. It's the kind of unglamorous engineering work that determines whether a model runs well on real hardware.

How the two-stage eviction and attention process works

When a large language model (LLM) generates text, it stores a key-value (KV) cache, a running memory table that records compressed representations of every token (word or word-fragment) it has already processed. Accessing that cache is one of the biggest bottlenecks in AI inference (the process of actually running a trained model to produce output).

Nvidia's patent introduces a KV management process with two consecutive stages:

  • KV cache eviction (coarse-grain): An algorithm with an adaptive pooling size decides which cached entries to discard entirely before the fine-grain step runs. Think of it as bulk-deleting old notes that are statistically unlikely to influence the next output.
  • Hybrid attention (fine-grain): This stage groups the remaining cache pages, stores their minimum and maximum values as quick reference summaries, then accumulates query data to find the most relevant entries. Finally it performs sparse attention, meaning it only fetches and processes the top-K most relevant KV pairs rather than every single one.

Critically, this is described as training-free: the technique is applied at inference time, so it doesn't require re-training an existing model. That makes it much easier to drop into a production deployment.

What this means for running large AI models on tighter hardware

The KV cache memory problem gets worse as conversations and documents get longer, which is exactly the direction AI applications are heading. Systems that can summarize a contract, analyze a codebase, or hold a multi-hour conversation need to process hundreds of thousands of tokens without running out of memory or slowing to a crawl. A method that reduces memory bandwidth and capacity use without sacrificing accuracy directly affects what hardware can run a given model and at what speed.

For Nvidia, which sells both the GPUs that run these models and the inference software stack around them, owning the IP on memory-efficiency techniques is a natural extension of its AI infrastructure business. If this approach lands in Nvidia's TensorRT-LLM or related tooling, it could let the same GPU serve more users simultaneously, which translates directly to lower cost per query for anyone deploying AI at scale.

Editorial take

This is deep infrastructure work, not a flashy demo, but it addresses a real and pressing constraint in production AI deployments. The training-free angle is the most practically useful part: operators don't want to retrain models to get better hardware efficiency. Whether Nvidia's specific two-stage approach becomes a standard technique or gets superseded is an open question, but the problem it's solving is very real.

The drawings

8 drawing sheets from US 2026/0220041 A1 · click any drawing to enlarge

Patent filing page

Which company should we read for you?

We track 17 companies here. Pro is the same weekly breakdown for any company you choose, delivered privately. Type a name and we'll scope it and send you a quote.

Get one Big Tech patent every Sunday

Plain English, intelligent commentary, no hype. Free.

Source. Full patent text and figures from the official USPTO publication PDF.

Editorial commentary on a publicly published patent application. Not legal advice.