Most coverage of NVIDIA's patent estate stops at the die. The grants the company received on June 2, 2026, suggest the more interesting story is what surrounds the die. Among the 38 patents issued to NVIDIA that day was US12648114B1, a method for controlling the flow of coolant inside a datacenter cooling system — the plumbing of the AI buildout, not the processor at its center.

The grant describes a secondary cooling loop that interfaces with a primary loop, governed by a hierarchy of proportional-integral-derivative (PID) controllers that set coolant flow rates from a coolant distribution unit out to a set of downstream flow controllers. The stated problem is a familiar one in operations: secondary flow controllers that "toggle" — oscillate — when they fight each other for the same coolant budget. The claims tie the flow decisions to feedback, and the specification reaches into NVIDIA's own neural-network toolkit for the inference step.

one or more neural networks to receive the temperature input, to infer the heat generated from one or more computing devices, and to enable the flow— PID control to address toggling of secondary flow controllers in datacenter cooling systems, US12648114B1

For a reader who tracks the AI infrastructure spend, that single claim is worth pausing on. Liquid cooling has become a gating factor for high-density GPU racks; the heat a modern accelerator shelf rejects is the reason the cooling plant exists at all. A granted claim that ties coolant-flow control to an inferred heat load is coverage on the facility-side response to NVIDIA's own thermal output. It maps the company's issued patents onto a part of the datacenter that, until recently, was someone else's problem.

The rest of the week: silicon, memory, autonomy

The cooling grant did not arrive in isolation. The same week's NVIDIA issuances cover the more expected territory of an accelerator company. US12645455B2 claims a tensor prefetch instruction that stores tensors directly into GPU caches — a memory-movement method aimed at the data-shuffling cost that dominates large-model workloads. US12645596B2 covers hardware-assisted page migration across a multi-dielet processing system, the kind of chiplet-to-chiplet memory plumbing that matters as single dies give way to multi-die packages.

On the model side, US12645944B2 claims a technique for training energy-based variational autoencoders, and US12645942B2 covers replacing corrupt neural-network gradient values with surrogate values inside transmitted data packets — a fault-tolerance method for the distributed training runs that span thousands of accelerators. A separate grant, US12645917B2, claims a neural architecture for self-supervised event learning and anomaly detection. And a sizable slice of the week's grants stay in NVIDIA's automotive and robotics lane, including US12646292B2 on surround-scene perception using multiple sensors.

What the spread of CPC classes shows

Taken together, the week's filings sort into three buckets that rarely share a portfolio: machine-learning method claims in the G06N classes, GPU and memory-architecture claims in G06F, and — the outlier — the cooling grant in the H05K 7/20 family that covers electronic-equipment cooling. The presence of all three in a single week of issuances is the data point. NVIDIA's coverage is not confined to what runs on the chip; it now includes the gradient-recovery method that keeps a training run alive across a failed link, the page-migration scheme that moves data between dielets, and the controller that decides how much coolant reaches the rack.

For the infrastructure reader, the cooling grant reframes a recurring question. The AI datacenter is usually discussed as capex — land, power, racks, and the accelerators inside them. A granted claim on coolant-flow control indicates NVIDIA is also assembling enforceable coverage over how that facility operates around its hardware, not only over the hardware itself. Whether that coverage is ever asserted is a separate matter; what the June 2 record shows is that the patents now reach the cooling plant.

The cluster also underlines how broad NVIDIA's issued estate has become. A company known for a graphics architecture held grants the same day spanning autonomous-vehicle perception, generative-model training, multi-die memory coherence, and datacenter thermal management. Each is a distinct technical area with its own competitors; each is now documented in a patent number with NVIDIA on the assignee line. The footprint described by a single week of grants is wide enough that no single label — chip company, AI company, infrastructure company — quite contains it.