—
title: “LG and NVIDIA’s Physical AI Bet: The $50B Infrastructure Race Nobody Is Talking About”
description: “LG and NVIDIA are quietly building the infrastructure for physical AI systems. Here’s what their talks reveal about the massive capital expenditure reshaping data centers in 2026.”
publishDate: 2026-05-02
category: AI News
tags: [Physical AI, NVIDIA, LG, Data Centers, AI Infrastructure, 2026]
—

Table of Contents

1. [What’s Happening Between LG and NVIDIA](#whats-happening-between-lg-and-nvidia)
2. [The Physics Problem Nobody Talks About](#the-physics-problem-nobody-talks-about)
3. [Why $50 Billion Is Conservative](#why-50-billion-is-conservative)
4. [LG’s HVAC Bet](#lgs-hvac-bet)
5. [What Physical AI Actually Needs](#what-physical-ai-actually-needs)
6. [Timeline and What’s Coming](#timeline-and-whats-coming)

—

While most AI headlines focus on models and benchmarks, something quieter and more consequential is happening in data centers around the world. LG and NVIDIA are in exploratory talks about physical AI infrastructure—and the numbers involved are staggering.

What’s Happening Between LG and NVIDIA

In April 2026, LG CEO Ryu Jae-cheol met with Madison Huang, NVIDIA’s Senior Director of Product Marketing for Omniverse and Robotics, in Seoul. The topic: physical AI systems, data center infrastructure, and mobility.

Physical AI—the branch of AI dealing with real-world robotic systems, autonomous vehicles, and industrial automation—requires compute densities that conventional data center infrastructure was never designed to handle.

Unlike pure software AI (language models, image generators), physical AI systems need to:

• Run inference on sensor streams in real-time

• Process massive amounts of spatial and video data

• Make millisecond decisions that interact with the physical world

• Maintain persistent awareness of changing environments

This is computationally expensive in ways that cloud AI serving simply isn’t.

The Physics Problem Nobody Talks About

The densification of compute clusters required for complex machine learning models creates what engineers call an unavoidable physics problem. NVIDIA’s data center business is generating record revenues—but operating high-density server racks is pushing conventional cooling infrastructure past safe operating limits.

Here’s the issue: A single NVIDIA H100 GPU generates up to 700 watts of heat. A dense rack with 8 GPUs produces 5,600 watts—5.6 kilowatts—in a space smaller than a household oven. Traditional air cooling can’t handle this at scale.

Physical AI compounds this because the systems need GPU clusters running continuously, not the burst-y usage pattern of typical cloud AI workloads. The thermal load is constant, not variable.

Why $50 Billion Is Conservative

Industry analysts estimate the physical AI infrastructure buildout will require $50-100 billion globally by 2028. Here’s why:

• Each autonomous vehicle development program needs 500+ GPUs

• A single smart factory requires 200+ GPUs for real-time control

• Robot-controlled logistics warehouses need 100-300 GPUs

• EachLevel 4+ autonomous vehicle on the road generates 4TB of data daily that needs real-time processing

Tesla’s Dojo supercomputer reportedly cost $1B+ for a single campus. Waymo’s infrastructure is estimated at $3-5B. Amazon’s robotics and autonomous logistics infrastructure is rumored to exceed $10B.

And these are just the leaders. Traditional manufacturers—BMW, Foxconn, CATL—are all building physical AI infrastructure now.

LG’s HVAC Bet

At CES 2026, LG positioned its commercial divisions to supply high-efficiency HVAC and thermal management solutions engineered specifically for AI data centers. This isn’t accidental—LG’s air conditioning and industrial cooling divisions see the GPU density problem as a massive opportunity.

The math is compelling: A hyperscale data center with 10,000 GPUs needs $50-100M in cooling infrastructure. With NVIDIA selling hundreds of thousands of GPUs annually into data centers, that’s a $5-10B annual cooling market opportunity.

LG isn’t alone—Schneider Electric, Honeywell, and Mitsubishi are all positioning for this. But LG’s advantage is vertical integration with its existing manufacturing footprint and relationships with Korean semiconductor fabs.

What Physical AI Actually Needs

Physical AI systems need three things conventional AI doesn’t:

1. Deterministic Latency
A robot arm in a factory can’t wait 500ms for an inference result. Physical AI needs inference latencies under 10ms, sustained 24/7. This means dedicated GPU clusters, not shared cloud infrastructure.

2. Continuous Sensor Processing
Cameras, LIDAR, IMUs—all generating data simultaneously. A physical AI system processing 8 camera streams at 4K resolution needs 10x the inference compute of a text-based AI assistant.

3. Edge Deployment
Most physical AI can’t run in the cloud—the latency is unacceptable and the bandwidth is insufficient. This means edge GPU clusters at factories, warehouses, and vehicles. Each deployment needs its own compute infrastructure.

Timeline and What’s Coming

2026: NVIDIA and LG finalize partnership terms. LG thermal management products integrated into NVIDIA’s reference data center designs for physical AI workloads.

2027: First physical AI-optimized data centers come online. Expect dedicated GPU clusters at major manufacturing sites (Foxconn, TSMC, BYD).

2028: Physical AI infrastructure spending hits critical mass. Expect major announcements from Google (TPU v5 physical AI workloads), Microsoft (Azure dedicated physical AI clusters), and AWS (Trainium-based physical AI instances).

The physical AI revolution is real—but it won’t be won in model benchmarks. It’ll be won in data center efficiency. And that’s a race LG and NVIDIA intend to lead.