So...NVIDIA is Building AI Factories Now?

Jensen Huang has left AI Models behind. What's his next play?

Author

Sebastian Muehl
March 20, 2025

Over the last 24 hours, I’ve been thinking about a good analogy for what Jensen Huang described at GTC—AI factories, token generation, and inference.

Here’s what I came up with.

Winter Sundays in my house mean one thing: movie time. Everyone gathers in front of the TV. We scroll through Netflix or Disney+ for what feels like forever—everyone wants something different, and no one can agree.

Eventually, I type in a search. The system retrieves pre-existing results. Another list of movies. More scrolling and debating.

Now imagine we don’t have to pick a movie. Instead, we call a Hollywood studio and say:

👉 Give us a Pixar-style Indiana Jones, but with Christoph Waltz as the villain and Jonah Hill as the eccentric sidekick he played in Wolf of Wall Street. (I’d watch that)

And in real-time, the studio creates it, just for us.

🚀 This is the shift happening in AI.

We’re moving from retrieval (searching for pre-made content) to generation (AI factories producing content on demand).

This is what Jensen Huang described in his keynote. NVIDIA isn’t just selling chips—it’s supplying the hardware and software for these AI factories.

And AI factories, read this slowly:
🟢 "Will be the new industrial infrastructure of intelligence itself."

Like electricity in the early 20th century, which powered everything that followed, AI factories will power the next era of business.

High Signal Alert: Nvidia’s Master Plan

"AI factories will be the new industrial infrastructure of intelligence itself."

Jensen Huang

When Jensen Huang, CEO of Nvidia, talks, I pay very close attention. Yesterday, he delivered one of the most important (tech) keynotes of the year at GTC.

Although he said “AI” 147 and “scale” 66 times in his speech (yes, I pulled the transcript and counted with CTRL + F) this wasn’t just hype. It is where he sees the world going. And he’s been right in the past — a lot.

From Training to Production: The Shift to AI Factories

AI isn’t about training models anymore — it’s about running them at scale. AI is everywhere now — constantly generating, optimizing, and deploying itself in real-time.

AI is becoming an infrastructure layer at a large scale, like electricity or the internet.

Model training does not scale

We’ve discussed this before:

  • Training is expensive and slow

  • Nobody makes money from “training the models.” The real value comes from running the models and supporting use cases.

  • Keeping AI updated requires constant re-training, which is unsustainable at scale.

To solve this, we need more compute and better hardware

Huang laid out his new product line and roadmap to solve exactly these bottlenecks:

  1. Explosion of AI inference workloads — AI is a power hog. Running these models burns through compute, electricity and the bank account.

  2. Need for efficient AI processing — Reducing energy costs for running AI models.

  3. AI-driven industries are emerging — Robotics, enterprise automation and all kinds of AI driven business processes are taking off, AI will become the backbone of digital infrastructure.

“The computer has become a generator of tokens, not a retrieval of files.”

Jensen Huang

Nvidia’s Master Plan - the AI Production Line

Full-stack AI - from chips to software. In this evolution to the AI era, Nvidia wants to control the hardware and critical software layers.

The blocks of AI output - Tokens are created by datacenters (factories). And Nvidia is the supplier for them.

  1. AI compute power: Blackwell & Vera Rubin GPUs

    • New GPU: Blackwell Ultra with another 1.5x increase of performance. And Vera Rubin (2026/2027) even more gains.

    • These chips are not just raw speed but are designed to reduce inference costs.

  2. Dynamo Software - The Factory OS

    • The Operating System for the AI factory. Ensuring AI workflows run efficient and minizing cost and latency. The big shift: Nvidia is now offering the entire AI operating system on top of just chips.

  3. AI-Powered Robotics

    • New tools that bring AI reasoning into the physical world. This will speed up robotic leaning massively. Nvidia starts with an open-source model for humanoid robots: GR00T N1.

The third point here fits into the layer as a transition from the digital world (tokens, decisions, predictions) into the physical world. Automation of physical tasks and the full end-to-end cycle is complete.

AI Inference is Nvidia’s New Moat

Nvidia’s strategy? Lock in customers with a full-stack ecosystem. (hw and sw), making it hard to leave. The ecosystem path that Apple has so wonderfully paved for the end consumer. (See post “Why Apple’s Innovation Feels Different Now”).

No company wants to run around and Frankenstein their infrastructure from 10 different vendors. I know, I played the B2B software game for a while. They want one stack that works seamlessly. Companies that can make AI think faster without breaking the bank, will win.

Nvidia is positioned to be the ONLY player right now that can get you there.

Nvidia is turning its AI factory into a service that everyone will need.

OR IS IT?

The open-source challenge to Nvidia’s ecosystem

What if Jensen Huang’s vision about AI isn’t the only future? What if AI development shifts away from massive, central data centers and this AI factory need?

This would deem Nvidia’s vision at risk.

If it turns out that the rise of smaller, open-source AI models can run efficiently without Nvidia’s expensive HW, there might be a second reality in which Nvidia’s future will look different.

Or could it be both? 

In other news, last July, we analyzed Google's possible acquisition of Wiz. Now it happened. They said no to $23 billion last year. It turns out that $32 billion was the number. Wiz is an amazing example of fast scaling—since it’s timely, here is the deep dive.

Breaking Records: How Wiz Reached $500M ARR in Just 4 Years

The Game-Changing Strategies Behind Wiz's Meteoric Rise

www.techtalesandtactics.com/p/wiz-meteoric-rise

PS: Instead of focusing on the specific announcement (like most news outlets, I took the chance to zoom out. I hope this is as interesting for you as it is for me.

PS: I am currently working on some new branding for the newsletter. I am still taking suggestions and would appreciate feedback—just reach out, and I’ll show you what’s coming soon.

Have a great rest of the week,

How did you like this edition?

Login or Subscribe to participate in polls.

Reply

or to participate.