Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
From The Floor
A Brief History of Civilian Drones
Scan of the Month
Apple AirPods Pro (2nd Generation) CT Teardown
Materials World
Apple Rethinks Paper Packaging
Design to Reality
Apple vs. Meta: Same Problem, Different Answers
The Quality Gap
Behind the Battery Report
The Quality Gap
Blind Spots in Electronics Quality
Design to Reality
Building for the Brain: Pioneering a Long-Term Neural Implant
Design to Reality
CT Teardown: AirPods Pro (3rd Generation)
Recall Radar
Defective On Arrival
The Quality Gap
Do Water Filters Actually Work?
Design to Reality
Eight Years to Redesign a Ketchup Cap
Design to Reality
Evolution of the Plastic Bottle
Recall Radar
Fall on Fire
From The Floor
Finding Lead in Stanley's Quencher
Design to Reality
Furbo and KONG: Two Ways to Give a Dog a Treat
Recall Radar
Hidden Failures of Everyday Interfaces
Design to Reality
How Does a Car Cigarette Lighter Work?
Scan of the Month
How Four Pens Solve the Same Problem
From The Floor
How Ground Truth Data Builds Trust Between OEMs and Suppliers
Design to Reality
How I Think About R&D (and Turning Ideas Into Products)
The Quality Gap
How People Drive Quality
From The Floor
How Saucony Uses CT to Build Better Running Shoes
Design to Reality
How SawStop Stops a Saw Blade in 5 Milliseconds
Materials World
How the Wine Industry Engineered Around Cork's One Flaw
From The Floor
How to Read a Plastic Bottle
Materials World
How We Learned to Hold a Bit Still
From The Floor
Inside a 12‑Month Sprint from Concept to Factory‑Ready Product
The Quality Gap
Inside The Best Smoke Alarms
Scan of the Month
Inside the Nintendo Switch 2 Joy-Cons
From The Floor
Jensen Is Wrong. The Cake Has Six Layers.
From The Floor
Malicious Hardware Hidden in Plain Sight
From The Floor
Manufacturing in 2026: Less Disruption, More Discipline
Materials World
Materials That Make or Break a Shoe
Design to Reality
Not All USB-C Cables Are the Same
Design to Reality
Nothing Ear 3 Has Nothing to Hide
Recall Radar
Parts Under Pressure
Design to Reality
Pulling the Thread on Talenti’s Stubborn Lids
Materials World
Reading the Rock
The Quality Gap
Reshoring's Hidden Constraint: Quality
Recall Radar
Routine Uses, Real Risks
Recall Radar
Safety Gaps That Keep Shipping
Recall Radar
Same Mistake, Same Results
Materials World
Seashell Architecture
Materials World
Speaking in Steel and Sapphire: MING’s 20.01 Series 5
Recall Radar
Stored Energy Meets Soft Spots
From The Floor
The Missing Middle in Battery Manufacturing
Design to Reality
The Pink Tax: Are Men's and Women's Razors Actually Different?
The Quality Gap
The Price of Trust: Behind the Takata Recall
Recall Radar
Tolerance Tested
Recall Radar
Too Hot, Too Sharp, Too Loose
Scan of the Month
We CT Scanned a Bag of Chips and 3D Printed the Results
The Quality Gap
What Counterfeit Apple Products Look Like on the Inside
The Quality Gap
What Food Manufacturers Can't See
Scan of the Month
What Medical Connectors Have to Get Right
Design to Reality
What QMSR Means for Medical Device Product Lifecycle Management
Materials World
What Roasting Does to a Coffee Bean, Seen From the Inside
The Quality Gap
What’s Hiding Inside Haribo’s Power Bank and Headphones?
Scan of the Month
What's Inside a Contactless Credit Card
Design to Reality
What's Inside the World's Fastest Marathon Shoes
The Quality Gap
What Went Wrong Inside These Recalled Power Banks?
The Quality Gap
What You Don’t Measure Will Cost You
From The Floor
Your Toner Cartridge Is Lying to You
Materials World
バッテリーの中身は?
From The Floor
August 2026

Jensen Is Wrong. The Cake Has Six Layers.

In this Article:

  • Jensen Huang's five-layer AI infrastructure stack, energy, chips, infrastructure, models, applications, is missing a critical layer: data, the raw material every model learns from.
  • Manufacturing has no equivalent of the internet. The data that matters is rare, often hidden inside a part or assembly, spread across incompatible systems, and mostly proprietary.
  • Lumafield's scanners capture the physical truth inside manufactured products, and next week the company will launch an AI quality agent powered by that Lumafield CT data, an application built on the data layer Jensen left out.
  • 8.26.2026

    Jensen Huang has been touring a five-layer AI infrastructure cake since February: energy, chips, infrastructure, models, applications. It is a good cake. It’s the kind of cake that drives the economy to new heights. 

    Jensen is right that AI is not an app or a single model. He is right that it runs on real hardware, consumes real energy, and rests on a vast physical buildout. He is right that every successful application pulls on every layer beneath it, all the way down to the power plant.

    He is missing a critical layer though. 

    In Jensen’s telling, AI factories “are not designed to store information. They are designed to manufacture intelligence.” Intelligence is not manufactured from electrons and GPUs alone. It is manufactured from data.

    The cake has six layers:

    Energy → chips → infrastructure → data → models → applications.

    Nvidia's AI 5-layer cake.

    The data comes first

    The most capable AI systems did not just appear when somebody finally found the right model architecture. They required a sufficiently large corpus of data to complete training.

    Large language models were trained on the accumulated text and images of the public internet. The transformer could have been invented in 1985. ChatGPT could not have been trained then, because the training data did not exist.

    Self-driving cars required a different set of training data. Tesla’s fleet passed 10 billion miles driven with FSD (Supervised) this year, a milestone Elon Musk has identified as roughly the amount of data needed for safe unsupervised driving. Someone had to record the miles to train the model. 

    AlphaFold required decades of experimental biology. It learned from structures in the Protein Data Bank, protein sequences and annotations in UniProt, and metagenomics databases including MGnify. John Jumper, who led the AlphaFold team, put it plainly: “Public data were essential to the development of AlphaFold.” No PDB, no AlphaFold.

    Where the training data is missing, AI lags.

    Robot manipulation has no equivalent of the internet. A video of someone folding a shirt does not contain the joint torques, contact forces, actions, or failures needed to teach a robot hand. The largest open real-robot dataset, Open X-Embodiment, pools more than one million trajectories from 22 kinds of robots. That is an impressive start, but still tiny next to internet-scale training. Companies like Generalist are now building embodied models and generating the physical experience those models need.

    Materials science has a similar hole. The Materials Project contains computed data for more than 200,000 materials, yet Berkeley Lab says experimental properties are available in the open literature for fewer than one percent of compounds. A model can propose a stable crystal on a computer. That does not mean anyone can make it, or that it will have the same properties in the real world. Lila Sciences and Periodic Labs are building autonomous laboratories because simulation alone cannot supply the missing ground truth.

    Data is a layer, not a detail

    It is tempting to treat data as something contained inside the model layer, but it is much more than that. 

    Energy and chips are portable across domains. The same GPU can train a language model, a protein model, or a robot policy. Data does not transfer so easily. The text of a Reddit thread cannot teach a robot to grip a glass. Self-driving car training data cannot reveal the internal porosity of a casting. Protein structures cannot tell you if a battery is going to start a fire. 

    Each new class of AI application needs its own instrumented view of reality.

    That is also why data belongs in Jensen’s reinforcing loop. Better instruments produce better data. Better data produces better models. Better models enable new applications. Successful applications create a reason to deploy more instruments and capture more data. Every turn of that loop creates more demand for infrastructure, chips, and energy.

    The data layer is not sitting passively between infrastructure and models. The stack is incomplete without it. 

    Manufacturing has no internet

    “Physical AI” has become a catch-all for humanoid robots, autonomous cars, industrial systems, and almost anything else where software touches the real world. Manufacturing is a more specific problem.

    The data that matters is difficult to collect. Important manufacturing defects are rare. Many are hidden inside a part or assembly. The relationship between design, process, internal structure, and field performance is usually spread across incompatible systems, if it is recorded at all. And most manufacturing data is proprietary. There will be no Common Crawl of the world’s factory floors.

    CAD describes what engineers intended to make. But manufacturers need to know what was actually made, inside and out. Paired with process information and inspection outcomes, that information becomes the verifiable ground truth a manufacturing model can learn from.

    Internal porosity in an aluminum casting, visualized in Lumafield’s Voyager software.

    This is the layer Lumafield has been building since we started the company. Our scanners capture the physical truth inside of manufactured products. Our software makes that data usable across teams and over time. Our custom AI models can then learn from the geometry, defects, and decisions contained in that data specific to each individual customer.

    Next week, we will launch an AI quality agent powered by Lumafield CT data. It is an application at the top of the cake. Its capability comes from the layer Jensen left out.

    Factories will not realize the full benefit of AI until they can see, record, and learn from what they actually make.

    The cake has six layers. We are building the sixth one for manufacturing.

    Scan on,

    Eduardo

    Citations
    No items found.