The "AI cake": five layers, decreasing capital intensity, and the value at the top

One-line summary: Jensen Huang's five-layer framing — base (power/cooling/land) → chip → infrastructure (data centers) → model (LLMs) → app — turned by jonathan-thomas into an investing discipline: capital intensity falls as you go up, value creation concentrates at the top, and the only question that matters per name is which layer are you in, and is that position sustainable, expandable and defensible?

The insight

"AI" is used so broadly that it hides the fact that its layers have opposite economics. The framing (Huang's, relayed by jonathan-thomas, CEO of American Century Investments):

Thomas's discipline: "when we talk about AI investing, you got to figure out which level of the cake you're in... What level of the cake are you in? And is that position that they currently hold sustainable, expandable and defensible."

Why this is useful rather than just tidy. It supplies a structural explanation for something this book has been circling from several directions at once:

• The bottom two layers require "massive investment, tons of capital" — which is why they show up in this wiki as capex, financing, and depreciation problems (mega-issuance-peak-to-ai-capex-derate, steroid-era-earnings-inflation, nvidia-gpu-backstop-to-neocloud-financeability). Capital intensity is the cause of the financing fragility, not a coincidence beside it.
• The top layer requires the least capital and is where Thomas locates value creation — which is the same claim open-source-share-shift-bullish-for-compute makes from the opposite end (margin displaced from the model layer routes down to compute) and ai-creators-to-adopters-rotation makes from the demand side.
• It explains why "is AI a bubble?" is unanswerable as posed. Layers 1-2 can be over-capitalised while layer 5 is under-priced, simultaneously. The cake is not one asset.

Where it cuts against the book. Most of this project's AI exposure sits in the bottom two layers — the most capital-intensive, most financing-dependent, and (on Thomas's read) not where the value ultimately accrues. That is not a reason to exit; those layers are where the bottlenecks are, and bottleneck rent is real (hbm-cowos-as-binding-bottleneck, cowos-packaging-capacity-crunch). But it is a reason to hold the exposure with an explicit expiry: bottleneck rent is a function of scarcity, and scarcity is a phase, whereas Thomas is describing where value settles after the phase.

Thomas is candid that this is hard to operationalize. Josh Brown: "And I think on a stock, by stock basis... Very hard to do." Thomas: "It is, but we use the words AI so broadly, and there's just so much more to it than that." The framework is a sorting discipline, not a screen.

The scope creep is real and it is the interesting part. sean-russo: "It's nuts to me how many sectors it affects. Industrials, utilities, energy. Like, almost every single sector, maybe, except for financials and healthcare. We wrote about Caterpillar." And Brown's inversion — financials and healthcare are precisely where the app-layer (adopter) upside sits: "healthcare and financials arguably have companies that could be among the biggest beneficiaries of AI as users, as adopters."

Evidence

All from 2026-07-17-podcast-the-compound-and-friends-you-re-about-to-see-the-real-ai-winners-stand-up.

• The five layers — jonathan-thomas: "He talks about the AI is a cake and he says it has five layers to it... There's the base level, which is really the ground, the cooling, the electricity. You've got the chip level, right? The nvidias of the world, you got the infrastructure level, which are the data centers themselves. You have the model level, which are the LLMs. And then sitting on top of all of that is the app level, where the actual value is created from an investing point of view... those first two levels, the base level and the infrastructure level require base and chip require massive investment, tons of capital. As you go up that cake, it requires less and less capital."
• The discipline — jonathan-thomas: "when we talk about AI investing, you got to figure out which level of the cake you're in. Okay. Because it's not just the word is used so broadly and there's so much underneath it. What level of the cake are you in? And is that position that they currently hold sustainable, expandable and defensible."
• It was offered as the answer to a concentration question — Josh Brown asked directly: "Do you worry about bubbles in pockets of the market like semiconductors, memory stocks? Are there areas where you guys maybe think... we ought to do something about concentration here?" Thomas's answer was the cake — i.e. he uses it as the substitute for a top-down bubble call.
• Where the app layer lands — jonathan-thomas: "Going to that last level of the cake, the apps and stuff. That is where the drug discovery comes from, the financial and services."
• The adopter examples — josh-brown: "healthcare and financials arguably have companies that could be among the biggest beneficiaries of AI as users, as adopters. Like what? Think about insurance underwriting. Think about... drug discovery."
• Cross-sector scope — sean-russo: "It's nuts to me how many sectors it affects. Industrials, utilities, energy. Like, almost every single sector, maybe, except for financials and healthcare. We wrote about Caterpillar."
• Live app-layer adoption data, from an adopter's own call — patrick-conway in 2026-07-16-earnings-unh-q2-fy2026 (⚠ partial source): ambient-listening AI at "70% of employed Optum Health providers" targeting "exceed 90% by year-end"; Value Connect AI credited with pharmacy cost savings of "17% for early clients." Thomas's own hedge on the same: "I don't think it's gonna come anytime soon where you're gonna remove completely human judgment from the process... But all the things that feed into that ultimate decision are going to be fed by AI. And I agree with you. I don't think it's hype. I think it's promise."

Implications

• Sort every AI name by layer before comparing them. NVDA (chip) and an insurance underwriter deploying AI (app) are not the same trade and should not share a risk budget line. This is a useful cross-check on the cluster field in the trader Signal contract.
• Capital intensity is the through-line for the book's fragility theses. Layers 1-3 are where the financing chains live. That is structural, not incidental.
• Most of this book sits at the bottom of the cake — the capital-intensive, financing-dependent layers where Thomas says value does not ultimately accrue. Hold with an expiry: bottleneck rent is a phase.
• The app layer has no clean expression in this book yet, and Thomas/Brown both point at healthcare and financials as adopters. See ai-creators-to-adopters-rotation.

Contradictions / tensions

• ⚠ It is Jensen Huang's framework, relayed second-hand, and Huang sells layer 2. Brown's own reaction: "What a salesman." A taxonomy authored by the chip vendor that places "massive investment, tons of capital" at the chip layer is not a neutral map — it is also a sales argument for why everyone must buy his product first. Thomas relays it approvingly without interrogating that.
• Not operationalizable as stated, and the participants say so. Brown: "Very hard to do." No source gives a method for deciding whether a position is "sustainable, expandable and defensible," which is where all the work is.
• The "value is created at the app layer" claim is asserted, not evidenced. It is the framework's most load-bearing and least supported element. The dot-com analogue Thomas himself offers cuts both ways: value did eventually accrue above the infrastructure, but "the Internet took a long time to really crystallize its full effect" — he dates full realization to 2020, twenty years after the capex bubble. An investor who bought the app layer in 2000 waited two decades.
• The layer boundaries are already dissolving. Nvidia sells across chip, infrastructure and model; hyperscalers span base through model; frontier labs are integrating vertically (frontier-lab-vertical-integration-to-sovereign-ai-stack). A five-layer map of a stack that is re-integrating may describe 2023 better than 2026.
• Single-source (one podcast interview), and Thomas runs an asset manager whose products benefit from the "broadening out" story he is telling.

Open questions

• Is there a citable version of Huang's cake framework from Huang directly, rather than relayed?
• Does capital intensity per layer hold quantitatively — is there a figure for capex-to-revenue by layer?
• If layer boundaries are dissolving via vertical integration, does the framework survive its own premise?

Related

• ai-creators-to-adopters-rotation — the demand-side version of the same rotation
• open-source-share-shift-bullish-for-compute — margin displaced from the model layer routes down to compute
• steroid-era-earnings-inflation
• mega-issuance-peak-to-ai-capex-derate
• ai-roi-reckoning
• frontier-lab-vertical-integration-to-sovereign-ai-stack — the layers re-integrating
• jonathan-thomas
• jensen-huang
• josh-brown