Two Layers, Not Five

I'm Tony Uphoff, and this is the Uphoff on Media podcast, a practical take on how technology's rewriting the way B2B companies get built and run and how to get ahead of it for those founding, running, and funding B2B information marketing and technology companies, the businesses that run business. Okay, let's get to work. This episode is titled 2 layers, not 5. Energy, chips, and infrastructure are nice to have. Models and applications are need to have. That's where B2B leaders actually compete. Most B2B leadership teams are spending their AI planning time in the wrong layer of the stack. This is not a criticism of effort. It's a diagnosis of where the attention goes. Model selection, infrastructure vendor comparison, chip and compute questions that feel urgent because they're the ones dominating the trade press. Meanwhile, the layer where your actual competitive advantage lives is the layer getting shortchanged. Jensen Huang gave the industry a useful diagram at Davos this year, a 5-layer cake showing where the money in AI actually goes. Energy, chips, infrastructure, models, applications stacked bottom to top. Nvidia's blog explained it, and a wave of trade press and LinkedIn posts followed. By now, you've probably seen some version of it. What none of that coverage tells you is which layers you actually operate in. That's the only question that matters if you're building, running, funding, or starting a B2B information marketing or technology company right now. And it's a question Huang's cake was never built to answer. He built it for Wall Street and the hyperscalers writing 9-figure checks for data centers. You need a different read. Here's the split. 3 of Huang's 5 layers are nice to have, 2 are need to have. Confusing the 2 is costing operators time, budget, and attention they don't have to spare. The nice to have layers, energy, chips, and infrastructure are the foundation of the cake. The electricity, the GPUs, the data centers, and fiber that make real-time intelligence possible at all. They're also someone else's balance sheet. Unless you're running a utility, a chip company, or a hyperscaler, you don't build this layer, you rent it already priced into your cloud bill. Understanding these 3 layers makes you conversant at a conference or a cocktail party. It doesn't make you dangerous in a board meeting. Knowing that a frontier model run consumes as much power as a small city is a good fact to have in your pocket, It won't change a single decision you make Monday morning. Models and applications are where your decisions actually live and where the ground has shifted under most B2B teams without anyone announcing it. For the last 2 years, which model felt like the strategic decision? GPT, Claude, Gemini, or an open weight model fine-tuned in-house? That question is commoditizing fast. Model quality keeps converging, switching costs keep dropping, and every vendor keeps making the choice easier, not harder. Treating model selection as your differentiator in 2026 is like treating your choice of cloud provider as one in 2016. It matters operationally. It isn't the moat. The moat moved up one layer to applications. That's the shift underneath every B2B team's move from Q&A bots to agentic AI, the value stopped being, can we answer a question, and started becoming, can this thing complete the work inside our workflow on our data without a human catching every mistake? That's a harder problem than picking a model, and it's the one most companies are still underinvesting in while they debate their model roadmap. The AI stack, nice to know versus need to know. Energy, chips, and infrastructure are the foundation layers you rent. already priced into your cloud bill. Models and applications are where you build, buy, and compete. What this means by seed, for founders, stop pitching your model choice. Nobody's buying, we use GPT-5 as a differentiator anymore. And by the time you finish the sentence, the model landscape will have moved again. Pitch what your application does with the model that a general purpose assistant can't. The workflow it owns, the data it's trained on, the trust it's built up inside a customer's process. For operators, audit where your team's AI attention actually goes. Most B2B organizations still spend the bulk of their AI planning meetings on model selection and infrastructure vendor decisions, layers a competitor can copy in a weekend. Move that conversation up the stack. The question worth a full day of your leadership team's time is what an agent needs to see, touch, and decide inside your business before you trust it to act without you. And for investors, stop scoring pitch decks on which foundation model a company has integrated. That tells you almost nothing about defensibility, and every founder raising money right now knows it's the easy slide to fill. Score on the application layer instead. Proprietary data flywheels, workflow lock-in, and evidence the team understands the difference between a wrapper and a moat. Huang built his cake to explain infrastructure economics to people who write checks for power plants and data centers. You don't need to bake the whole thing. You need to own the top 2 layers before someone with more capital decides to own them for you. Well, that's the episode. If this was useful, please do one thing. Follow the show on Apple or Spotify and leave a rating. It's the single best way to help another operator find this. I'm Tony Uphoff, back at it next week.