The AI Stack Is Inverted: Why the Real Money Is in Infrastructure
As AI apps vanish into the model layer, and models commoditize, money and profits flow into infrastructure and real estate, not software
Traverse the AI Stack
There are three fundamental layers in the AI stack:
Infrastructure: the physical substrate: power generation, transmission lines, substations, transformers, cooling systems, data centers, networking equipment, and permitting regimes.
Models: the weights and architectures: LLMs, diffusion models, multimodal models, and the teams that train and fine-tune them.
Applications: chatbots, copilots, SaaS wrappers, and vertical-specific tools built atop foundational models.
The conventional wisdom is that value accumulates at the top of the stack. In other words, applications capture the most value, models command a premium, and infrastructure is a low-margin, high-CAPEX commodity. This is why venture capitalists are dumping money into “AI-native SaaS” and other layer 3 plays.
That view is not just wrong. It’s backwards. Read on for an explanation of how the AI stack really works.
Layer 3: A Doomed Frontier
The top of the stack—apps—is being eaten from below. As foundation models become more powerful, flexible, and context-aware, the AI-native app layer becomes unnecessary.
Why use a separate customer service chatbot product when GPT-6 can be fine-tuned to match a company’s tone and can be plugged directly into the company’s database? Why build an AI scheduling assistant when the model can directly interface with calendars and messaging APIs?
AI app companies are often compared to SaaS startups, but their unit economics are worse:
They don’t own the models, so they pay per-token or per-inference fees
They rely on brittle prompt engineering and shallow wrappers
They’re subject to margin compression as models gain capabilities and APIs undercut their offerings
This is a key misperception in the venture world: SaaS economics do not cleanly port over to AI-native apps. SaaS assumes control of the underlying logic and computation. In AI apps, the computation is outsourced, to OpenAI, Anthropic, or Mistral, and the logic is emergent rather than engineered.
Moreover, as multi-modal models evolve and build memory and long-term planning, the boundaries between “application” and “agent” dissolve. The orchestration logic moves into the model. And the developer's value prop becomes redundant.
The closer apps are to the model’s native affordances, meaning language, vision, or code, the more likely they are to be absorbed by the model itself. In the limit, the application layer becomes little more than a thin orchestration layer. Or it disappears altogether.
Layer 2: Models Under Pressure
The builders of foundational models are in a strategic race against commodification. Open weights, alternative architectures, and inference-time optimizations are eroding moat strength. Every model that launches is met with dozens of distillations, quantizations, and derivatives.
What’s more, user expectations are shifting. No one cares about your model’s architecture; they care about what it does and how quickly and cheaply it can do it. This pushes models into a functionally undifferentiated zone. Useful, yes, but generic.
To avoid becoming the Red Hat of AI, model companies will:
Lock in distribution by owning API endpoints, UX layers, and custom GUIs
Verticalize into specific domains with fine-tuning, retrieval augmentation, and compliance tooling
Integrate horizontally by buying or building inference infrastructure, chips, or datacenters
Some, like OpenAI and Anthropic, are already pursuing all three strategies. But even these are palliative measures, not cures.
Because the gravitational pull is toward commodification. Models are incredibly expensive to train and maintain, but their marginal cost to copy and serve is falling fast. This is a dangerous inversion: high CAPEX, low marginal pricing power.
And it’s accelerating. As more jurisdictions (China, France, UAE) push open weights for sovereignty reasons, and as distillation techniques improve, the notion of a defensible model becomes tenuous. Much like search or cloud infra, the market may tend toward a few players with scale advantages.
Even if models retain some economic power, their capital intensity makes them fragile. They are capital sinks, not cash cows. If the apps vanish and margins compress, the model builders will face an unpalatable choice: vertically integrate downward into infrastructure, or bleed out slowly. OpenAI’s Stargate project is both ambitious and indicative: Sam Altman has read the tea leaves and is acting accordingly.
Layer 1: Where the Bodies Are Buried
Value is consolidating at the base of the stack.
Why?
Because the physical constraints are binding. You can’t train a model without megawatts. You can’t deploy inference at scale without cooling and fiber. You can’t build datacenters without permits, electrical transformers, and substations.
And none of this scales on software timelines. LLM architectures can iterate every 6 months. Substation buildouts take 5-7 years. The timelines are desynchronized, and that asymmetry creates pricing power.
This mismatch means that:
Power purchase agreements (PPAs) become strategic assets
Permitting and zoning rights become defensible moats
Datacenter footprints are the new railroads: capital-intensive, slow-moving, and essential
Moreover, there is a geopolitical layer to all this. Control over compute is increasingly synonymous with control over intelligence. Sovereigns are waking up to this and scrambling to secure domestic capacity. That means public-private partnerships, permitting accelerators, and direct subsidies, none of which go to app companies. All of which benefit infra players.
You can’t copy a substation. You can’t distill an electrical transformer into a cheaper analog. The only way to scale physical infrastructure is to build it. And this is done slowly, expensively, and in the face of enormous regulatory friction. This is what makes it valuable.
Conclusion: The Inverted Stack
AI looks like software, but it behaves like infrastructure. The deeper you go into the stack, the more durable the value becomes. Apps will fade, models will fragment, but substations will keep humming.
The smartest investors and operators in AI aren’t chasing the next app wrapper. They’re buying land, signing energy contracts, and securing transformer capacity. They’re building where the constraints are hardest because that’s where the durable returns are.
AGI won’t be born in a laptop. It’ll be forged in steel, silicon, and substations. And the people who control the substations will be the ones who shape the intelligence economy of the 2030s and beyond.
Postscript: If You Have to Build at the App Layer, Do This
The only defensible app-layer moats are:
Deep domain coupling (e.g., EMR systems, CAD workflows, industrial control)
Proprietary data flywheels
Offline/edge deployment needs
Trust, compliance, or explainability mandates
Integration into legacy systems closed models won’t touch
But in general: foundational models are increasingly predatory. “Build on top” is not a stable strategy unless your app adds something truly irreplicable or your distribution moat is massive.
Great insight - so how does one get invested into this layer? For example, Stargate is owned by the parent companies. I can buy shares of Oracle and get a fraction of that, but I would prefer to find away to get integrated at that layer directly.