The Five-Layer Cake of AI: Why Vertical Startups Win at the Application Layer
As the infrastructure race hits the limits of capital and power, economic value is migrating to domain-specific software.
AI is often discussed as if it is a single, monolithic market. It is not. AI is a deeply integrated industrial stack. To understand where sustainable venture returns will be generated, we look to NVIDIA’s "five-layer cake of AI" framework. Every single intelligent output is the physical product of electrons, heat, silicon, data centers, frontier models, and software working in tandem.
But there is a sharp economic divide slicing through this stack. The first four layers: energy, chips, infrastructure, and models are where capital expenditure compounds. The fifth layer, applications, is where domain expertise compounds. The bottom layers manufacture intelligence. The top layer monetizes it.
For smaller funds, startup studios, and early-stage founders, the application layer is where vertical AI startups have a structural right to win.
AI is a Stack, Not a Single Market
To evaluate opportunities cleanly, investors must separate the layers of the industrial stack:
Energy: The foundational layer. Electricity, liquid cooling, grid access, power generation, and high-voltage transmission.
Chips: The hardware engine. GPUs, specialized accelerators, high-bandwidth memory, networking silicon, and semiconductor supply chains.
Infrastructure: The physical and operational architecture. Hyperscale data centers, cloud platforms, cluster orchestration, and high-performance storage.
Models: The algorithmic core. Frontier foundational models, open-source architectures, fine-tuned variants, and research laboratories.
Applications: The software delivery mechanism. Workflow-specific interfaces that translate raw model capability into measurable customer outcomes.
Each layer operates under entirely different economic realities, rewards different capabilities, and crowns different types of winners.
The First Four Layers Reward Capital Intensity
The bottom four layers of the AI stack represent an unprecedented infrastructure buildout. However, they are increasingly dominated by a handful of incumbents with massive balance sheets, sovereign backing, or entrenched hardware monopolies.
The infrastructure race requires billions of dollars in upfront capital. Hyperscalers outspend emerging funds by orders of magnitude, chip giants control the hardware allocation, and training frontier models requires ever-larger clusters and power grids.
While these layers are foundational and will produce immense enterprise value, they do not offer a structural advantage for early-stage venture capital or startup studios. In a game where raw scale, compute volume, and capital depth determine survival, smaller players cannot out-compute or out-spend the giants.
The Application Layer is Where AI Becomes ROI
AI only matters economically when it changes a real-world business outcome. Outside of tech hubs, enterprise buyers do not care about parameter counts or context windows; they care about margins, risk mitigation, and operational velocity.
The application layer is where raw intelligence is converted into financial return. Consider how value is captured when AI is applied to overlooked, complex industries:
Healthcare: A hospital system deploying predictive intelligence to reduce missed patient safety signals and lower clinical liability.
Operational Technology (OT): An industrial asset operator identifying cyber vulnerabilities across manufacturing infrastructure before causing costly downtime.
Supply Chain & Distribution: A wholesale distributor using automated line-item matching to eliminate short-pay claims and accelerate cash conversion.
Compliance & Finance: A highly regulated compliance team using deep-domain parsing to eliminate 90% of false-positive transaction alerts.
Agriculture: A multi-state agricultural network establishing immutable quality traceability to command higher premiums from enterprise buyers.
These are not generic chatbots or thin wrappers. These are workflow transformation engines built for highly specific, high-stakes environments.
The Shift from AI Adoption to AI Accountability
The broader enterprise market is entering a phase of strict economic discipline. For the past several years, large corporations spent aggressively on AI experimentation, tracking internal adoption metrics and token consumption. Today, boards and CFOs are demanding financial accountability. Public company commentary highlights this growing friction between AI usage and business value.
Uber has been open about the challenges of mapping massive token consumption directly to product velocity. As Uber President and COO Andrew Macdonald noted:
“That link is not there yet. It’s very hard to draw a line between one of those stats and, ‘Okay, now we’re actually producing 25% more useful consumer features.’”
Reporting from The Verge highlights that enterprise buyers are realizing that infrastructure costs must be weighed against tangible labor or efficiency gains:
“We’re going to have to start talking about token consumption and the associated cost versus headcount... If you’re not actually able to draw a direct line to how much useful features and functionality you’re shipping to your users, that trade becomes harder to justify.”
Similarly, Duolingo CEO Luis von Ahn recently walked back the company's approach of tying internal AI tool usage directly to employee performance reviews, recognizing that forcing technology adoption without a clear link to quality can backfire:
“The most important thing in your performance is that you are doing whatever your job is as well as possible. It felt like, rather than being held accountable for the actual outcome, we were trying to just push something that in some cases did not fit.”
This corporate skepticism is backed by broader data. PwC’s Global CEO Survey revealed that more than half of CEOs stated their companies had not successfully reduced costs or increased revenue through AI deployments over the prior 12 months. Only about one-third saw revenue expansion, roughly one-quarter achieved lower operational costs, and a mere one in eight realized both.
Furthermore, an RGP survey of 200 U.S. finance chiefs published by CFO.com revealed that only 14% had seen a clear, measurable impact from their AI investments. The primary culprit? A lack of data trust and reliability, only 10% of CFOs fully trusted their internal enterprise data.
Broad, horizontal AI deployments are stalling out because generic models cannot handle the messy, fragmented, and unvetted data environments typical of traditional enterprise operations.
"Tokenmaxxing" Is Not a Strategy
The early wave of AI software development was dominated by what we call "tokenmaxxing", the flawed assumption that maximizing prompt volume, adding more agentic steps, and consuming more tokens automatically equates to building more enterprise value. Emerging research into multi-agent workflows reveals that while complex agent architectures consume exponentially more tokens than simple chat interfaces, they often suffer from compounding error rates and ballooning inference costs.
The market is shifting from token volume to token efficiency. The next generation of vertical AI winners will not be judged by how much compute they consume, but by their precision. They will be measured on:
Cost per completed, audited workflow.
Time saved per expert operator.
Direct revenue uplift or cash conversion cycle acceleration.
Reduction in human review burden for highly regulated processes.
Gross margin improvement of the customer's core business.
Key Takeaway: The winners of the next phase of AI will be the companies that convert the fewest necessary tokens into the most measurable customer value.
Why Startups and Vertical Funds Have a Right to Win
While hyperscalers own the infrastructure, agile startups and focused venture studios possess structural advantages at the fifth layer that capital alone cannot buy:
Domain Specificity: Startups can obsess over narrow, painful, historically under-digitized workflows that are too niche for big tech to prioritize.
Customer Intimacy: Deep integration with design partners allows founders to build software that matches the exact cognitive patterns of real-world operators.
Proprietary Data Loops: By embedding directly into legacy systems, vertical applications create specialized, customer-labeled feedback loops that constantly improve model performance for that specific industry.
Workflow Ownership: Winning companies become the core system of action. Once an AI application owns the actual execution of a workflow, it becomes incredibly sticky.
Measurable ROI: Selling a highly targeted outcome (e.g., "we reduce compliance audit time by 60%") creates a clear, frictionless purchasing decision for enterprise buyers.
Trust & Human-in-the-Loop Design: In high-stakes fields like healthcare, critical infrastructure, and industrial supply chains, blind automation is a liability. Startups can design elegant human-in-the-loop interfaces that build trust rather than create operational risk.
Highway Ventures’ Point of View
At Highway Ventures, we are explicitly not in the business of funding massive compute clusters or competing in the foundational model arms race. Our mandate is to build and back vertical AI companies tailored for overlooked, operationally complex industries. We look for sectors where the data is messy, the regulations are strict, and the workflows are deeply entrenched. Whether it is patient safety intelligence in healthcare, operational technology (OT) cybersecurity, protein supply chain, or specialized industrial construction, our focus remains on the application layer.
By pairing world-class technical talent with deep domain experts and enterprise design partners, we build companies that transform how core industries operate. We believe that domain expertise, workflow integration, and deep customer alignment are the ultimate forms of defensibility.
Investment Framework for Application-Layer AI
When evaluating vertical AI companies, we use a rigorous, outcome-driven framework:
Workflow Pain: Is the targeted workflow frequent, expensive, highly regulated, or prone to catastrophic operational risk?
Budget Authority: Is there an explicit corporate buyer who owns the budget and feels the pain directly?
Data Edge: Does the startup have a unique path to access, structure, and leverage messy or proprietary industry data?
Integration Defensibility: Does the product embed deeply into existing systems of record, moving from a tool of curiosity to a system of action?
Quantifiable ROI: Can the product’s value proposition be explicitly measured in dollars saved, risks mitigated, or margin cleared?
Data Flywheel: Does the system naturally become more accurate and defensible as the customer uses it?
Go-To-Market Efficiency: Is there a repeatable, capital-efficient distribution path to reach fragmented enterprise buyers?
Token Efficiency: Is the architecture designed to minimize unnecessary inference costs, or is it burning margin through unoptimized "tokenmaxxing"?
Trust Architecture: Does the product include robust human-in-the-loop safeguards necessary for high-stakes, expert environments?
Wedge-to-Platform Potential: Can this narrow, initial product wedge scale into a comprehensive platform that owns adjacent industry workflows?
Conclusion
The AI era will be built by giants, but it will be applied by specialists. The first four layers of the AI cake manufacture raw intelligence at scale. The fifth layer turns that raw capability into economic value. For startup studios, vertical founders, and early-stage investors, the application layer is where defensible software companies are built.
The next generation of defining technology companies will be the ones that understand a complex workflow better than anyone else, design for trust, and turn raw tokens into clear, undeniable business value.
Source References
NVIDIA Blog: The Five-Layer Cake of AI — nvidia.com/blog/ai-5-layer-cake/
Fortune & The Verge: Corporate AI Spend and Token Metrics at Uber and Duolingo
PwC Global CEO Survey: AI Financial Performance Metrics
CFO.com / RGP Survey: Enterprise Data Trust and AI ROI Realities
