The artificial intelligence industry is often discussed as a single sector, but its economics vary enormously depending on which layer of the stack a participant occupies. There are four broad layers, each with its own competitive structure, margin profile, and cyclical sensitivity. Understanding where any given company sits — and how the layers connect — is the entry point to any serious analysis of the sector.
Layer one: hardware
At the base of the stack are the chip designers and manufacturers whose products power AI computation. Nvidia dominates leading-edge AI accelerators (H100, H200, and successor architectures). AMD holds a growing but still small share with its MI300 series. Custom silicon from the hyperscalers — Google's TPUs, Amazon's Trainium and Inferentia, Microsoft's Maia — represents a small but rising fraction of total capacity, primarily consumed internally.
Nvidia's economic position is the most-discussed in the entire technology sector. Its data-center revenue has grown from roughly $15 billion annually in 2022 to more than $150 billion in 2026 — an order-of-magnitude increase in four years. Gross margins on its AI products have expanded to the 70%+ range, extraordinary by hardware-industry standards. Whether these margins are sustainable is one of the most important open questions in technology.
The margin sustainability question hinges on three factors. First, whether alternative accelerators (custom silicon, AMD, various startups) can achieve competitive performance-per-dollar. Second, whether workload characteristics shift over time in ways that reduce the winner-take-all dynamics of the current generation. Third, whether Nvidia's software moat (CUDA, the developer ecosystem, the reference implementations of every major model) proves durable as more of the ecosystem matures.
None of these factors is settled. The bull case rests on Nvidia's engineering execution and ecosystem dominance persisting; the bear case rests on any of the three factors shifting materially.
Layer two: infrastructure (data centres and cloud)
Above the chips is the layer of infrastructure — the data centres, the cooling, the power, and the cloud platforms that abstract raw compute into consumable services. The hyperscalers (AWS, Azure, GCP) sit at the top of this layer. Independent operators like Coreweave, Lambda, and others provide capacity to customers that either cannot access hyperscaler allocations or prefer the pricing terms of alternative providers.
The economics of this layer have been transformed by AI demand. Historical cloud infrastructure was a scale business with modest returns on capital; incremental investment produced steady but not spectacular growth. AI compute demand has upended that pattern: hyperscalers are committing to capital expenditure programs in the hundreds of billions of dollars over multi-year windows, at return-on-capital assumptions that would have been considered aggressive in the pre-2023 environment.
Whether these commitments prove profitable depends on the demand curve for AI compute over the next five to ten years — a topic on which reasonable people disagree substantially. If demand continues its current growth trajectory, the capital being deployed today looks conservative; if demand plateaus or the workloads become dramatically more efficient, the capital deployed at 2025–2026 pricing may not earn its cost of capital.
Layer three: models and platforms
Above the infrastructure sit the model developers — OpenAI, Anthropic, Google DeepMind, Meta, xAI, various open-source ecosystems — and the platforms that host and deliver models to end users. This layer has the most rapid competitive dynamics of any in the industry: leadership positions have shifted materially within twelve-month periods, and the technical barriers to entry are lower than in the hardware or infrastructure layers.
The economics of frontier model development have become extraordinary. Training runs for leading models now cost hundreds of millions of dollars each. The number of organisations that can fund such runs is small and growing more concentrated. Simultaneously, the marginal cost of inference (running an already-trained model) continues to fall rapidly as engineering optimisation accumulates.
The competitive question at this layer is whether frontier capability is sustainable as a differentiating asset. If frontier models remain meaningfully better than commodity ones, model developers can charge premium prices; if the practical difference between frontier and second-tier models narrows for most use cases, the pricing power at this layer erodes toward the cost of inference.
The evidence to date suggests both dynamics operating simultaneously. Frontier capability continues to advance meaningfully with each generation, but the "good enough" threshold for many practical applications has also risen sharply, meaning commoditised offerings satisfy an increasing share of demand at prices well below frontier-model rates.
Layer four: applications
At the top of the stack are the applications that integrate AI into user-facing products — enterprise software, consumer applications, and vertical-specific tools. This layer is the most fragmented and has the most competitive uncertainty.
Some incumbents are integrating AI into existing product lines with meaningful revenue impact — Microsoft's Copilot suite, Salesforce's Einstein, Adobe's Firefly and generative-AI features. Some startups are building AI-native products in specific verticals: legal research, medical imaging, customer service, code generation, content creation. The winners at this layer are not yet clear and may not be for years.
The economic question at the application layer is who captures the value that AI enables. If enterprise customers derive substantial productivity gains from AI, some fraction of those gains flows to the application providers, some fraction to the model providers, some fraction to the infrastructure providers, and some fraction to the customers themselves. The relative shares are being contested actively.
The connections that matter
The four layers are connected by supply-and-demand relationships that determine the health of the ecosystem. Application demand pulls model demand. Model demand pulls infrastructure demand. Infrastructure demand pulls hardware demand. When each layer is growing rapidly, the cycle reinforces itself. When any layer stalls, the effect propagates.
The concentration at each layer also matters. If the model layer commodifies but hardware remains concentrated, most of the value accrues to hardware. If application winners emerge with defensible positions, they capture disproportionate value. If none of the layers commodifies and all remain highly competitive, the aggregate returns to the ecosystem are lower than the current investment levels imply.
What this frame does not answer
Nothing above predicts which specific companies will succeed at each layer or which layer will capture the most value. The frame's utility is not predictive; it is structural. Any news story or research report about "AI" can be located in the frame — which layer, which relationship, which competitive dynamic — and reading it with the layer in mind produces sharper analysis than reading it as generic AI commentary.
Educational content only. Not investment advice.