TL;DR / The Quick Version
The AI technology stack runs from physical silicon all the way to the apps on your phone — seven distinct layers, each with its own physics, economics, and constraints. Most AI discourse — hype, skepticism, investment theses, and hot takes alike — operates exclusively at the top two layers: application and UX. That's like having an opinion about skyscrapers while only ever looking at the lobby.
- —The physical layer is controlled by a single company with an unbreakable monopoly.
- —The compute layer is experiencing the largest capital buildout in the history of technology.
- —The data layer is already being exhausted. The model layer is commoditizing faster than anyone predicted.
- —And the application and UX layers — where 99% of the conversation lives — are built entirely on assumptions about what the six layers below them can sustain.
If you want to understand what AI can and cannot do, where the real value concentrates, and which companies are genuinely defensible vs. just well-funded, you need a map of the full stack. This is that map.
I worked at ASML — the company I describe in Layer 1 as the sole manufacturer of the machines the entire AI industry depends on. I've also built and shipped a production AI application with real customers, from scratch, solo. Almost no one writing about this stack has stood at both ends of it. This post is written from that vantage point.
What Is the AI Technology Stack?
The AI technology stack is the full set of layers required to go from raw physics — atoms, light, electricity — to the AI products people actually use. There are seven of them. Each layer depends entirely on the one below it. A breakthrough at layer two unlocks new capabilities at layers four and five. A constraint at layer one caps everything above it. They are not interchangeable, they are not equally understood, and they are not equally covered.
Here they are, bottom to top:
- —Physical / Silicon
- —Compute / Cloud Infrastructure
- —Data
- —Model / Algorithm
- —MLOps / Tooling
- —Application
- —UX / Distribution
Most people live in layers six and seven. Most of what shapes the future of AI is determined at layers one through four.
Why 7 Layers, Not 3, 5, or 6: Defending This Framework
Before going deeper, a legitimate question deserves a direct answer: why seven layers? Different frameworks exist, and they disagree on the number. NVIDIA CEO Jensen Huang has articulated a widely-cited five-layer model — energy, chips, infrastructure, models, and applications — described at Davos as the structure of "the largest infrastructure buildout in human history." Other frameworks collapse the stack into three layers (infrastructure, model, application) or propose six. Which is correct?
All of them are internally consistent. But they are answering different strategic questions.
The five-layer framework (Huang's model) is optimized for understanding geopolitical and industrial competition. It separates energy as a distinct layer because power generation is a genuine competitive advantage at the national level — the US-China AI race is partly an energy race. It separates chips from infrastructure because NVIDIA's own business depends on distinguishing the GPU from the data center around it.
The three-layer framework is optimized for simplicity. If you collapse everything below the model into "infrastructure," you get a cleaner picture for venture investors or early-stage founders who need orientation, not granularity.
**This seven-layer framework is optimized for a different purpose: understanding where constraints actually live, and therefore where defensibility can actually be built.**
The key distinction is the separation of physical silicon manufacturing (layer one) from cloud infrastructure (layer two). Most frameworks merge these. This one does not — because they operate under completely different physics, timelines, and bottleneck profiles.
ASML, the Dutch company that produces the machines required to manufacture advanced AI chips, shipped 48 EUV systems in 2025. It is targeting more than 60 in 2026. The entire global supply is allocated years in advance. A new fab cannot begin operation until ASML delivers a machine — no amount of capital spending upstream can accelerate this. Each machine contains roughly 100,000 parts and takes months to assemble in specialized cleanrooms.
Hyperscalers, by contrast, deploy thousands of servers per month. They iterate rapidly, rebalance capacity, and respond to demand shifts in weeks. The binding constraint is capital and power, not manufacturing lead time.
These are not the same constraint. Merging them into one layer obscures the single most important bottleneck in the entire AI supply chain.
Similarly, data (layer three) and models (layer four) are separated here because the competitive advantage they offer is structurally different. A company with proprietary data cannot be displaced by someone with a marginally better model. A company with a marginally better model can be displaced by someone with better data. They are not interchangeable inputs — they are different types of moats.
And MLOps (layer five), application (layer six), and UX/distribution (layer seven) are separated because most builders conflate them, creating strategic blindness. You can have a technically superior application with no distribution and lose to an inferior product with better reach. You can ship to production without production-grade MLOps and discover at scale that your system is unreliable or too expensive to operate.
The point of this framework is not to be more correct than others. It is to be more useful for a specific set of questions: where are the real constraints? Which layers can I own? Which am I renting? Which layers can kill my business, and which can make it defensible?
Different frameworks give different answers. This one is built to give the right answers to the questions that matter most for anyone building, investing in, or forming a serious opinion on AI.
Layer 1: Physical / Silicon — The Layer Nobody Talks About
At the foundation of every AI system is a chip. That chip exists because someone manufactured it. And that manufacturing process is among the most constrained industrial operations in human history.
To produce a modern AI chip, you need specialized fabrication equipment capable of extreme precision at a scale most people find difficult to visualize. The only company in the world that makes the most advanced of these systems is ASML, based in the Netherlands. Each machine contains roughly 100,000 parts, weighs 180 tons, and takes months to assemble in specialized cleanroom environments. ASML shipped 48 EUV systems in 2025 and is targeting more than 60 in 2026 — with demand far exceeding supply and lead times stretching years into the future.
These machines cost between $200 million and $380 million each, depending on capability level. The production constraint is real and inescapable: no factory anywhere in the world can operate faster than ASML can deliver the equipment to build it. As ASML CEO Christophe Fouquet told CNBC in May 2025: "We barely made it. It was a very risky investment because when we started, there was no guarantee the technology would work."
The chips themselves are then fabricated almost exclusively by TSMC, which controls roughly 72% of the pure-play semiconductor foundry market. In 2025, TSMC generated $122.4 billion in revenue — up 35.9% year-over-year — as demand for AI compute drove an unprecedented surge in orders.
One further constraint; the optical systems required for this manufacturing come from a single German supplier, Carl Zeiss, in a relationship so tightly integrated that ASML holds a significant equity stake in the company. There is no backup, no redundancy, and no alternative supplier.
This is layer one. It is a dual chokepoint — ASML for the machines, TSMC for the chips — sitting at the base of every AI product ever shipped. When people debate whether AI is overhyped, they are rarely accounting for what it costs — in capital, time, and geopolitical exposure — to physically manufacture the hardware that makes any of it possible.
Layer 2: Compute / Cloud Infrastructure — The Layer Being Built at Warp Speed
Above the chip sits the infrastructure required to aggregate its power at scale: GPU clusters, data centers, cooling systems, power grids, and the cloud platforms that abstract all of it away for developers.
This is the layer where money is moving fastest. Goldman Sachs projects AI infrastructure spending will exceed $700 billion in 2026 — a baseline estimate of $765 billion, with actual hyperscaler guidance from Amazon, Microsoft, Google, Meta, and Oracle pointing even higher. Numbers that rival and may surpass the peak of the 1990s telecom buildout.
The key dynamic at this layer is that compute is currently the primary constraint on AI capability. More compute equals more capable models — the scaling laws established by OpenAI and others have held with remarkable consistency. The entities that control compute infrastructure control the ceiling of what's possible at every layer above them.
A developer building an application at layer six is renting compute from someone operating at layer two. That dependency shapes pricing, latency, reliability, and ultimately competitive position. Most application-layer builders treat this as background infrastructure. It is not. It is an active variable.
Layer 3: Data — The Layer That's Already Running Out
AI models learn from data. The quality, quantity, and diversity of training data directly determine a model's capabilities and limits. This layer is less visible than compute but arguably more consequential.
The central tension at the data layer today is exhaustion. The common estimate among researchers is that high-quality human-generated text on the internet — the primary training fuel for large language models — is approaching its limits as a scaling resource. Models have consumed Wikipedia, Common Crawl, GitHub, books, and most of the public web. Synthetic data generation and multimodal data (images, audio, video) are now being explored as the next frontier, but these come with their own quality and bias challenges.
Data is also the layer where competitive moats are most underappreciated. A company with proprietary, high-quality, domain-specific data has a structural advantage that cannot be replicated by simply buying more compute. In specialized verticals — medical records, financial transactions, legal documents, industrial sensor data — the data layer is where durable defensibility actually lives.
Layer 4: Model / Algorithm — The Layer Commoditizing Faster Than Expected
This is the layer most people think of when they hear "AI": the neural network architectures, the training processes, the foundation models. GPT, Claude, Gemini, Llama. The transformer architecture. Reasoning models. Multimodal systems.
The defining trend at the model layer is rapid commoditization. What required a frontier lab and hundreds of millions of dollars in compute two years ago can now be replicated or approximated by open-source models running on consumer hardware. Meta's Llama series demonstrated this clearly: releasing model weights publicly collapsed the moat that closed-source labs assumed they had.
This has profound implications. Model capability itself — the raw intelligence of the AI — is becoming a commodity. Competitive advantage is migrating away from "who has the best model" and toward "who has the best data" (layer three), "who controls compute access" (layer two), and "who has the best distribution" (layer seven). Building your entire business on being marginally better at layer four is a fragile strategy when the gap compresses every six months.
Layer 5: MLOps / Tooling — The Layer That Makes Everything Actually Work
Training a model is not deploying a model. Between the trained weights and a production AI product sits an entire engineering discipline: evaluation frameworks, fine-tuning pipelines, vector databases, orchestration tools, monitoring systems, guardrails, and versioning infrastructure. This is MLOps.
Layer five is unglamorous, coverage of it is sparse, and most non-technical AI commentary ignores it entirely. This is a mistake. The graveyard of AI projects is full of teams that produced impressive demos but could not ship a reliable production system. Latency, cost, drift, hallucination rate, and observability are all layer-five problems. Companies that solve them build compounding operational advantage that is invisible from the outside.
Layer 6: Application — The Layer Where Almost Everyone Is Building
This is the layer that generates most of the startup activity, venture investment, and media coverage in AI. Products, APIs, agents, copilots, wrappers. The application layer is enormous, growing, and genuinely valuable.
It is also the most crowded and the most exposed to risks created by the layers below it. An application that depends on a third-party model API has no control over pricing changes, capability changes, or model deprecations. An application built on a commodity model with no proprietary data or distribution has no structural moat. The application layer creates real value, but it is the layer where strategic clarity matters most and is most frequently absent.
The question worth asking about any AI application is not "does it work?" but "which layers give it defensibility?" If the answer is none, the business is a feature, not a company.
Layer 7: UX / Distribution — The Layer That Determines Who Wins
At the top of the stack is the interface through which real people interact with AI — and the channels through which products reach them. ChatGPT crossed one billion monthly active app users in May 2026, becoming the fastest application in history to reach that milestone — outpacing Google Maps, TikTok, Instagram, and YouTube, according to Sensor Tower data reported by Reuters. That is not a model achievement. That is a distribution achievement.
Layer seven is where habits form, where network effects accumulate, and where the economic value created across all six layers below eventually gets captured. The irony is that distribution is also the layer that requires the least technical understanding to discuss, which is why it dominates AI discourse. Talking about ChatGPT's growth, Apple Intelligence's integration, or the next AI assistant category requires no knowledge of chip fabrication or training data exhaustion. So that is where most commentary lives.
But distribution without a coherent understanding of what you're distributing — and what constraints exist below your layer — is how you build fragile products and make bad investment decisions.
Why Your Layer Determines What You Can See
The single most important insight from this map is that each layer has a radically different vantage point.
Someone operating at layer one thinks about precision manufacturing and supply chain logistics. Someone at layer four thinks about parameter counts and benchmark scores. Someone at layer six thinks about user retention and API costs. They are all talking about AI, and they are barely talking about the same thing.
Most AI opinion is formed entirely within layers six and seven. Most AI risk and opportunity is shaped at layers one through four. This mismatch between where attention concentrates and where leverage actually lives is not a small gap. It is the central source of confusion in how AI is discussed, invested in, and built.
Understanding the full stack does not make you an expert in every layer. But it gives you a map. And without a map, you are not analyzing AI. You are describing the lobby of a very large building and calling it architecture.
Next: Layer 1 in depth — the physical constraints that will determine who wins the AI race over the next decade.
Frequently Asked Questions
What are the 7 layers of the AI technology stack? The seven layers of the AI technology stack, from bottom to top, are: (1) Physical/Silicon — the chip fabrication hardware and semiconductor manufacturing; (2) Compute/Cloud Infrastructure — GPU clusters, data centers, and cloud platforms; (3) Data — training datasets and data pipelines; (4) Model/Algorithm — the neural network architectures and foundation models; (5) MLOps/Tooling — the deployment, monitoring, and operational infrastructure; (6) Application — the AI-powered products and APIs; and (7) UX/Distribution — the interfaces and channels through which users reach AI products.
Why do different sources describe different numbers of AI stack layers? Because different frameworks are answering different questions. NVIDIA's Jensen Huang uses five layers (energy, chips, infrastructure, models, applications) to analyze geopolitical AI competition. Three-layer models (infrastructure, model, application) prioritize simplicity for early-stage founders and investors. This seven-layer framework separates physical silicon manufacturing from cloud infrastructure, and data from models, because those distinctions reveal different types of constraints and different types of defensibility. There is no single "correct" number — the right framework depends on the question you're trying to answer.
Why do most people only talk about the top layers of the AI stack? Most AI commentary focuses on layers six and seven — applications and user interfaces — because these are the most visible and accessible parts of the stack. Discussing ChatGPT's user growth or a new AI product requires no technical knowledge of semiconductor fabrication or model training. The lower layers operate under NDAs, within specialized industrial contexts, and at timescales that don't generate daily news. So that is not where attention goes — even though it is where most of the leverage lives.
Who controls the physical layer of the AI stack? Two companies effectively control the physical layer. ASML is the world's only supplier of EUV lithography systems — the machines required to manufacture advanced AI chips — shipping 48 units in 2025 and targeting more than 60 in 2026, with multi-year lead times. TSMC holds roughly 72% of the pure-play semiconductor foundry market, generating $122.4 billion in revenue in 2025. Carl Zeiss is the sole supplier of the optical systems ASML's machines require. Every major AI chip — including NVIDIA's GPUs — is produced within this tightly constrained structure.
Is the AI model layer a sustainable competitive advantage? Increasingly, no. The model layer is commoditizing rapidly. Open-source models from Meta (Llama) and others have demonstrated that frontier-level capabilities can be replicated and distributed freely, compressing the performance gap between proprietary and open models. Sustainable competitive advantage in AI is migrating toward proprietary data (layer three), compute access (layer two), and distribution scale (layer seven) rather than model capability alone. Building a business exclusively on model superiority is a fragile strategy in a market where the gap between open and closed models narrows every six months.
About the Author
Software Engineer · CEO · Board Member · Investor
Obaid Ghafoori is a software engineer, CEO, board member, and investor who has operated at the intersection of deep technology and business — from engineering at ASML, the world's most critical semiconductor equipment company, to founding and investing in technology startups across multiple industries. He writes about technology, leadership, and what it actually takes to build something that lasts.