There’s a version of the AI story most people are still telling.
It goes like this: the company with the best model wins. The team that ships the most features fastest wins. The product with the most users wins. Just build something useful on top of GPT or Claude, grow fast, and figure out defensibility later.
That story made sense in 2023. It’s becoming dangerous in 2026. And by 2030, the companies that built their entire strategy around it will be looking back, trying to understand what happened.
Here’s the version that actually maps to where things are heading.
The next decade of AI isn’t going to be defined by who builds the smartest model. It’s going to be defined by who understands — and owns — the chain through which intelligence flows. From raw energy and silicon, through data and models, through workflows and trust gates, all the way to memory that compounds over time. That chain is the real competitive battleground. And most product leaders, founders, and investors are still focused almost exclusively on the surface.
The First Era Is Already Over
The first era of generative AI was about access.
Getting access to a capable large language model was genuinely hard in 2022. It required either enormous capital to train your own or early API access to OpenAI’s closed systems. The companies that got in early had a real structural advantage — not because they were smarter, but because they sat at a scarcity point. When something is scarce and valuable, proximity to it is enough.
That era ended faster than almost anyone predicted.
Open-source models — Llama, Mistral, Falcon, and their descendants — collapsed the access barrier inside eighteen months. Competition between OpenAI, Anthropic, Google, and a dozen other labs drove frontier capability costs down by orders of magnitude. What cost hundreds of dollars per million tokens in 2023 costs pennies in 2026. The model layer, as Marc Andreessen noted at the a16z January 2026 LP meeting, is no longer the moat. The moat is what you build around it.
This isn’t an accident or a temporary dip. It’s structural. Intelligence is commoditising downward, the same way every enabling technology before it has. Electricity was once a competitive advantage for manufacturers who had it. Then it became a utility. Computing power was once a moat. Then it became a cloud subscription. The pattern is always the same: the enabling layer gets absorbed, and value migrates to what sits above and below it — to whoever controls the bottlenecks, the data, the workflow, the trust, and the memory.
AI is following the same path. We’re just two or three years into a ten-year transition.
What a Supply Chain Actually Means Here
When most people hear “supply chain of AI”, they think logistics. Chips shipping from Taiwan. Servers arriving in Virginia data centres. That’s one supply chain, and it matters enormously for infrastructure investors. But it’s not what product leaders and founders need to be thinking about.
The supply chain of intelligence is the chain through which value flows in an AI product — from the raw inputs at the bottom to the user outcome at the top.
Think of it the way you’d think about gold mining. Ore in the ground doesn’t become a wedding ring by magic. There’s a chain: extracting the ore, refining it, assaying its purity, transporting it, working it into jewelry, getting it into a store, and finally selling it to someone who wants to mark a moment in their life. Every stage in that chain has different economics, different barriers to entry, and different margins. The miners rarely get rich. The refiners do better. The jewellers capture the relationship. The stores capture the distribution. The brand owns the desire.
AI has the same structure. The framework we’ve built at Supply Chain of Intelligence maps it across 10 layers, from L−1 (the energy and physical resources the whole thing consumes) up through L0 infrastructure, L1 data, L2 foundation models, L3 trust and gatekeeping, L4 access and integrations, L5 execution and domain skills, L6 orchestration, L7 surface interfaces, to L8 memory. Every layer has different defensibility characteristics. Most AI companies, when you map them honestly, sit primarily at L7 – the surface. That’s the most exposed position in the chain.
The next decade will reward the companies that understand which layer they actually own, why that layer is hard to replicate, and what adjacent layers they can reach to compound their position. It will punish everyone who confuses “building on top of AI” with “building an AI business”.
The Four Laws That Will Govern the Decade
Over the course of researching and mapping hundreds of AI companies, four structural laws emerge that hold across categories, company stages, and geographies. These aren’t predictions. They’re patterns that are already visible if you look at the full chain rather than just the surface.
Law I: Intelligence commoditises downward.
Every layer of the AI stack will eventually be absorbed by the layer below it or the platform beside it. This is already happening at L7. Jasper was building an AI writing assistant at a $1.5 billion valuation. When OpenAI and Anthropic shipped comparable capabilities natively, and Microsoft embedded it into Word and Outlook, Jasper’s entire value proposition became a feature of a larger surface. The company didn’t fail because it was poorly built. It failed because it sat on a layer that was always going to be absorbed.
Chegg is a sharper version of the same story. Fifteen years of educational content, packaged at the surface layer, collapsed 99% in market value when ChatGPT made the same layer free. Not because ChatGPT was better at education — it often wasn’t, at first — but because free beats paid when the capability is equivalent enough.
The law applies to every layer over a long enough horizon. But the timeline varies enormously. L7 surfaces commoditise in months. L1 proprietary data takes years or decades. L8 memory, once accumulated, may be practically impossible to replicate. Understanding the commoditisation timeline of the layer you’re on is one of the most important strategic questions any AI company can answer.
Law II: Value accrues at bottlenecks.
This is the oldest law in economics dressed in new clothes. Wherever there is scarcity in the chain, value pools. Right now, the bottlenecks in the AI supply chain are shifting. In 2023-2025, the scarcity was computed — NVIDIA’s position as the near-monopoly supplier of the GPUs required to train and run frontier models was a genuine structural bottleneck. The IEA projects that global data center electricity consumption will more than double by 2030, which means energy — nuclear capacity, transmission infrastructure, and cooling systems — is becoming the next bottleneck layer.
But for application-layer companies, the relevant bottlenecks are different. Proprietary behavioural data that can’t be purchased is a bottleneck. A compliance gate that requires years of regulatory approval is a bottleneck. Institutional memory embedded in thousands of customer workflows is a bottleneck. Distribution relationships that took a decade to build are a bottleneck. These things don’t show up in a product demo, which is exactly why they’re durable.
Law III: The surface captures attention. The chain captures power.
This is the one that trips up the most product leaders, because it cuts against the intuitions that make you good at product management. Beautiful interfaces attract users. Smooth onboarding drives retention. Delightful AI interactions create word of mouth. All of that is real, and none of it builds the kind of structural power that survives a platform shift.
Cursor at $9 billion isn’t valuable because it has a better chat interface than GitHub Copilot does. It’s valuable because it owns the IDE workflow (L5), the project indexing pipeline (L4), the agent orchestration loop (L6), and the accumulated context about your specific codebase (L8). Four layers reinforcing each other. The surface is what users see. The chain is why they can’t leave.
Law IV: Generation and verification must be separate.
Wherever AI output carries weight — legal, medical, financial, regulatory, reputational — the system that generates the output and the system that validates it need to be architecturally distinct. This isn’t a limitation of current AI. It’s a structural requirement of any high-stakes decision-making system, regardless of how capable the generation layer becomes.
This law creates durable value opportunities at L3 — the gatekeeping layer — that are systematically undervalued right now because they’re less visible than the flashy generation capabilities at L2 and L5. Adobe Firefly’s defensibility isn’t primarily that it generates better images than Midjourney. It’s that it generates images with cleared IP provenance, making it the only image model most enterprise legal teams will actually approve. The compliance gate IS the product.
What Actually Shifts in the Next Ten Years
The model wars become infrastructure wars.
The race to build the biggest frontier model will continue, but it will increasingly resemble the race to build the biggest nuclear plant — strategically important, enormously capital-intensive, and relevant to a shrinking percentage of actual business decisions. What will matter more for most companies is the infrastructure layer around models: the orchestration systems (L6) that coordinate multiple agents across complex tasks, the memory systems (L8) that accumulate context over time, and the access protocols (L4) that govern which agents can touch which systems.
Gartner predicts that task-specific AI agent adoption will jump from under 5% in 2025 to 40% of enterprise applications by end of 2026. That’s not a model story. That’s an orchestration and memory story. The companies that own the agent loops, the routing logic, and the accumulated task history will be in structurally stronger positions than the companies that own the prettiest chat interface sitting on top of someone else’s model.
Data becomes the only truly non-fungible layer.
Every other layer in the chain can eventually be replicated with enough capital, talent, or time. Data is the exception — specifically, the categories of data that can’t be purchased, generated synthetically, or scraped from the web.
Proprietary behavioural data (L1c) — the actual recorded patterns of how real users accomplish real tasks — is extraordinarily hard to replicate. Outcome data (L1d) – what actually happened as a result of a decision – is even harder. This is why Bloomberg’s fifty-year accumulation of financial data creates a moat that a $100 billion model investment can’t easily overcome. It’s why John Deere’s sensor data from millions of acres of real farmland is more strategically valuable than any agricultural AI model trained on public data. The data doesn’t just train better models. It defines the problems the models are applied to in ways that require real-world presence to accumulate.
The companies that will dominate the next decade are, right now, building data flywheels — systems where usage generates data, data improves the model, improved models drive more usage, and more usage generates more data. Klarna’s 700-agent customer service deployment isn’t primarily a cost story (though the $40M in reported savings is real). It’s a data accumulation story. Every resolved interaction becomes training signal. The gap between Klarna’s model and a generic customer service model widens every quarter, invisibly, at the data layer.
Memory becomes the new lock-in.
Enterprise software built its moats on switching costs — get your data into our system, and the pain of moving it somewhere else is prohibitive. AI is creating a deeper version of the same dynamic through memory.
L8 memory — the accumulated context about a specific user, organisation, workflow, or domain — is the layer that makes AI systems genuinely hard to replace. Not because the competitor couldn’t build a comparable model, but because the competitor would be starting from zero on context that took months or years to accumulate. Sierra’s customer service AI compounds with every resolved ticket. Glean’s enterprise search gets more useful every time a document is accessed, a query is answered, and a preference is recorded. The switching cost isn’t migrating data. It’s losing the institutional memory that makes the system actually know your business.
This dynamic will intensify over the next ten years as AI systems become more deeply embedded in organizational workflows. The companies that understand this — and architect their products around accumulating irreplaceable institutional memory rather than delivering impressive demos — will build the most durable positions in the AI economy.
Trust gates will be worth more than model capability.
The most underappreciated structural shift of the next decade is the rising value of L3 — the gatekeeping layer that controls whether AI output can be trusted, approved, and deployed in high-stakes contexts.
EU AI Act provisions become fully applicable in August 2026. Forrester predicts 60% of Fortune 100 will appoint dedicated AI governance heads this year. The regulatory pressure on AI deployment in healthcare, finance, legal, and government is accelerating, not retreating. Every constraint on deployment is a structural advantage for the companies that already operate inside those constraints — because it raises the bar for competitors to enter.
The companies that will win in healthcare AI aren’t the ones with the best clinical reasoning models. They’re the ones with the HIPAA compliance infrastructure, the FDA approval pathways, the physician workflow integrations, and the liability frameworks that let a health system actually deploy the model without ending up in a lawsuit. That’s an L3 play, not an L2 play. The model is table stakes. The gate is the business.
The Defensible Triangle — and Why Most Companies Are Building Outside It
At Supply Chain of Intelligence, the most consistent pattern we’ve found across AI companies with durable positions is what we call the Defensible Triangle: owning L1b proprietary data, L5 deep domain execution, and L8 compounding memory simultaneously.
Each corner reinforces the others. Proprietary data trains better execution models. Better execution generates more behavioral data. Accumulated memory makes the system harder to leave. Together, they create a flywheel that compounds faster than a competitor starting from zero on any single layer.
Harvey AI demonstrates this in the legal vertical. It sits on L1b (licensed case law and matter history), L5b (legal reasoning scaffolds built around actual attorney workflows), and L8d (institutional memory of specific client matters). A generic legal AI built on the same foundation model is categorically different from Harvey — not because of the model, but because of three years of accumulated layer ownership.
Bloomberg has owned this triangle in financial data for decades, before “AI” was even the right word for what they were doing. The data nobody else can buy. The terminal workflow that every professional in the industry knows. The accumulated history makes switching unthinkable.
Most companies building in AI right now are building at L7 only – the surface. They have a product-market fit signal, genuine user value, and a good team. What they don’t have is a layer that compounds. When the platform below them ships the same capability for free, there’s nothing left.
This is the strategic question of the next decade: not “how do we build a better AI product?” but “which layer are we actually building on, and what does it look like when the layer below us gets smarter?”
What This Means If You’re Building Right Now
A few concrete implications, not for 2035, but for decisions you’re making this year.
Name your layer before you name your product. Before you write a line of code or design a single screen, be able to answer: which layer of the intelligence supply chain are we actually owning? If the answer is “We have a great interface on top of GPT”, that’s an L7 answer, and you need to be honest about what that means for your defensibility horizon.
Treat data as the product, not the input. The companies that will have the strongest positions in 2030 are, right now, making deliberate choices about what data they accumulate, how they accumulate it, and what structural advantages it creates. That’s not an engineering decision or a data science decision. It’s a product strategy decision that belongs in the board conversation.
Design memory as an architecture, not a feature. Most product teams think about memory as a feature — “remember my preferences” or “recall previous conversations”. The companies building durable positions think about memory as architecture — a system that gets structurally better with use and creates irreversible advantages over time. That distinction changes everything about how you build.
Don’t confuse distribution with defensibility. Growing fast is real. Users loving you. High NPS is real. None of it is a moat if the platform below you can replicate your value proposition in a release cycle. Distribution buys time. Layer ownership buys survival.
The Real Race
The first decade of modern AI — roughly 2017 to 2026 — was a race to prove that intelligence at scale was possible and commercially valuable. That race is settled. The answer is yes.
The next decade is a different race. It’s not about whether AI can do impressive things. It’s about who controls the chain through which intelligence flows from raw resources to real outcomes — and captures the value that chain creates.
The gold analogy at the supply chain of intelligence holds for a reason. In the original California gold rush, the people who got rich weren’t primarily the miners. They were the people who owned the railroads that moved the ore, the assay offices that verified its purity, the supply companies that sold the shovels, and the banks that stored the wealth. The surface of the rush attracted attention. The chain captured the value.
AI is the same. The companies that win the next ten years aren’t the ones with the shiniest demos at the surface. They’re the ones who understand, map, and own their position in the intelligence supply chain — and build from there.
That’s the map we’re building at supplychainofai.com. Not because the framework is elegant, but because the decisions that follow from it are the ones that will actually matter.
Anand Arivukkarasu is an Ex-Meta (Instagram) Product Leader and the creator of the Supply Chain of Intelligence™ framework — a 10-layer defensibility map for AI companies. The framework and all case studies are free at supplychainofai.com.
