Since the launch of ChatGPT in late 2022, capital has poured into artificial intelligence faster than into almost any technology in modern history. Hyperscalers are spending hundreds of billions of dollars a year on data centers, foundation model labs are raising rounds at valuations that rival public companies, and thousands of startups are building on top of the same handful of models. For investors, the question is no longer whether AI will create value, it clearly will , but who will keep it.

This report looks at the AI value chain in four layers: (1) foundation models, (2) applications, (3) integrators and services, and (4) proprietary data and workflows. For each one we ask a simple question taken from classic strategy: is there a source of advantage that competitors cannot easily copy, and that lets the company keep pricing power over time?

Our main conclusions:

For private equity investors, especially in the lower and core middle market, the implication is encouraging: the most durable AI moats tend to sit inside established businesses with proprietary operational data, entrenched customer relationships, and complex workflows. Those are exactly the kind of assets PE already owns or buys. The opportunity is to use AI to deepen those advantages, not to compete with model labs.

Every major technology wave goes through the same debate. In the early internet, investors argued over whether value would go to the network providers, the browser makers, or the websites. In cloud, the question was infrastructure vs. SaaS. In both cases, the answer was not obvious at the start, and investors who backed the wrong layer lost a lot of money even while the overall technology succeeded.

AI is following the same pattern, with a bigger price tag. Sequoia Capital's David Cahn framed the problem in his 2024 essay "AI's $600B Question": if you take the annual capital expenditure on AI infrastructure and estimate the revenue needed to earn a reasonable return on it, there is a large gap between that required revenue and the AI revenue actually being generated today. The gap does not mean AI is a bubble; it means that someone in the stack is going to earn less than they are currently spending.

This makes the moat question practical, not academic. For a venture investor, it decides which layer to back. For a growth or PE investor, it decides which portfolio companies are threatened by AI and which can use it to strengthen their position. For an operating company, it decides whether to build, buy, or partner.

We use Hamilton Helmer's 7 Powers framework as a reference. Helmer argues that a business has durable advantage only when it has both a benefit (higher prices or lower costs) and a barrier (something that stops competitors from copying it). His seven powers are scale economies, network economies, counter-positioning, switching costs, branding, cornered resource, and process power.

Applied to AI, the question becomes: which of these powers are available at each layer of the stack? As we will see, the answer differs a lot by layer, and the most tempting sources of advantage (technical superiority, being first) are often the least durable.

Andreessen Horowitz's 2023 essay "Who Owns the Generative AI Platform?" offered one of the first clear maps of the generative AI stack. Its early conclusion was that infrastructure vendors were capturing most of the money, while application companies were growing fast but struggling with retention, differentiation, and gross margins, and model providers had not yet shown durable lock-in. Much has changed since then, but the structure of the stack remains a useful starting point.

For this report we simplify it into four layers:

These layers are not fully separate. Model labs are moving into applications (ChatGPT Enterprise, coding agents), applications are building their own models, and integrators are productizing their services. But the separation helps clarify where each type of advantage comes from.

Where is the money going today? Menlo Ventures' annual "State of Generative AI in the Enterprise" survey has tracked enterprise spending across layers since 2023. The general trend in these surveys has been a fast increase in total enterprise spend on generative AI, with a growing share going to the application layer as companies move from experimentation to deployment.

The bull case for model companies rests on scale economies. Training a frontier model requires enormous compute, rare research talent, and billions of dollars. Epoch AI's data show that the compute used to train frontier models has grown by several times per year over the last decade, and training costs for the largest models now run into the hundreds of millions of dollars. Only a handful of companies, backed by hyperscalers or sovereign-level capital, can keep up. In theory, this should create an oligopoly with strong pricing power.

Model companies also benefit from distribution and brand. ChatGPT became one of the fastest-growing consumer products in history, and a brand that becomes a verb has real value. Enterprise relationships, safety track records, and developer ecosystems add some switching costs.

The evidence over the last three years points the other way for anyone who is not at the very top of the frontier.

First, the price of intelligence is collapsing. Andreessen Horowitz's "LLMflation" analysis (2024) estimated that the cost of using a model with a given level of capability has been falling by roughly 10x per year. Data from Artificial Analysis, which tracks price and quality across model providers, shows the same pattern: performance that cost premium prices a year ago is now available for a small fraction of the price, often from several competing providers.

Second, open-weight models have closed most of the gap. The Stanford HAI AI Index has documented how the performance difference between the best closed models and the best open-weight models on major benchmarks has narrowed dramatically. In early 2025, DeepSeek's R1 model showed that a lab with far less capital could reach performance close to the frontier, which briefly erased hundreds of billions of dollars of Nvidia's market value and forced the market to question how much scale really protects incumbents.

Third, the insiders themselves have doubted it. The leaked 2023 Google internal memo, "We Have No Moat, and Neither Does OpenAI," argued that open-source models were improving so quickly that proprietary labs would struggle to maintain a lasting lead. Not everything in the memo proved true, but its central point — that model capability alone is hard to defend — has held up well.

The model layer looks less like a classic software business and more like a capital-intensive industry such as semiconductors or airlines: huge fixed costs, fast depreciation of the core asset (a model is state-of-the-art for months, not years), and price competition among a few large players. A small number of frontier labs may still earn very large revenues, particularly if they build strong consumer brands and enterprise distribution. But for most investors, and certainly for almost all portfolio companies, competing at the model layer is not a realistic strategy.

The positive side is important: commoditization of models is good news for everyone else. Every drop in model prices lowers the cost base of applications, integrators, and companies using AI internally. Value tends to shift to the layer that controls the customer.

The application layer has produced the fastest revenue ramps in software history. Stripe and several venture firms have reported that leading AI-native companies are reaching $1 million and then $30 million or more in annualized revenue much faster than the typical SaaS company of the previous decade. Coding tools such as Cursor, legal platforms such as Harvey, and medical documentation tools such as Abridge became category leaders in a few years.

But speed of growth is not the same as durability. Bessemer Venture Partners' "State of AI" report distinguishes between "Supernovas" — companies that grow extremely fast but often with thin margins and fragile retention — and "Shooting Stars," which grow more like excellent SaaS companies, with solid gross margins and strong retention. The first group shows that demand exists; only the second shows that a moat exists.

An application that simply puts a user interface on top of a third-party model faces three threats at once:

The winning applications share a few characteristics that map well onto Helmer's powers:

The key insight is that, at the application layer, AI is a feature that strengthens an existing moat, not a moat by itself.

There is a large gap between what AI can do in a demo and what it delivers inside a real organization. MIT's NANDA initiative, in its 2025 report "The GenAI Divide: State of AI in Business," found that the large majority of enterprise generative AI pilots — around 95% by their measure — had not produced measurable P&L impact. The report also found that companies that bought specialized solutions or partnered with external vendors succeeded more often than those that tried to build everything internally.

McKinsey's annual "State of AI" survey shows a similar picture: adoption of AI is widespread across business functions, but only a minority of organizations report a significant impact on enterprise-level EBIT. The constraint is rarely the model. It is data quality, process redesign, change management, integration with legacy systems, and trust.

Who captures the implementation value?

This gap creates an opportunity for three kinds of players:

Traditional services businesses are not known for strong moats: they scale linearly with headcount, margins are modest, and talent walks out the door every evening. AI changes some of this. An integrator that standardizes its delivery into reusable software, accumulates domain-specific data from many deployments, and becomes embedded in client operations can build real switching costs and process power.

The risk is that AI also makes services cheaper to deliver, which can pressure pricing for traditional consultancies. The integrators most likely to build lasting advantage are those that are vertical (deep in one industry), outcome-priced, and that turn their repeated work into proprietary tools and data.

For private equity, this layer is especially relevant. Many lower-middle-market service businesses — accounting, insurance brokerage, IT services, healthcare administration, logistics — can be transformed by AI into higher-margin, more scalable businesses. Buying a service company with strong customer relationships and applying AI to its delivery model is one of the clearest value creation plays available today.

If models become a commodity, the main thing that differentiates one AI system from another is what it knows and where it is embedded. Public internet data is now largely used up as a source of differentiation: every major lab has trained on roughly the same corpus. Epoch AI's research on data limits projects that the stock of high-quality public human-generated text could be fully used for training within this decade.

The market is already putting a price on proprietary data. Reddit's data licensing agreement with Google was reported at around $60 million per year, and News Corp's multi-year deal with OpenAI was reported at more than $250 million over five years. These deals show that data that cannot be scraped from the open web has real economic value.

A balanced view needs the counterargument. In "The Empty Promise of Data Moats" (2019), Andreessen Horowitz argued that most data does not create a lasting advantage. Data often has diminishing returns: the first million examples improve a model a lot, the next ten million much less. Data can go stale, can be acquired or generated by competitors, and the cost of cleaning and maintaining it can be high. Synthetic data and better models also reduce the amount of data needed to reach good performance.

The conclusion is that a static database is not a moat. Owning a large archive is helpful but rarely decisive.

Data becomes a durable advantage when it has several of the following properties:

In short, the moat is rarely the data alone. It is the combination of proprietary data, the workflow that keeps generating it, and the customer relationship that makes it hard to leave.

The table below summarizes our assessment of each layer using the 7 Powers lens.

Two patterns stand out. First, durability tends to increase as we move closer to the customer and the real-world process, and away from general-purpose technology. Second, the most durable positions combine layers: a vertical application that owns a workflow, generates proprietary data, and is delivered with a strong implementation capability is much harder to displace than a company strong in only one layer.

The moat analysis is especially favorable for private equity, particularly in the lower and core middle market:

A useful set of diligence questions for any company:

Any analysis of AI moats has to stay humble, since the technology is changing quickly. Key uncertainties include:

The AI moat is unlikely to live where most of the capital is going. Foundation models absorb enormous investment but face some of the fastest commoditization ever seen in technology. Thin applications grow quickly but are easy to copy. The more durable advantages appear closer to the customer: in vertical applications that own workflows, in integrators that turn implementation know-how into repeatable products, and above all in proprietary data that is continuously generated by real-world operations.

For investors, this suggests a simple rule: follow the workflow and the data, not the model. For private equity in particular, the AI era is less a threat than an opportunity. The assets that create durable advantage — proprietary data, entrenched customer relationships, and complex operational processes — are already present in many established middle-market businesses. The winners will be the investors and operators who use increasingly cheap and capable AI to deepen those advantages.