The AI App Layer Thesis - Who Captures the Value?

Back in early 2023, a16z published the reference essay on the whole 'app layer' thesis as it pertains to the AI boom - Who Owns the Generative AI Platform?

The stack can be divided into three layers:

• App Layer - the 'generative' AI Apps that people interact with on the front end; they either run their own model pipelines or rely on 3rd party APIs (OpenAI, Claude, etc)
• Models - the frontier models (Astra, Fable, Grok, open-source, etc.) that are made available through various API endpoints
• Infrastructure - vendors who run training, inference, and scaling infrastructure for the AI Apps built on top of them

Three years after the a16z essay was written, the core thesis remains hotly contested; it's also clear that AI as a whole is only in the early innings, and the pace of innovation across models, applications and agents is unlike anything we have seen in history. Furthermore, the leading model companies - Anthropic and OpenAI - are on the cusp of IPOs projected to be in the $Trillions, putting a public market value on the top end of the market for the first time.

The original Lumos post (Generative AI Business Model), written almost exactly two years ago, was a simple analysis on the same premise - at that moment in history, very little of the app layer was visible to end users, so it was a question of where the value was accruing. It's clear now, in hindsight, that the value accrued to the hardware companies, neoclouds, and hyperscalers first.

Now, however, the tools to build virtually any type of application are available to anyone. Claude Code, Codex, and Cursor have kicked off a dev frenzy with the inception of 'vibe coding,' where traditional coders and non-coders alike can now build end-to-end applications by prompting. Open-source models have started to catch up, creating cheaper onboarding ramps and the development of several products such as Open Router, Open Code, and Hermes for development across multiple models. The app layer is being built with haste.

The counterclaim to the thesis is that the labs themselves are moving up the stack, and this claim has been partially vindicated. The frontier labs have moved up the stack with their own apps: OpenAI has launched Codex and several iterative experiments in commerce as one example, like Instant Checkout (since walked back); Anthropic has launched Claude Code, Cowork, and most recently in commerce, Claude Commerce Agents.

The potential value capture and opportunity in the app layer is enormous, as demonstrated by a slew of companies such as:

• Cursor- coding, acquired by SpaceX for $60 Billion in an all-stock deal
• Harvey - legal, estimated $350M ARR according to Sacra
• Glean - enterprise search agents, estimated $300M ARR according to the company

These are just a few examples of the billion dollar companies that have gone parabolic in the last couple of years by building their own apps. And the gold rush continues, so let's dig into the business model mechanics that underpin the app layer.

The Margins and Mechanics of the AI App Layer

The typical reference point to the 'app layer' is Saas (software-as-a-service). The Saas giants were built in the previous decade and many have done extremely well given the back-of-the-napkin math for most Saas companies - huge gross margins (80 - 90%) and high switching costs, meaning retention stays high.

The mechanics of apps in the AI layer are different, since they rely on frontier model intelligence via APIs and must pay for inferences as a COGS (cost of goods sold).

The reference numbers for the thesis, from the original a16z essay, showed "there's a wide range of gross margins — as high as 90% in a few cases but more often as low as 50-60%, driven largely by the cost of model inference." App companies spent between twenty and forty percent of revenue on inference and per customer tuning. The TLDR is less margin, but much faster growth rates.

Since those early estimates in 2023 by a16z, there is some variance in these numbers due to the fact that the number of infrastructure and model opportunities has increased exponentially. What matters to the app thesis is that the companies building out these apps capture the core customer workflow and governance layer surrounding that.

If we are talking about enterprise agents, for example, it's not just the creation and implementation of the agent in the enterprise, but the full management of the tech and human elements in the workflow. The customers expect the result, and the entity delivering the solution answers to the core customer metrics, it doesn't simply provide the tools.

In terms of margin and capture, Iconiq's 2026 'The Execution Era of AI' report cited gross margins as 52% on average, and inference as 23% of COGS at scale. Jevons paradox states that while, yes, inference costs are dropping rapidly, usage grows exponentially, which makes the net outlay grow. This is captured by the frontier labs themselves, and feeds into their trillion dollar valuations.

Since these companies are in the process of fine tuning their business models, they vary in their current state and are subject to iteration: end product pricing is typically subscription (52%), while consumption (39%) and outcome-based (18%) continue to grow according to Iconiq.

Low gross margins, compared to Saas and the typical software model, are not a bellwether for whether the business model itself is sustainable, since these companies are growing exponentially and have the ability to turn on multiple levers to capture additional value in the future.

The Labs - Moving up the Stack?

As mentioned above, the frontier labs such as Anthropic and OpenAI have started themselves moving up the stack to build out apps and becoming vertically integrated across certain verticals.

Claude Code and OpenAI Codex are examples of applications that literally harness the core capabilities of the models for the purpose of coding.

Cowork and ChatGPT Work are applications aimed broadly at enterprise productivity, while ChatGPT and Google's Gemini remain core to the consumer app stack for search and related activities.

Given that OpenAI recently raised $122B in Q2 '26 and are slated to be on the cusp of an IPO in early 2027, while Anthropic is gearing up for its own IPO at a $1 - $2 Trillion valuation, the labs have the cash to experiment. Much like Claude Code and Codex have ignited and uprooted the entire coding/development model, we can expect more spontaneous innovations from the labs themselves in different verticals.

What we don't know about the labs yet, given the absence of public disclosures and S-1s, is their numbers on the models. While they are making tens of billions in gross revenue - $65B for Anthropic, $40B for OpenAI (August '26 estimates) - they are also simultaneously investing into compute and capacity.

The app layer itself is the future revenue layer of many industries, we are in a gold rush. It makes sense that the landscape will continue to rapidly evolve and blur the lines across industries. Regulation will likely also play a large factor in how the use of AI evolves in many important verticals due to the inherent social and economic conflicts it creates; whether or not labs remain 'trusted' intermediaries in these industries may reduce their capacity to innovate on the app layer itself.

Claude Commerce Agents and the CPCT

Anthropic's recent launch of Commerce Agents is a live test of the whole post.

What are Commerce Agents?

• Commerce brands and platforms such as Shopify are building 'shopping agents' and 'merchant agents' that work with the customer on the brand's website to reduce purchasing friction
• Claude Commerce Agents act as a blueprint with the harnesses, patterns, and guardrails that teams need to embed them on a company's site
• Shopping agents can fetch items from the catalog, remember customer preferences, serve them a product comparison, build the cart, and hand it to checkout
• Merchant agents support the people running the store and can help them track inventory, recommend pricing, and draft marketing campaigns
• Anthropic has created the agent rails but not the wallet, so checkout stays with the merchant in this case

The CPCT (Cost Per Completed Task) was built to help test the creation of these types of agents in a workflow from a unit economics perspective. The thesis of the site is to not only write about these types of developments, but test them.

In Agentic Modelling - What Does it Cost to Run an Agent? I outlined the general why behind the CPCT and how it works:

• Agents run off of API credits, which is billed based on the agent loop for completing a specific task
• Those loops can sometimes become complicated and highly variable over time, which makes the cost predictability difficult to assess without running a real test
• The CPCT simulates that test and traces the agent tasks in dollar terms, creating a benchmark through simulation of multiple tasks for the cost to run an agent
• The results showed the engineered stack (caching, compaction, etc) at about 10% of the naive stack's cost per task

The CPCT version of Commerce Agents is coming next.

The App Layer Thesis

Claim	Verdict	Evidence	What would flip it
Models commoditize and prices fall	Confirmed	Equivalent capability roughly 10x cheaper per year (a16z, "LLMflation"); the frontier converged by mid-2026	A lab opens a durable capability gap nobody closes for six-plus months, with pricing that resists deflation
Infrastructure concentrates	Confirmed	Nvidia, the three clouds, and a handful of labs still take the majority of dollars	Open-weight models on commodity hardware genuinely decentralize inference
The labs stay out of applications	Refuted	Claude Code, Codex, Cowork, Instant Checkout (launched, then walked back), and the Claude Commerce Agents blueprint (September 2, 2026)	Nothing, fact
App-layer revenue is real	Confirmed	Cursor roughly $4B ARR by mid-2026 and a ~$60B acquisition; Harvey, Glean past $100M–$450M ARR — all company- or press-reported, none audited	A wave of these companies decelerating or being absorbed after a lab ships their core capability
App-layer margins are durable	Too early	~52% gross margin in ICONIQ's 2026 survey (survey figure) versus 75%+ for classic SaaS	Margins climbing toward 60%+ as inference deflates, or staying pinned as usage growth eats the savings
Distribution and data are the moat	Moved	Distribution alone is now a lab advantage; what survived is proprietary workflow data loops plus governance in regulated verticals	Models adding native memory and workflow context that erodes the data-loop advantage from inside

Verdicts as of September 2026; each row's flip condition is its tripwire.

Right now, the consensus in the market is that both the apps themselves and the frontier labs can win.

As outlined in Avoiding Death on the Yellow Brick Road by a16z, the labs will win because as they move horizontally, they own distribution. The apps can win because they own the data capture, workflow, and governance within the specific vertical or customer subset; new models can be swapped in as they become available, so labs launching new models increases what they can offer.

Regulation may end up being the app layer's biggest moat, rather than a constraint, since those companies are building out the governance layer. The governance layer is what makes the AI model's output permitted within the specific vertical, whether it be finance, health or legal. This is where the app companies themselves can step in and work with regulators, and where the labs have declined to build.

On top of that, a whole new generation of builders are only learning and experimenting with the tooling, the real ripple effects across multiple industries with AI agents and apps has only just begun.

Overall, the app layer thesis is alive and well, but it isn't over yet. It's a barbell, where the distribution is weighted heavily at both ends and the little is left in the middle. On one end, you have the labs, who have the horizontal path; they can continue to plug their frontier models into specific tasks and use their distribution power to scale. On the other end, is the messy, complex vertical path where the apps live; here, the scaffolding and guardrails around the workflow are more important than the model itself. The only way to win is to know your model inside and out, and in this era that includes inference and agentic numbers modelled down to the workflow and specific tasks.