Foundation models are commoditizing. Six organizations ship competitive frontier models, and capabilities leapfrog every few weeks. Ben Evans argues there is “no mechanic we know of for one company to get a lead that others could never match.” ChatGPT has 800-900M users but only 5% pay and engagement is shallow. The value in AI isn’t the model layer — it’s the vertical agent layer that turns raw capability into domain-specific execution.

Key Takeaway: The best AI model doesn’t win. The best distribution and domain expertise wins. Foundation models are becoming commodities like TSMC chips — essential but undifferentiated. The value accrues to vertical agents that encode specific workflows (legal, CX, marketing) on top of whichever model is best this month.

Benedict Evans — one of the sharpest technology analysts alive — just published “How will OpenAI compete?”. His answer, essentially: it’s not clear that it can.

The piece is nominally about OpenAI’s competitive position. But read through the lens of where AI value actually accrues, it’s a roadmap for why vertical agents — not foundation models — will capture the market.

The Moat That Doesn’t Exist

Evans’ opening argument is devastating. Six organizations now ship competitive frontier models. Capabilities leapfrog every few weeks. There is, in his analysis, “no mechanic we know of for one company to get a lead that others could never match.”

ChatGPT has 800-900 million users. Sounds like a moat. It isn’t.

Only 5% of users pay
80% of users sent fewer than 1,000 messages in all of 2025 — roughly 3 prompts per day
US teens use chatbots “a few times a week or less” rather than multiple times daily

Evans calls this “very narrow engagement and stickiness, and no network effect.” The users are aware. They’re not hooked. And the data shows something uncomfortable: if someone can’t think of anything to do with ChatGPT today, a better model won’t fix that.

The Netscape Problem

Evans draws a parallel that should make every AI lab nervous: Netscape vs. Internet Explorer.

Netscape had the best browser. It had the users. It had the early-mover advantage. Then Microsoft shipped IE with Windows, and distribution won. Not because IE was better — because it was already there.

Google is doing the same thing with Gemini. Meta is doing it with Meta AI. Both have existing distribution — search, social, messaging — that lets them deliver equivalent AI capabilities without asking users to switch products.

Evans notes that Anthropic’s Claude “regularly scores at the top of benchmarks” but has “close to zero consumer awareness.” The best model doesn’t win. The best distribution wins.

This is the foundation model layer in a single chart: commoditizing capability competing against entrenched distribution.

Where Value Actually Accrues

Here’s where Evans’ analysis gets interesting for anyone building on top of foundation models. He invokes the TSMC analogy:

TSMC has a monopoly on cutting-edge chip manufacturing. It provides “little to no leverage further up the stack.” People built Windows apps, not TSMC apps. The infrastructure layer captures value in its own right but doesn’t extend that value upward.

The same is happening with foundation models:

Layer	Examples	Competitive Dynamic
Infrastructure	NVIDIA, TSMC, cloud providers	Capital-intensive, oligopoly, defensible
Foundation models	OpenAI, Anthropic, Google, Meta	Commoditizing, no moat, leapfrogging
Application layer	Vertical agents (Harvey, Sierra, Lane)	Domain expertise, workflow integration, sticky

Evans’ key observation: when developers call foundation model APIs, “customers don’t know or care what model you used.” Users of Snap on AWS versus GCP never know the difference. The model is the engine, not the product.

This maps to what a16z’s George Sivulka argued in “In Defense of Vertical Software”: the value isn’t in the model. It’s in the process engineering — knowing how to apply the model to a specific domain’s actual workflow.

The Capability-Usage Gap Is the Opportunity

Evans identifies something crucial: there’s a massive gap between what AI models can do and what people actually do with them.

Most ChatGPT users send a few messages a week. The models can write code, analyze data, compose emails, plan projects, research markets. But users treat them like a slightly fancier search engine.

Why? Because a chatbot interface doesn’t do the work. It helps you do the work. Evans puts it precisely: “if you can’t think of anything to do with ChatGPT today, that won’t change if you give them a better model.”

This is the copilot ceiling from a different angle. The model is capable. The interface — a text box where you type prompts — limits what people actually accomplish with it.

Vertical agents solve this by removing the user from the loop. You don’t prompt Lane to write a blog post. Lane identifies that a blog post should be written, writes it, formats it for the target channel, distributes it, and measures the result. The capability-usage gap closes because the agent uses the capability, not the user.

Approach	User’s Role	Capability Utilization
ChatGPT/Claude (horizontal)	Prompt, review, execute manually	Low — limited by user’s imagination and time
AI copilot (integrated)	Direct, review, approve	Medium — limited by user’s availability
Vertical agent (autonomous)	Set strategy, review results	High — agent utilizes full model capability continuously

The “New Experiences” Question

Evans asks the most important strategic question in AI right now: who invents the “second step of generative AI experiences”?

The first step was the chatbot. Ask a question, get an answer. That’s been commoditized.

The second step — the one that creates real value — is purpose-built applications that use AI to do specific jobs. Not “ask AI a question about marketing” but “AI runs your marketing.”

Evans notes that “the entire tech industry is trying to invent” these new experiences. And he points out that OpenAI can’t invent all of them. Neither can Anthropic or Google.

This is fundamentally a bet on the application layer. The foundation model providers build the engine. Thousands of vertical companies build the vehicles — each one tailored to a specific job, a specific workflow, a specific industry.

In marketing, that vehicle needs to:

Understand 19 distinct traction channels
Know which channels work for which business type
Create channel-specific content (not generic copy)
Distribute across platforms with native formatting
Measure results and reallocate effort
Do all of this autonomously, without a marketing team

No foundation model does this out of the box. It requires the vertical layer — the process engineering, the domain knowledge, the orchestration logic. That’s where the value is.

Evans’ Implicit Prediction

Evans doesn’t quite say it explicitly, but his argument leads to a clear conclusion:

Foundation models commoditize. Distribution incumbents (Google, Meta) absorb the horizontal use cases. Vertical agents capture the domain-specific value.

This is the same pattern that played out with cloud computing. AWS, Azure, and GCP commoditized infrastructure. The value moved to applications built on top: Salesforce, Shopify, Stripe. The infrastructure providers made money. The application companies captured disproportionate value.

AI is following the same arc. OpenAI, Anthropic, and Google are the AWS/Azure/GCP layer. The vertical agents — Harvey for legal, Sierra for customer service, Lane for marketing — are the Salesforce/Shopify/Stripe layer.

Evans’ question — “how will OpenAI compete?” — has an answer. OpenAI will compete at the foundation layer, where margins compress and commoditization is relentless. The real question is: who wins at the application layer?

For marketing, we think we know.

Source: How will OpenAI compete? — Benedict Evans, February 2026

References

How will OpenAI compete? — Benedict Evans, February 2026
Related: a16z Says the Last Mile Is the Entire Problem. — Why marketing AI must be vertical
Related: Anthropic’s Marketing Team Saves 100+ Hours With Claude. Here’s What They Still Can’t Automate. — The copilot ceiling
Related: Dario Amodei Predicts a Country of Geniuses in a Data Center. — The capability exponential
Related: The 19 Traction Channels, Explained — Domain-specific channel framework
Related: Software Is Eating Marketing Labor — a16z’s broader take