AI's State of the Union: Application and Service Layers

Writer RAG tool: build production-ready RAG apps in minutes

• Writer RAG Tool: build production-ready RAG apps in minutes with simple API calls.

• Knowledge Graph integration for intelligent data retrieval and AI-powered interactions.

• Streamlined full-stack platform eliminates complex setups for scalable, accurate AI workflows.

Learn more about our production ready RAG tooling here.

FUTURE
Your Full Guide to AI Investing

Last week, we explored the sweeping changes coming to AI in 2025, focusing on hardware, infrastructure, and models. This week, we shift to the application and service layers—the ones poised to disrupt the status quo and challenge investors thanks to the emergence of ‘AI wrappers’.

These businesses, some of the fastest-growing companies in history, are the hot new form factor for start-ups, one that takes you from 0-to-1 (millions in revenue in days or weeks) much faster than at any time in history.

But most wrappers are actually death traps that won’t scale despite exceptional initial traction. They are, essentially, born to die (early).

All in all, today, you’ll learn:

• What are wrappers, and why you should be equally excited and scared about them

• How to categorize AI start-ups according to their wrapper level

• The unit economics of start-ups in an AI world

• Why most AI start-ups won’t make it, and how to pick the winners & avoid the losers using my AI-adjusted Value Formula, a non-technical, simple-to-apply way of evaluating AI companies

• And the Big Service Opportunity. Why are most investors looking in the wrong direction when identifying the most significant opportunity in AI.

Let’s dive in!

A Wrapped World

The vast majority of start-ups today, especially those that leverage AI (the majority, at least according to their pitches), are, in some shape or form, AI wrappers.

But what does that mean? And why is it such a crucial detail?

AI-Native, But at What Cost

Until further notice, the truth is that the most powerful AIs in the world are not public. Sure, DeepSeek-R1, the model we saw on Friday, gives everyone hope that these tools will continue to be commoditized and could be stored in the safety of your organization, but the rapid progress of the field will force you to readapt your AI stack every few weeks, months at best.

And even if you plan to use open-source models, you will still be required to run them outside your organization (at least physically) unless you own a multiple-million state-of-the-art GPU cluster in a data center site with at least 40 kW per-rack power capacity (if you aren’t knowledgeable in AI energy discussions, you can read this piece, but I’ll save you the hustle: your company most likely does not have such a data center or the capacity to build it).

Long story short, you’re using technology developed by third parties for most of your AI use cases. And, as open-source will require you to be a much more mature organization, most companies trying to become—or are born as—AI-native opt for a much simpler option: outsource the AI from private companies like OpenAI, Anthropic or Google, becoming what most people in the industry know as wrappers.

But what exactly makes you a wrapper?

To me, a company becomes an AI wrapper when the outsourced AI is a part of your product or service. Or, to be more specific, the moment your AI API costs are part of your COGS (cost of goods/services sold). If your outsourced AI costs are OPEX, aka operational, you aren’t an AI wrapper; you’re using AI to cut costs, make employees more productive… etc.

Of course, this generates a certain degree of dependency on that party, and companies must be categorized accordingly:

1. Level 1 wrapper: You use AI to reduce costs or simply as a productivity tool for your employees. You’re not really a wrapper, but you’re still using third-party AI.

2. Level 2 wrapper: Some of your revenue-driving features are based on outsourced AI technology. Without them, your product loses some of its appeal.

3. Level 3 wrapper: What most people think of when referring to a wrapper. Your entire company is built on outsourced AI, which usually acts as the backend of your product or service. Without this AI, there’s nothing.

Independently of your wrapper type, all of their financials are considerably influenced by these third-party AIs.

And that’s a huge problem.

The Costnomics of Wrappers

First of all, it’s important to understand that outsourced AI costs aren’t cheap and take a considerable toll unless the wrapper company engages in highly active AI engineering efforts.

In fact, their costly nature could even lead to you adapting your entire business model, as we’re seeing with Salesforce (even if they aren’t a wrapper as they run on in-house AI).

• If you’re a Level 1 wrapper, you may have some features in your product with outsourced AI, like a text summarizer, which aren’t moving the needle. Therefore, your dependency on AI is minor and, thus, doesn’t necessarily affect your business model that much, whether costs increase or decrease.

In that case, paying attention to the costs of running these services and deciding whether to include them as part of your price is enough. As it’s not a crucial part of your product, you will probably absorb the costs for your customers, offering more value for your price, although erasing your gross margins if you’re a subscription business or erasing your customer lifetime value (LTV) metric if it was a one-time payment (probably not recommended unless this number is very high).

• If you’re a Level 2 wrapper, that means your value proposition would be seriously affected if the AI features are turned off.

This means that features powered by third-party AIs are exhaustively used by your customers, representing a considerable portion of your revenues and, importantly, of your COGS.

In this case, you need to start seriously considering moving into usage-based pricing, where the value of your product/service is proportional to the customer’s usage. Otherwise, you could see yourself in a similar situation to Microsoft in the past with GitHub Copilot, where they were losing up to $20 per customer. However, you don’t have almost $300 million daily free cash flow, do you.

Or, even worse, you could find yourself in a situation like Canva, which faced tremendous backlash when it was forced to increase its charges to customers multiple-fold (in some cases 300%) because its AI features’ COGS were eating it alive.

• But if you’re a Level 3 wrapper, your entire business depends on the outsourced AI (think of products like Bolt.new, Cursor, or Harvey).

Here, there’s no way around it: You need usage-based pricing. If you consider your product to be differential, you can try subscriptions with a decent markup like Perplexity or Cursor.

However, this is very complicated, and seeing Perplexity’s recent ventures into monetizing ads signals to me that this subscription service isn’t going well. Level 3 Wrappers are severely exposed to their AI provider, and some crucial measures must be taken to avoid getting killed. Importantly, as we’ll see in a minute, building a moat as a Level 3 wrapper is almost impossible, as you’re building your business on top of technology accessible to others.

But if managing costs is complex, things only get trickier when talking about the revenues of these companies.

Valuing My Products as a Wrapper

As The Information explains in one of its recent articles, many start-ups in this category have apparently booming businesses:

• Cursor, a tool to code with AI, reached $100 million in annual recurrent revenue

• Anysphere, another AI coding tool, has also reached the $100 million ARR mark

• Harvey just crossed $50 million ARR

• And Bolt.new reached $20 million in annualized revenue in just 8 weeks, basically unheard of.

But despite these astonishing numbers, most of them are struggling to raise capital from investors who argue that their business will be ‘short-lived.’ And those that have raised have done so at a considerable discount compared to model layer companies (between 20 and 30 forward PE vs the 40-60 range model layer companies are in).

But why are investors so wary?

For starters, several of these companies, some even in the billions of revenue, have displayed poor cost and price strategies. For instance, as mentioned earlier, Canva suddenly had to ramp its prices without previous notice.

This should not happen.

Unless you’re a Level 1 Wrapper, levels 2 and 3 should be going full usage-based pricing right now, aka margin businesses. In that regard, your business is basically adding a mark-up over the costs of running the AI APIs, making margin your business model.

While usage-based pricing is preferred by customers, who only pay for what they use, defining your margin is tricky. In most cases, your entire product is a set of workflows/business rules sitting on top of an OpenAI API, which means that:

• Your product isn’t really that unique or defensible, as in the age of AI, workflows will rarely be complex to replicate, exposing most SaaS companies to an inevitable death (as expressed by Microsoft’s CEO, not me).

• If your markup is very high because you’re in a high-ticket market, you are simply on a countdown toward extinction, as LLM providers will eventually penetrate your market with new features that obsolete your product.

• If your markup is low, your business is unattractive to LLM providers, which takes the risk of being cannibalized off the table. However, you’re still a shitty, low-margin business with huge COGS and, worse off, little to no room for reducing them because you can’t control API prices and little room to increase prices because your value just isn’t there.

For all those reasons, the application layer will showcase the most significant number of huge winners and an equal number of massive losers.

Funnily enough, most investors, including high-profile VCs, are pouring money into digital wrappers whose value is inversely proportional to the progress of AI; every time AI improves, their differentiation capabilities, their intrinsic value, falls.

While they appear to be firing in all cylinders, these comapnies are just surfing one huge tsunami that, in reality, is about to break into the White Cliffs of Dover, as invaders who thought they had it all experienced for centuries when trying to enter Britain.

So, the question is: how do we value these companies?

Identifying winners and losers

The typical way to measure the value of level 3 wrappers (which are most AI companies today) is as follows:

{Value = Utility x Degree of Abstraction x Difficulty of the Abstraction}

As previously mentioned, level 3 wrappers are businesses whose revenue is the gap between the final cost the customer pays and the cost of running the AI models. Therefore, your value is the workflows or features you’ve built on top of the foundation model, abstracting the difficulty of creating them on the user’s behalf.

Consequently, the resulting value of your product is how much abstraction you’ve done multiplied by the difficulty of replicating that abstraction (assuming utility, of course). Put another way, your company’s value is the quality of your features times the defensibility of your business model.

Now, in all honesty, this formula is pretty much identical outside of AI; a company’s values are always the quality of what it offers and times how defensible the business is. This is textbook investor maths, which is nothing we didn’t know already.

But what if I told you that the formula is incomplete and doesn’t account for the most important piece of the puzzle?

But what do we mean by that?

Become a paying subscriber of Full Premium package to get access to this post and other subscriber-only content.

Already a paying subscriber? Sign In.

A subscription gets you:

• • NO ADS
• • An additional insights email on Tuesdays
• • Gain access to TheWhiteBox's knowledge base to access four times more content than the free version on markets, cutting-edge research, company deep dives, AI engineering tips, & more