AI Wrappers and the Coming Compression — A PM Perspective
What remains and what gives, as a market practitioner sifts through the AI complex
TL;DR
• The market prices all “AI” off one multiple. The application layer is about to split into platforms and features, and it will not be gentle.
• Wrappers lose pricing power the moment the model provider ships their feature natively. Jasper is the template: pricing erodes first, margins follow, the multiple re-rates last.
• The real mispricing is indiscrimination. The consumer search box trades near 100x revenue; embedded, data-rich businesses like Harvey trade at half that. The richest multiples sit on the thinnest moats.
• The screen: would the model provider build this itself, and why hasn’t it? With the equity risk premium near zero, you are paid to be right where opinion is still dispersed, in the application layer, not where durability is already consensus.
Two artificial-intelligence companies, each generating roughly $200 million in annual recurring revenue. One is valued at $11 billion. The other at $20 billion. The cheaper of the two sells AI agents that draft contracts and run diligence inside a majority of the AmLaw 100, backed by a content alliance with LexisNexis and switching costs measured in organizational years. The more expensive one is a consumer search box.
This is not a quirk of two data points. It is a preview of the largest mispricing in the AI complex, and it has almost nothing to do with whether artificial intelligence creates value. It will. The question is who keeps it.
Approached from the desk rather than the sidelines, the question changes. The strategist asks who wins; the practitioner asks what is already in the price, how the view can be expressed, and what it costs to be early. Sift the complex with those three questions and it sorts into two piles: what remains, and what gives.
The label flattens everything. “AI” now describes chips, data centers, hyperscale cloud, enterprise software, coding tools, search, and the consumer chat box, and the market has been content to price the entire stack off a single multiple. The economics underneath the label could hardly be more different, and the gap is widest at the application layer, where a large share of the most celebrated companies are, at bottom, wrappers: a clean interface and some workflow logic placed on top of a model that someone else trained and someone else pays to run.
There is nothing dishonorable about a wrapper. Many are excellent, and several have introduced millions of people to a better way to search, write, research, or code. But the relevant question is not whether a wrapper solves a problem today. It is whether it can still charge for solving that problem once the model provider decides to solve it too.
That is the distinction the market keeps collapsing. A company can be early and still not last. It can have a good product, grow quickly, and end up as a feature inside someone else’s platform anyway. The market will eventually sort AI companies into two camps, and it will not be gentle about it. Platforms control distribution, compute, data, identity, or the workflow itself. Features improve the experience until the platform absorbs them.
History already ran this experiment, recently enough that the participants are still in business. In October 2022, Jasper, an interface that turned OpenAI’s models into marketing copy, raised $125 million at a $1.5 billion valuation and was being described as one of the fastest-growing software companies on record. One month later, OpenAI released ChatGPT. Customers quickly found that they could get most of what Jasper sold for $20 a month, or for nothing. Revenue that had reached roughly $120 million in 2023 fell to somewhere near $55 million the following year. The company cut its internal valuation and pivoted, surviving as an enterprise marketing tool rather than the category-defining platform it had nearly become.
The sequence matters more than the casualty, because it is the sequence every exposed wrapper will follow. Pricing power goes first: the moment the provider ships the same capability natively, the wrapper can no longer charge a premium for convenience, and discounting begins. Margin follows, because the wrapper still pays the provider for inference on every query while its own price falls toward the provider’s. Only last, and most violently, does the multiple re-rate, when the market stops paying for growth it had assumed was durable and reprices the business as the feature it turned out to be. The tell that the cycle has begun is rarely a revenue miss. It is a change in language, the quiet moment when a company that used to call itself a platform starts calling itself a copilot.
The serious objection to all of this runs the other way, and it deserves to be stated at full strength. As models commoditize, with open weights closing the capability gap and the cost of a given level of capability falling by an order of magnitude a year, intelligence itself becomes the cheap input, and scarcity migrates to whoever owns the customer. On that logic the wrapper is not the victim but the winner. It holds the user relationship and the data while the model collapses into an interchangeable commodity bought by the token. The bridge does not become unnecessary. The bridge owns the traffic.
The argument is right about everything except who is standing on which side. In previous platform shifts, the aggregator and the supplier were different companies. The operating system belonged to Microsoft; the applications belonged to everyone else. This time the model providers own the distribution as well. ChatGPT is not a wholesale model dressed up for developers. It is the most-used consumer interface of the cycle, holding precisely the user relationship, memory, and default position that the aggregation thesis says are scarce. When the supplier already controls the channel, the wrapper cannot retreat up the value chain to safety, because the provider is already there. That is the mechanism that turns a product into a feature, and it is why this cycle compresses the application layer rather than rewarding it.
But “the application layer compresses” is too blunt to be useful. The deeper error is indiscrimination. The market is pricing the application layer as a single asset class at the moment that class is about to split in two.
Consider Cursor, the strongest case for the durable wrapper. It began as a clean shell around GPT-4 and Claude, the textbook thin wrapper, and now runs at roughly $2 billion in annual recurring revenue with more than half the Fortune 500 inside it. Its last closed round valued it near $30 billion, and it is reported to be valued at around $50 billion. What separates it from Jasper is not the original idea, but what has accreted beneath it: enterprise contracts, the developer’s entire workflow, and a model of its own, now trained for the task. Cursor is racing to escape the wrapper because it can see the providers’ own coding agents, Anthropic’s Claude Code and OpenAI’s Codex, coming straight at it. That durability is not a reward for arriving early. The company is digging the moat now, spending billions to do it before the provider arrives. Still, the veteran venture investor Bill Gurley, echoing Marathon Asset Management’s famous critique of the dot-com bubble, made the point this spring: being right about the product is no protection against being wrong about the price. Cursor’s product-market fit is not in doubt. Its multiple is the open question.
Compare Harvey, the legal company on the cheaper end of the pair we began with. At $11 billion on revenue approaching $200 million, it carries roughly half the multiple of Perplexity, the consumer search box, despite owning far more that a model provider cannot trivially replicate: a content alliance with LexisNexis, tens of thousands of customer-built agents, the compliance and ethical-wall machinery that regulated firms require, and a seat-expansion flywheel inside institutions where changing vendors is a governance event. The market, in short, is paying its richest application-layer multiples for the businesses with the least to defend, and a discount for the ones embedded in how regulated work actually gets done. That is the mispricing.
The screen that separates the two is not the one usually offered. “Does the company own something the model providers cannot replicate” is too soft, because almost anything sounds unique in a pitch deck. The sharper test inverts it: would the model provider build this itself, and if the answer is yes, why hasn’t it already? Sometimes the honest answer is a real obstacle, regulatory liability the provider would rather not carry, or proprietary data it cannot reach, or a distribution channel it does not own. Then the moat is real. When the honest answer is that it simply hasn’t gotten around to it, there is no moat, only a head start, and head starts compress.
This is also where the reflexive conclusion, sell the wrappers and own the picks and shovels, quietly fails. The infrastructure layer is durable, and almost no one disputes it, which is exactly the problem. Nvidia and the hyperscalers are the most-owned assets on the planet because the entire market agrees they will endure, and that agreement is already in the price. Being right about durability earns nothing when durability is consensus. With the equity risk premium compressed toward zero, the market is no longer paying anyone to take the obvious risk. It is paying, to the extent it pays at all, to be correctly positioned where opinion is still dispersed, and in AI that dispersion lives in the application layer, in the gap between the wrappers the market is treating as features and the embedded businesses it is pricing as though they were the same thing.
Which returns the argument to the book, because a thesis that cannot be expressed is only an opinion. The cleanest names in this story are private. One can neither short the consumer search box at 100 times revenue nor own the embedded legal platform in the size the view deserves. So, the position has to be built where it is investable: across the listed application-software complex, the public infrastructure layer, and the few platforms that own both a model and its distribution. That makes it a relative-value posture far more than a directional bet. Timing is the harder discipline, and here the derivatives reflex is the correct one. Compression is a path, not an event. It runs through pricing, then margin, then the multiple, and can take quarters to traverse while the consensus rating holds. The risk is not being wrong but being right and early, paying carry while a coordinated market stays coordinated. Jasper unwound in months once it began, but nothing told you in advance which month. The expression wants to be convex rather than levered, and patient rather than anticipatory. There is no premium for standing unhedged in front of the most-owned names on the tape merely for having noticed they are expensive.
Sorted that way, the complex separates into what gives and what remains. What gives is the part of the application layer whose only moat is convenience: the thin consumer wrappers, and anything for which the honest answer to whether the provider would build it is yes, it simply hasn’t yet. That is where the multiple is least defensible and the eventual re-rating most violent. What remains is twofold. Scarce infrastructure endures, but its endurance is consensus and already paid, so it earns ballast rather than edge. The edge lies in the embedded businesses: those with proprietary data, regulated distribution, and workflow stitched so deeply into the work that removing it is organizational surgery. These are the companies a model provider would have to want to become rather than merely decide to build.
Valuation, in the end, is less a measurement than a coordination mechanism. Whichever rubric enough capital agrees to becomes, for a while, the price. The market has coordinated around a single “AI multiple” that pays generously for revenue growth and asks too little about its source. That coordination will not break gently. It will break the way Jasper’s did, first in pricing, then in margin, and only at the end in the multiple, and it will break unevenly, sparing the businesses with something genuinely their own and discovering one by one which of the rest were ever more than a convenient layer over someone else’s model.
AI will create enormous value. It will not distribute that value evenly, and it will be least generous to the companies whose only advantage was being the first comfortable way to reach a model the provider can now reach directly.
Everything else is a wrapper.
And in technology, wrappers have a habit of becoming features.
#AI #Markets #Investing #Macro #Valuation
© Macro Fireside
The Macro Fireside
macrofireside.com · @Macrofireside on X
The Macro Fireside is a practitioner’s publication, written at the intersection of markets, policy, and geopolitics by someone who has spent decades managing money across multiple market cycles. Analysis here is earned, not assembled.
Disclosure: The author may hold positions, personally or through managed vehicles, in the public securities and themes discussed. This is commentary and analysis, not investment advice or a recommendation to transact in any security.
For professional inquiries: gs@macrofireside.com