The Risk of Vertical Integration

This is one of the most interesting questions facing the AI industry: where is more value created, at the underlying technology layer or at the interface? Which will subsume the other? And what will that interface even look like?

Today, the leading frontier model providers have two basic business models: a consumer chat app, and an API offering (which allows enterprise customers to build products on top of the LLM foundation.) This creates an interesting competitive tension between suppliers of the LLM technology and their customers.

Take the coding industry. In our previous post, we mentioned Cursor which provides an AI-powered interface for programmers. It became one of the most-loved products in the software industry, as evidenced by its remarkable revenue ramp to a $3B run rate three years after launch (and its $60B acquisition by SpaceX). Since Cursor offers its users a choice of LLMs, its users can choose to work with ChatGPT or Claude within the Cursor app. As such, Cursor has also become one of the largest customers of companies like Anthropic.

Yet coding is such a large and strategic domain that each of Anthropic, OpenAI, and Google Gemini have also launched specific products for programmers, called Claude Code, Codex, and Gemini Code Assist which target that same market. Even if these products use different interfaces and entry points to the market, they are still competing with their customer - Cursor. More recently, Cursor also introduced its own coding model, Composer (based on one of the Chinese open-weight models), competing back with the LLM providers at the technology level, and giving users the option of a lower-cost model instead of Claude, ChatGPT, or Gemini.

Furthermore, Cursor’s interface subtly promotes Composer over other models, switching to it when budgets have been exceeded and presumably favoring it when the user has not made an explicit choice (i.e., in auto mode).  Based on benchmarks, Composer is comparable, or even better, than Anthropic and OpenAI’s various frontier models for certain kinds of tasks, though certainly not all.  Cursor has been able to leverage the wealth of data gathered from users’ interactions with the app to inform the way it trains and fine-tunes Composer, optimizing it for the kinds of tasks its millions of users demand every day. 

What if, as a next step, Cursor (now owned by SpaceX) forced users onto its own model, vertically integrating its entire technology stack and depriving Anthropic and OpenAI of this revenue source?

The risk for Cursor is that such a move could backfire. Claude enthusiasts might abandon Cursor for Anthropic’s Claude Code even if it meant learning a new interface. The opportunity cost of not having access to Claude’s ever-improving capabilities is just too high.

Alternatively, Anthropic could launch a competing interface. Certainly, it has the resources to build an app like Cursor’s if it chose to. One risk of doing this – directly taking on a customer like Cursor – is that Anthropic could erode the confidence of its other API customers, pushing these app developers to build on top of alternative models rather than partnering with a potential competitor.

Anthropic may also have a longer-term view on all of this. One could argue that programmers may not even need a Cursor-like interface in the future, because agents will handle most coding tasks without requiring direct supervision or editing –– abstracting the interface layer entirely. In such a scenario, Cursor, the app, as it exists today, may have little or no value. All the value would accrue to the frontier model providers.

Of course, Cursor has not been standing still. For example, a Cursor Developer-agent can now autonomously build and test new software features. A Cursor QA/QC-agent (i.e., a Quality Assurance /Quality Control) can then record on video how the feature behaves when used and send that video back to the Developer-agent, which will not only be able to inspect the code but also watch a screencast of the user-experience (i.e., Cursor’s agents validating a software they just built!).

At a more fundamental level, Cursor is also using a host of algorithms and smaller models to optimize usage. If a user asks it to design a new feature, it does not simply send the entire codebase to the LLM’s context window and ask the LLM to do the work. Cursor’s interface software does crucial preliminary work to ensure that only the most relevant pieces of code are transmitted, improving both cost efficiency and performance, and adding significant value on top of the frontier models’ core capabilities.

This kind of LLM overlay – often categorized as a “harness” – deepens the integration of the interface layer within specific sectors and allows companies to build their own moats in an increasingly competitive environment.

The example of Cursor also highlights an area where open-weight models can gain significant market share. The availability of alternatives for enterprise AI adopters means that leading AI players will continue to face strong competitive pressure to continuously improve both model capabilities and pricing. In practice, it is unlikely that the market for AI-generated code ends up with a single dominant winner anytime soon. Chances are that both the interface layer and the underlying LLM’s will continuously evolve, with each player capturing a portion of an enormous and rapidly growing market.

A little over a year ago, many laughed when Anthropic’s CEO, Dario Amodei, claimed that AI would write all software within a year. Critics mockingly asked why his company was still hiring software engineers. Today, engineers at companies like Anthropic, OpenAI, and Spotify openly state that AI is, in fact, writing 100% of their code, and that engineers’ roles are increasingly shifting toward supervision, orchestration, and direction. GitHub, a code-hosting platform, reports that 4% of public code is generated by AI today – a figure GitHub expects will rise to 20% by the end of 2026. This shift will profoundly disrupt traditional software companies, leading to what has been termed a SaaSpocalypse.

In recent months, every time Anthropic or OpenAI announced a new industry-specific feature, the stock prices of companies in related sectors plunged as investors tried to grapple with its impact. For example, in February 2026, when Anthropic announced new security capabilities, the stock prices of software security companies plunged – CrowdStrike fell 18%! – as investors questioned whether Claude would ultimately replace the functionality these companies had to offer.

The market reaction implicitly raised the question: Why pay for CrowdStrike when Claude can do the same for you?

At the same time, perhaps companies like CrowdStrike will integrate AI tools into their offering and strengthen the set of industry-related features and integrations they have already built, and which cannot be easily replaced by Claude’s technical abilities alone. The question facing incumbents, as reflected in the volatility of their stock prices, is not whether Claude matches their feature sets one-for-one – it is whether these new capabilities will change the operating model of the industry, ultimately rendering some of them obsolete.

As frontier model capabilities evolve and improve, this question of industry disruption becomes even more pertinent.

In this light, the security industry may be the canary in the coal mine of how step changes in model capability can affect an industry.  In April, Anthropic announced its next-generation security capability, called Mythos – a model so powerful that Anthropic initially restricted its use to a small set of known companies that were allowed to use it to scan their own codebases. Anthropic was worried that, just as companies could use Mythos to find and patch vulnerabilities within their codebase, other bad actors could also use Mythos to exploit those same vulnerabilities. As a result, Anthropic gave various industry leaders a chance to use Mythos, patch their vulnerabilities and to share what they learnt with the good players in the industry as a whole, all before the bad actors got their hands on the same frontier model capabilities.

The release of Mythos raises some very interesting questions.

(1) First, it is instructive to think about the value of such new models in a thought experiment: If Anthropic had released the model to the highest bidder, how much would the leading banks pay for early access; or how much would they pay to stop bad actors from gaining access. Of course, there are a host of reasons why Anthropic would not make such a crude move, but the question re-emphasizes the power such models have to change industries. It also opens up the question of whether the current per-seat pricing model will persist going forward, and how the value created by the models will be captured by various players in each industry. (The US government’s banning of Fable, the public guard-railed version of Mythos, also raises the question of whether governments may play a similar gate-keeping role for frontier models in the future.)

(2) Second, Mythos serves as an example of how even the most capable models can be deployed. As Cloudflare, one of a select group of Mythos users, has explained: it is not optimal to just ask Mythos to “please find vulnerabilities in my system.” Cloudflare’s security team still needed to orchestrate Mythos’s capabilities by splitting the problems into a series of narrower tasks that are constantly tested and questioned by Mythos itself. Could Mythos build such harnesses to conduct this orchestration itself? Possibly, but (as discussed in our previous posts) each model generation is also trained on a set of questions that define the level of abstraction at which the model can excel. The need for orchestration will likely persist, be it at the frontier model level, at a vertical supplier level or by the end customer itself. Either way, this highlights a broader point about how current models operate: each model generation will bring new capabilities that will push us humans to generalize our knowledge to higher levels of abstraction and orchestration. This trend will likely persist across industries, transforming the nature of work done in each.

(3) Third, Mythos unlocks a host of new capabilities for enterprises to assess, manage, and foolproof their security infrastructure. The work that Cloudflare and others are doing using Mythos is not the work that any other software company could do or was able to do previously. In that way, Mythos is not competing directly with any of the incumbents. In fact, a host of companies are using the current generations of publicly released Claude models to scan their codebases and even the largest and most popular software products are proving to have vulnerabilities that had not been known. For example, even before Mythos, Claude discovered 22 new vulnerabilities in the popular Firefox browser, 14 of which were classified as high severity. As such, the models’ technical capabilities expand the market for software security as they make new categories of work economically viable.

(4) Fourth, of course the fact that the market is expanded does not mean that the incumbents are not in danger. That is reflected in the volatility of their stock prices, which follow a typical pattern when new technologies get introduced. First there may be panic that an incumbent would get replaced completely – the stock goes down. Then it becomes clear that things are not so simple because the current business model makes the elimination of that incumbent quite difficult – the stock price goes back up. Then, over time, a host of startups offer better products based on the new technical capability, and slowly certain incumbents’ value propositions become irrelevant, as do their stock prices.

To illustrate this, consider a hypothetical example.  Let’s imagine that a new security start-up offers a free(!) service by building a thin layer of security functionality on top of its clients’ Claude subscriptions. It targets companies that are outside the reach of the larger enterprise players – say, companies with < 10 FTEs. The core security functionality provided by Claude may be so effective for this target segment that the start-up could guarantee its customers’ security and run its service with very few resources, capturing a market that did not exist before.

At this point, it is not directly competing with any incumbents. It could then introduce a new business model, offering a full guarantee of software security to small companies for a flat fee – a sort of insurance policy. Such a business model would be structurally different from traditional per-seat pricing. (Per seat pricing is the prevalent SaaS revenue model where customers pay based on the number of users of a software, rather than the performance of the software, which an insurance policy more closely resembles.)

Over time the start-up can offer this security guarantee to larger clients and expand its addressable market. It starts to encroach on the incumbents’ territory, but its business model is hard for incumbents to adopt because it would require them to redesign their software and cannibalize their customer bases. Thus, the incumbents retrench to provide highly customized and expensive solutions for only their largest clients, making them increasingly irrelevant over time. Such a scenario would be a classic case of disruption, as originally defined by Harvard’s Clayton Christensen back in the 1990s.

AI has created a whole new set of risks and opportunities across the software industry that will play out in the years to come. When AI has so many capabilities baked in, why would you need specialized software at all? Or rather, which software companies would you still need? Which have built durable advantages, and can defend their moats over the long run? Which will become irrelevant following the patterns described above? SaaS stocks have been declining since the second half of 2025 as the markets grapple with these questions, discerning the long-term winners and losers of the monumental AI disruption.

Other industries are poised for similar turmoil and disruption, as more of their work gets delegated to AI agents.

For example, the legal industry is poised for AI-driven transformation.  Several startups are now providing new AI-based legal tools that do not necessarily replace lawyers but instead give them new superpowers to perform tasks they were not able to perform before. LLMs can read through all of a company’s contracts (much like they read through existing software codebases) and uncover vulnerabilities (or opportunities) in those contracts.  They can also be trained to understand legal concepts and provide structured guidance on complex legal issues to lawyers and non-lawyers alike. LLMs excel at being good-enough experts in multiple domains. They can hold large amounts of contracts, legal concepts, and case law within their context window during business meetings, ready to answer all the questions an executive may have. Previously this would have required an army of lawyers and been prohibitively expensive.

AI is playing a similar role in finance. It is simply not economical for human analysts to read thousands of financial reports and aggregate both quantitative metrics (like comparable analyses) and qualitative insights (like management outlooks). However, armed with some basic financial training, an LLM can perform such tasks at negligible cost. To date, investment banks have relied on MBA-educated analysts to perform such work, and have thus been constrained by scale (i.e., the limited number of documents each employee can review). AI changes this equation. That said, AI may not radically upend investment banking, given the relationship-driven nature of the business – but there is no doubt that AI will allow those same bankers to do much more with less.

Both legal and finance are examples of sectors that are considered “rules-based.” Each has a specific set of knowledge-based rules that an LLM can be trained-on relatively easily, and each has access to enormous amounts of data that have been underutilized to date – be they the thousands of financial statements sitting on the SEC website or all of the case law accumulated throughout history. AI suddenly makes all this data analyzable at a very low cost.

The generalized models offered by leading LLMs are in a good position to service these rules-based industries. For example, Anthropic has introduced Claude for Financial Services, a set of tools and capabilities packaged specifically for the finance industry. Another example is Harvey, a legal-tech company that has a partnership with OpenAI to deliver AI-based services to the legal industry. Like Cursor but for legal, Harvey’s product has deep integrations into the workflows of the industry.

Just as with code, the LLM leaders have carved early leads in these markets. However, the competitive questions remain and will become more pertinent over time. Just as in the case of Cursor, Harvey would have the choice to build its own legal-specific model by fine-tuning an open-source model and thus reduce its reliance on OpenAI. To maintain their leadership, frontier model providers will need to continue innovating capabilities, find solutions that allow their customers to fine-tune their models using proprietary data, and make integration within the enterprise environment easy and price-competitive.

For rules-based sectors like legal and finance, fine-tuning or customizing existing open-source models may be worthwhile, but there would be little justification for a company to develop a specialized frontier model for, say, finance only. As discussed in our first post, the costs of developing frontier models are just too high, and there is not enough model-specific data to differentiate the capabilities of a specialized model. However, there are functions and industries where specialized models make a lot of sense. Industries with very specialized skillsets where models can access existing data or create high-fidelity synthetic data in sufficient quantity to train the LLM are clear opportunities. Examples of such industries include self-driving cars and robots. Tesla’s unique dataset in both miles-driven and robotic manufacturing puts it in a great position to develop leading LLMs for autonomous cars as well as humanoid robots focused on manual tasks.

LLM leaders like OpenAI and Anthropic have decided not to compete in robotics and self-driving at this point, but there are other domains where the user base and datasets are much closer to home and where competition from specialized LLM providers may play out.

Voice or more specifically Speech-to-Text and Text-to-Speech is one of those areas where leading LLM providers will have to up their game if they want to hold onto it. For example, ElevenLabs, has managed to build a compelling business by labeling voice data for accents, textures, and personalities to train their models to offer clients truly customized voices, e.g., an energetic, clear, millennial with an American accent or alternatively a smooth, friendly, persuasive Englishman…and, by the way, both customized voices can communicate in 70+ languages. It turns out many enterprise customers are willing to spend more to provide these nuanced/branded voices.

Just as we have seen that milliseconds of response time and minor nuances on website design have outsized implications on customer satisfaction, it may be that the nuances of voice differentiation and customization will provide the next level of touch-and-feel needed for best-in-class customer service. Such capabilities can also open up hither-to unimaginable use cases: Elevenlabs is already automating audio books – ultimately one could imagine that any book could be read with any voice, customized to the listener’s preferences, and that the dialogues may be impersonated in different voices; Elevenlabs is also working with music labels, bringing a multitude of voices on-demand to music creation; Elevenlabs can also allow creators to tweak any audio publication, be it a podcast or a marketing video, to touch up the voice and get it just right; it can play a role in film distribution - imagine movies and videos being simultaneously translated in any language using the actor’s own voice.

The technology opens up so many possibilities, and so ElevenLabs’ models have an opportunity to create and dominate brand-new markets. This example also points to the possibility that other multi-billion niches emerge as entrepreneurs use LLM transformer technology to develop specialized, proprietary models able to outcompete the leading LLM providers within a specific domain.

In each of these cases, be they rules-based (like legal, coding, and finance) or skills-based (like voice and robotics), the capabilities of AI models are already transforming industries. In each case, if the industry is large enough, the leading LLM providers are training their models with specialized skills for those specific domains while vertically integrated competitors like Cursor and ElevenLabs are fine-tuning existing models or building specialized models from scratch to tackle the same vertical. The prize for being early in any given domain is huge because early movers benefit from a data flywheel as they can incorporate usage data to continuously improve their models and services and thereby secure a significant and sustainable share of a highly lucrative market. As such the competitive intensity between leading labs and domain-specific players will remain high.

The vertically integrated players are pulling on one side, and the generalized models on the other – a tug of war of epic proportions that is bound to redefine the structures of all sectors of the world economy. Far from commoditizing LLMs, these dynamics point to a future of highly differentiated solutions cutting across industries at multiple angles, redefining the structures of each segment of the economy in new and interesting ways.

In the fourth and final post in this series, we will examine potential disruptive threats to the current leaders from radically new technological approaches.