Every computing epoch has brought a slew of articles and posts decrying that the leading businesses were untenable due to low margins. This was an oft-used critique of web companies (Amazon and Netflix of particular note), mobile (Uber and Doordash popular targets), generative AI models, and now more recently, apps built on top of foundation models.
While these pieces are broadly penned, a common and particularly puzzling source of the negativity is investors. We’ll never understand why an investor would be comfortable disparaging startups, or offering general pessimism about innovation waves. However, the pieces do tend to get a lot of visibility, and as a result, are a relatively cheap win for clicks. The problem is that they’re quite often wrong, or at best short-sighted, and using them as investment guidance would have resulted in passing on many of the largest winners in tech history.
In this post, we address some of the ill-informed chatter about gross margins in AI app companies. We strongly believe (and history provides ample evidence) that lower gross margins at a moment in time are not a long term indicator of a lack of a sustainable business model. Further, many of the critics completely miss the nuance of today’s AI companies by drawing comparisons to DTC (direct to consumer) subscription companies in prior cycles that suffered from low retention and paid customer acquisition. Those cohorts look dramatically different from today’s AI applications, which deliver deep customer value, strong usage and retention, and explosive net expansion via spreading to teams and the enterprise. We’ll do what we can to correct that as well.
In what follows are some of the common criticisms of app companies built on foundation models.
Claim: AI apps must be permanently low margin at $X/month
A primary fixation of critics is to point out that introductory plans that offer unlimited usage tend to be low gross margin. And yes, that’s generally the case with most product-led companies. However, the standard approach that follows is for the company to introduce a pro plan that will offer limits which then convert to usage based billing. Or they’ll abstract away exactly which model is being used allowing them to direct a subset of traffic to smaller / cheaper models. Many of the leading gen AI apps companies continue to see strong continued growth after this tiering, underscoring that actual product value rather than subsidies drive usage.
Furthermore, it’s important to realize that, in many domains, the cost to serve a query is highly variable. If a subset of the more expensive queries can be identified and serviced by a cheaper model, the margin problem is mollified. Many of the leading app companies we’re aware of are quite sophisticated at routing high value queries to margin optimized inference.
The argument that users simply want the frontier model at all times suggests a blinkered understanding of how these apps are used in practice. Model preference varies by task, and in many use cases where performance differences are limited, a lower TCO (total cost of ownership) is preferred.
Claim: AI app companies’ unprofitable cohorts drive overall business economics indefinitely
For many AI subscription businesses, a small minority of users are driving the bulk of the usage and therefore underlying costs. This means that rate limiting a small percentage of total users (who tend to be loud on X) can meaningfully reduce costs without impacting revenue growth significantly. Even frontier models face this issue, typically seeing limited revenue impact but meaningful cost reduction after introducing rate limits for the top 5% of users.
Moreover, most app companies have profitable segments at higher level tiers (teams, enterprise, etc.) that are not visible in external data. In the AI era, pricing and packaging are more strategic margin levers than ever, and professional users continue to surprise even our more optimistic internal models with their willingness to pay for high value, heavily utilized AI products.
Claim: AI app companies don’t control model pricing and are at the mercy of monopolistic rent seeking
To date, there is no model monopoly. And there is unlikely to be in the foreseeable future, given the level of investment behind the big players and the effectiveness of the open source community. Model drops see massive movements of users, in part because apps are a valid distribution channel and can effectively steer users. Very recently, we’ve seen a massive swing of developers away from Claude to GPT-5 after OpenAI’s recent launch. And while drafting this piece, DeepSeek released its V3.1 API, which appears to again dramatically push the performance cost boundary, likely causing yet more shifts.
Further, there is little indication that the foundation model market will naturally consolidate into a monopoly. Being first to SOTA (state of the art) in a model modality has not guaranteed long-term leadership. There was a nearly year-long period where model providers couldn’t catch up to GPT-4. Midjourney wasn’t the first to image. Google and Kling weren’t first to video. In fact, one could argue leadership positions are historically short for AI models. And in the absence of a dominant position for a sustained period, pricing pressure increases in what has been a savage fight for market share over the past 18 months.
But that’s not the only driver of cost reductions. Inference optimizations are still in early innings, and the trend to date on inference costs is indisputable. Depending on what benchmark you’re looking at, for the same cost per million tokens, inference costs have dropped anywhere from 10x to 100x+ in the last 18 months. While we acknowledge that it’s too simplistic to tie dropping overall cost of inference to increased margins in apps, and we recognize the frontier models are typically more expensive than prior generations, it’s similarly short sighted to assume we’ve reached equilibrium given the ongoing competition and optimization at the model layer. Moreover, given the utilization of multiple models as previously discussed, drops in the cost of prior model generations still drive opportunity for margin expansion.
Claim: Subsidies are clouding true product market fit
At the moment, there is a lot of tired talk about “selling $1 for 50 cents” to describe AI app companies (particularly those touching code) and using VC subsidies to buy behavior from transient users that will immediately dry up when the music stops.
Drawing this conclusion based on short term gross margins is dangerous, as it ignores the more important metrics to gauge whether a free tier or prosumer strategy is working: 1) conversion to paid, 2) conversion to enterprise 3) cohort usage as well as revenue retention and expansion. If the top of funnel that a company is attracting converts to paid, walks the product into enterprises (often with record fast sales cycles), and not only retains but also expands revenue, the strategy is working, full stop.
The playbook is not a new one by any means, but we’re starting to see gen AI companies execute on this successfully, penetrating the enterprise faster than anticipated, and even passing incumbents with incredible distribution. AI companies are bootstrapping their own distribution by building a large user base that fuels their sales funnel, leading to higher-margin teams and enterprise contracts that will account for an increasingly larger share of revenue over time. As anyone who has been in enterprise long enough knows, the top 20% of contracts generally drive 80% of revenue. Having a small segment of revenue where a company trades off margin for pipeline into higher margin segments is a feature, not a bug.
Claim: AI apps are thin and add no value beyond access to the underlying frontier model
Another prevailing misconception is that gen AI apps are predominantly thin wrappers around frontier models. While this may be true of a subset, we’ve also observed that applications have a strategic control point that can quickly lead to differentiation. This happens primarily in two ways: 1) high-velocity teams use the initial entry point to rapidly expand their product surface area, and 2) leading applications compose multiple models (including some of their own) offering a service at either a capability or price point no single model could offer. Both contribute to higher long term margins, yet require winning the user in the near term.
The reality is that models themselves simply aren’t products or platforms. AI code gen companies have created massive value by adding capabilities such as hosting, custom domains, and security scans. High-value (and often high-margin) upsells increase ARPUs and retention.
The rise of product specific models creates room for even further potential differentiation, as products often have far more contextual data on user behavior and preference. As a result, they are in a position to create models, or fine tunes of models, that are uniquely effective in a way no model provider without deep app layer visibility could achieve. Take coding for example,from the vantage point of an API, it’s difficult to know to what extent generated code was accepted. And even if it was accepted, there is no visibility into whether it survived long enough to be submitted as a PR or merged into main. In the constant race to find ever more meaningful sources of data, few are closer to the truth than product data. And as much as the apps teams are training or fine tuning their own models, this is a clear path for model quality differentiation, another potential lever to drive longer term pricing power and cost reduction.
Conclusion
Investors looking for visibility at the expense of startups will continue to write these formulaic and generally ill informed critiques on short term startup margins. There is little we can do about that. But for those looking for a broader view on a complicated topic, we hope this piece adds positively to the discourse. Not all startups will succeed. Low margins will, at times, be a fatal aspect of the business. And of course, systemically low or negative margin businesses should be valued accordingly. That said, margins alone don’t tell the full story. We strongly believe that fixating on margins while ignoring a company’s value to its customers, retention and ease of acquisition misses the mark. Just as important, there are many effective strategies for AI app companies to drive healthy margins in the long term.
Our friend Nat Friedman said it best: “Being a pessimist makes you sound smart. Being an optimist makes you money”.
