The Future of Frontier Labs’ Revenue — Saagar Pateder

Revenue for OpenAI and Anthropic has grown wildly. OpenAI grew from $6B to $20B in annualized run-rate revenue over the course of 2025 and reached $24B in March; Anthropic grew its ARR from $1B to $14B over the same time frame and was at $47B in May. These are “never before seen in the history of capitalism” growth rates which make near trillion-dollar pre-IPO valuations seem rational (perhaps even underpriced).<br>But can that growth continue? I’m not talking about the supply constraints right now on compute (SemiAnalysis covers this extensively), I’m thinking about the demand side of the equation. Will companies continue to pay for models for Anthropic and OpenAI when models from China are seemingly so close behind in intelligence and so much cheaper?<br>The Demand Curve<br>Here’s the framework I’ve thought of. First, imagine a bunch of tasks, and arrange them from easiest to hardest.

Now imagine you could quantify the revenue opportunity for AI in fulfilling all of the instances of that task. That’s probably a product of quantity (how often that task is done) and value (you’re willing to pay significantly more for a high-alpha financial insight or cancer drug than you are for a pot roast recipe). We can add that as our second dimension, and imagine all of the tasks in the economy as blocks of various heights that stack on top of each other to form a nice curve:

We could argue about which specific tasks are where, and how big they are, but this is primarily illustrative. Perhaps token consumption on the y-axis would be a better fit, but again… illustrative!

Now let’s imagine a line that sweeps through the chart, moving from left to right over time that represents what AI is capable of today. This is the frontier line; everything to the left of it is automatable with some AI model + harness combination [1]. Behind that is another line representing the cheap model frontier. Now our tasks are divided into three categories:<br>Those which can be done by very inexpensive models (open or otherwise) and frontier models alike.

Those which can be done only by frontier models, inexpensive models aren’t yet capable of doing these tasks.

Those which can’t be done by AI today, for any price.

I think the frontier model capabilities line is too far to the right in this example, but we’ll touch on that in a bit.

What defines “cheap”? Broadly, I’m using it as a proxy for the open-weight frontier, though cheap closed-weight models (e.g., GPT-5.6-Luna post-price-reduction) may also compete here. Open-weight models have competition at the inferencing layer, and so you’ll have dozens of companies competing to serve a model at lower prices with higher reliability at faster speeds (and you’ll have companies whose whole job is to simplify that choice) [2]. Open-weight models (without restrictive licenses) are inherently commoditized; a cheap closed-source model that achieves similar price-performance ratios is fundamentally in the same boat.<br>Given the dynamics around inference-layer competition and cost concerns from enterprises, tasks that can be done sufficiently well by commoditized models generate consumer (application) surplus rather than producer (inference provider or model lab) surplus. Tasks that are only able to be done by a handful of frontier AI models, however, should generate producer surplus rather than consumer surplus; it’s this strip in the middle that truly drives frontier labs’ revenue [3]. (I’ve included an appendix section at the bottom that shows the difference between token share and spend share on OpenRouter.)

As a side note, I really like how Jesse Zhang explains the dynamics behind the transition from closed frontier models to open-weight ones:<br>“...When a use case is new, you want the smartest general-purpose model you can get. You don't know the shape of the problem yet, so you pay a premium for intelligence you may not end up needing. That's the right trade at that stage. But once the use case is fully built out, when you know the distribution of inputs, the behaviors you need, and the failure modes to guard against, the trade flips. Now general intelligence is overhead, and you want the smallest, fastest model fine-tuned to do your specific thing extremely well.”<br>Jesse’s point is reflective of Decagon’s needs: Decagon cares a lot about latency and has the technical expertise to fine-tune models. Strip away Decagon’s idiosyncracies, and I wager that most companies don’t care about closed vs. open weights but instead just want the cheapest model to do [insert your favorite task] at a sufficient performance level.<br>Back to the main point: I’ve been drawing this chart to look like a bell curve. But how do we know what shape this curve looks like? I think the left half is accurate: the frontier model labs have experienced exponential growth in their revenue, and inference providers are also growing exponentially. This best fits an exponential ramp-up in...