The paradox of OpenRouter
In May 2026, a company called OpenRouter raised $113 million in funding; it was valued at $1.3 billion, making it a unicorn.
The lead investor was CapitalG, a subsidiary of Google's parent company. NVentures, NVIDIA's venture capital arm, also participated, and existing shareholders such as a16z and Menlo Ventures increased their investments.
What is this company doing?
It can be explained in one sentence. It packages the APIs of over 400 large models worldwide into a unified interface. Developers can access all models with a single connection. OpenRouter charges a handling fee of about 5% for each call.
To put it simply, it's a middleman. It seems that people think this is an easy way to make money. However, after some research, I found the founder even more interesting.
Alex Atallah graduated from the Department of Computer Science at Stanford University and wrote code at Palantir. In early 2018, he founded OpenSea with Devin Finzer. It was the world's first and later the largest NFT trading platform.
In January 2022, OpenSea's monthly trading volume exceeded $4 billion. It was valued at $13.3 billion, with a market share of 97%. What does that mean? Out of every 100 NFT transactions, 97 were made through OpenSea. It was almost a monopoly.
In July of the same year, Atallah resigned from OpenSea. He remained on the board, saying he wanted to start something new from scratch.
In February 2023, Meta released LLaMA, and Stanford followed with Alpaca. After observing these two events, Atallah identified a trend: small teams could also develop competitive AI models. He predicted that a large number of models would emerge in the future.
Driven by this prediction, he founded OpenRouter with Louis Vichy, a collaborator on the Plasmo framework.
If you visit Atallah's personal website, you'll see two lines of introduction side by side:
「OpenRouter: the first and largest router for LLMs; OpenSea: the first and largest NFT marketplace.」
He emphasizes this symmetry himself: the first and largest NFT trading platform, and the first and largest AI model router, both created by the same person.
Atallah has a clear line of reasoning: When the supply side becomes highly fragmented, create a central hub for the supply.
In 2018, he noticed the fragmentation in the NFT market. Various digital artworks, game items, and virtual real - estate emerged, and people didn't know where to buy or sell them. So he created a trading platform.
In 2023, he observed the fragmentation in the AI model market. OpenAI, Anthropic, Meta, Google, and Mistral were all launching different models. The interfaces were incompatible, pricing was inconsistent, and stability varied. He created a router.
The structure of his two bets is exactly the same. As long as the premise holds, the central hub is the most stable position. Isn't it interesting?
01
But think about it. Just setting up a simple website and connecting a few models. Why is it worth $1.3 billion?
I checked the data from Sacra.
OpenRouter's annualized revenue was about $1 million by the end of 2024. It rose to $5 million in May 2025, $10 million in October 2025, and $50 million by early 2026.
It increased 50 - fold in a year and a half.
Sacra's explanation is simple. It's all about usage volume. The more people use it and the more calls are made, the higher the revenue.
I calculated that with an annualized revenue of $50 million and a valuation of $1.3 billion, the price - to - sales ratio (PS) is about 26 times. This figure is relatively high for SaaS companies and not cheap for infrastructure companies. What exactly are investors buying?
Let's start with a number: 25 trillion.
This is the number of tokens processed by OpenRouter in a week. It was 5 trillion six months ago, a five - fold increase. On average, about 100 trillion tokens flow through the platform each month, and there are over 8 million users.
Where does this usage volume come from?
A large part comes from AI programming tools such as Cursor and Claude Code. Developers call models multiple times a day when writing code. A developer may make dozens or even hundreds of calls in a day, and many of these requests go through OpenRouter.
Programming has a unique characteristic. It naturally requires multiple models. Different models are used for front - end development, code review, and long - context processing.
Developers don't want to register, pay, and maintain interfaces for each model. This is where OpenRouter's value lies. With a single API key, everything is taken care of.
Mo Jomaa, a partner at CapitalG, said something after this round of investment. He compared OpenRouter to Cloudflare, Stripe, and Databricks. He said that every time there is a platform shift, an infrastructure gap emerges.
Cloudflare filled the gap in the Internet era; Stripe did so in the payment era; Databricks in the data era. CapitalG is betting that OpenRouter will fill the gap in the AI era.
Atallah also used a similar analogy. He said that OpenRouter does the same thing as Stripe. Stripe abstracts payment processors, so developers don't need to directly interact with banks.
OpenRouter abstracts model providers, so developers don't need to directly connect with each AI company. This analogy is quite apt.
However, a 26 - times PS ratio can't be supported by just the "AI - version Stripe" story.
There's another aspect. OpenRouter has a public model ranking list. You can see in real - time which models are being called the most, which are rising, and which are falling.
This ranking list is becoming an industry benchmark. Investors use it to assess the market position of model companies, developers refer to it to choose models, and researchers cite it for industry analysis.
Once a platform becomes a standard of measurement, the data it accumulates becomes an asset. With 8 million users making choices on the platform every day, these choices form a real - time updated map of the AI industry. This is probably what the 26 - times PS ratio is really pricing.
02
The map is valuable. What's happening on the map is even more worth paying attention to.
I checked OpenRouter's ranking list from April 2026 and found some interesting things.
The top - ranked model is Xiaomi's MiMo - V2 - Pro, which processes about 4.92 trillion tokens in a week. The second - ranked is Claude Sonnet 4.6, with less than half of Xiaomi's volume. Four out of the top five models are from Chinese companies.
How fast has this change been? I found that a year ago, the traffic share of Chinese models on OpenRouter was less than 2%. Now it's over 45%. From less than 2% to 45% in 12 months.
You might think this means the quality of Chinese models has caught up. That's partly true, but not entirely.
I noticed another set of data. Anthropic's models account for only 12% of the token traffic on OpenRouter, but they generate 46% of the platform's revenue in dollars.
12% of the volume and 46% of the revenue. When you put these two numbers together, a clear picture emerges.
Two completely different markets are emerging on OpenRouter. One is a high - volume commodity market. Chinese models have aggressive pricing, with many costing less than $1 per million tokens. Developers are attracted by the low price, and the token volume has soared.
The other is a high - quality, high - end market. Models from Anthropic and OpenAI are more expensive. Although the number of users is relatively small in terms of volume, those who are willing to pay a high price are here.
One is like selling commodities, and the other is like selling luxury goods on the same shelf.
This isn't the most crucial part. I also dug into a detail. OpenRouter released a research report based on 100 trillion tokens. It found that on this platform, there is an almost inverse relationship between a model's benchmark score and its actual market share.
Models with high benchmark scores aren't necessarily used more. Cheap models are the ones that get more usage.
The companies that have gained market share the fastest, such as Xiaomi, Alibaba, MiniMax, and DeepSeek, have one thing in common: the lowest pricing.
If this finding were just a data bias on OpenRouter, it could be ignored. But another set of independent data points in the same direction.
Epoch AI conducted a study on the changing trend of AI inference prices. The conclusion is that, for models with the same performance level, the median inference price drops by 50 times per year. After 2024, this rate accelerated to 200 times.
Here's a specific example:
When GPT - 4 was first released in 2023, it cost $30 per million tokens to achieve its performance level on PhD - level scientific questions. Now, using an open - source alternative to achieve the same score costs less than $0.1.
A 300 - fold decrease in three years.
The AI Index Report from Stanford University in 2026 also made a judgment. The score gap between the top four global models (Anthropic, xAI, Google, and OpenAI) on the Arena ranking list is less than 25 points. Looking further down, Alibaba and DeepSeek are also catching up.
The capabilities of top - tier models are converging, and the prices at the lower end are dropping rapidly.
David Cahn, a partner at Sequoia Capital, said that this situation is starting to resemble the airline industry: high fixed costs, low marginal costs, and no monopoly. In this structure, competition will eventually drive prices down to the marginal cost.
All these trends, such as the sharp decline in inference prices, the convergence of model performance, and the market share being grabbed by cheap models, exist independently. But when they occur simultaneously on the same platform, they create a cumulative effect.
This is what I think is the most notable thing.
Everyone says that OpenRouter is a "neutral infrastructure", a "water - carrier", or an "API transporter". This description isn't wrong, but it misses one thing: OpenRouter is actively accelerating the commodification of models.
How can we understand this?
When over 400 models are placed on the same "shelf" and sorted by price, usage volume, and latency, a model is no longer a "brand". It becomes a commodity that can be compared in price.
When developers open OpenRouter's page, they see information like "this model costs $0.5 per million tokens, has a latency of 120 milliseconds, and its usage volume increased by 30% last week". The brand fades away, and the parameters remain.
What the router does is essentially the same as what a supermarket shelf does. A supermarket puts all shampoos together, attaches price tags, and indicates the capacity. Consumers stop being loyal to a particular brand and start comparing cost - effectiveness. Shampoo manufacturers are forced to lower prices, offer promotions, and introduce larger packages.
OpenRouter does the same thing to AI models. It makes all models measurable by the same standard, and the most important factor in this standard is price.
The trends emerging from the ranking list data, such as cheap models gaining market share and the inverse relationship between benchmark scores and market share, are amplified by this "shelf" structure.
The router gives developers the right to choose, and developers vote with their feet for cheap models. As the market share of cheap models increases, expensive models are forced to lower prices or accept a shrinking market share.
Once this cycle starts, all models are moving towards becoming "commodities".
I read an interesting statement this morning:
When five manufacturers offer near - cutting - edge performance at a price of less than $1 per million tokens, the basic AI capabilities have been commodified. The competitive advantage no longer comes from which model you choose, but from how you use the model.
From another perspective: when the differences between models are so small that they don't affect usage, the need for a router also diminishes. You don't need a router to choose from 400 almost identical models.
03
Someone has experienced this story before: Atallah himself.
I looked into what happened to OpenSea later. By the end of 2023, its monthly revenue dropped from a peak of $125 million to $3 million.
Investor Coatue directly wrote down the value of its holdings by 90%. The team was reduced from 175 to 60 people. By June 2025, the monthly trading volume was only $120 million, 3% of its peak.
The technology didn't fail, and there were no scandals in the management. The problem was on the supply side.
Data from 2024 showed that 96% of NFT projects had been declared dead. The seemingly ever - growing supply was essentially built on speculative capital.
Once the capital withdrew, the supply collapsed, and there weren't enough valuable items on the "shelf" to trade. The central hub lost its reason for existence.
The question now is: Is the fragmentation of AI models the same as this?
On the surface, it's not. AI models have real demand. Developers really need different models for different tasks. In March 2026, 12 large models were released within a week. This is real production capacity.
But think more deeply.
The fragmentation of NFTs was false. After the bubble burst, the supply disappeared, and the central hub ended up with an empty "shelf". The fragmentation of AI is real. But precisely because it's real, following its own logic, it's moving towards commodification.
This is a completely different kind of danger.
OpenSea faced a collapse in supply. The "shelf" was empty, and there was nothing to sell. OpenRouter may face supply homogenization. The "shelf" is full, but the items are becoming more and more similar.
When the performance and price of all models converge to a certain extent, developers will simply choose the cheapest one.
Two different paths, but the end may be the same.
This is probably the paradox that Atallah is facing. The more successful his central hub is, the faster the models will be commodified. The deeper the commodification, the lower the developers' dependence on him.
The way he wins may also be the way he loses.
The last time he worked on OpenSea, this paradox didn't have a chance to unfold because the NFT market crashed before the fragmentation was proven false. This time, with OpenRouter, the AI market has given him a longer runway. The fragmentation is real, the demand is real, and the growth is real.
But this also means that the paradox has enough time to play out.
In the future, if you see a business like this, where a company acts as a central hub in a fragmented market, you might want to ask a question first: Is this fragmentation temporary or irreversible?
If it's temporary, the ceiling will be reached when the fragmentation subsides. If it's irreversible, then ask a second question: Will the end of fragmentation lead to differentiation or commodification?
The answers to these two questions will determine whether the central hub can survive the cycle it serves.
Personally, I think OpenRouter isn't a great business model, but it's a good business. It makes money from the time and information gap.