He sold his 2.2 billion fortune before the NFT crash and then quickly moved into the hottest AI track.

Author: Di Yafan

In 2022, a Stanford graduate’s net worth once exceeded $2 billion.

He created OpenSea, the world’s largest NFT marketplace, valued at $13.3 billion.

Just a few months before the NFT bubble burst, he made a more critical decision: to leave.

Two years later, his new company grew tenfold in 7 months, secured investments from a16z, Sequoia, and Menlo Ventures, and reached a valuation of $500 million.

His name is Alex Atallah. His new company is OpenRouter.

This is a story about timing and methodology replication.

Who is OpenRouter? What does it do?

If you are an AI application developer, you’ve probably heard of OpenRouter. Its main function is to help developers solve the pain points of model switching:

• Want to code with Claude but find it often lacks capacity

• Want to analyze with GPT but the cost is painful

• Want to try open-source models but need to rewrite API integrations

Each model provider’s API is different. Every time you switch models, you have to modify your code.

OpenRouter’s function is similar to Ctrip, bringing all airlines into one app.

One API, access to over 300 models. More than 60 providers. Switch models? Change one line of code.

OpenRouter as a Multi-Model Aggregation Layer

Two startups, same methodology

Before starting his entrepreneurial journey, Alex Atallah had a solid software background: Stanford Computer Science, Palantir engineer, co-founder and CTO of OpenSea…

OpenSea co-founders Alex Atallah (left) and Devin Finzer (right)

He explained in a podcast that both startups shared a common approach:

“OpenSea organized this very heterogeneous inventory and put it together in one place… You see a lot of those similarities with how AI works today.” (OpenSea consolidated a very chaotic NFT inventory into one place… AI is the same today.)

What is his methodology?

Identify a “fragmented ecosystem” and build an “aggregation layer”.

• In the NFT era: metadata standards vary → OpenSea aggregates

• In the AI era: API standards vary → OpenRouter aggregates

Alex once said in a podcast that left a deep impression on me: If training a large AI model costs only $600, then in the future, there could be tens of thousands or even hundreds of thousands of models. They will need their own ‘market’.

Early 2023, this was an extremely contrarian view. The mainstream narrative then was: OpenAI is far ahead, and other models are just followers.

But Alex was right.

Today, there are over a thousand open-source models alone. Claude, Gemini, Llama, Mistral, DeepSeek… new players enter every few weeks.

In a world of explosive model growth, an “aggregation layer” is needed. That’s exactly where OpenRouter fits.

An underestimated huge market

Behind OpenRouter’s success lies a visible trend in the AI market: “Reasoning” will replace “Training” as the main driver.

The difference between reasoning and training, and the future trend of this market, was clearly explained in Groq’s recent analysis—worth checking out.

COO Chris Clark’s perspective is worth noting:

“We believe that inference costs will eclipse salaries as the dominant operating expense for most knowledge-based companies over the next five to 10 years.” (We believe that in the next 5-10 years, AI inference costs will surpass wages, becoming the largest operational expense for knowledge-based companies.)

This can be seen from OpenRouter’s own data.

OpenRouter’s token consumption approaches 80 trillion

A well-known “mass-market” AI model

As one of the earliest players in this space, OpenRouter has an exclusive advantage: the ranking list.

After processing over 100 trillion tokens, they know:

• Which model is best at coding

• Which model offers the best cost-performance ratio

• Which model suddenly excels at specific tasks

This ranking has become an industry benchmark and is highly recognized within the developer community.

Even more astonishing: in April 2025, a mysterious model called “Quasar Alpha” was launched on OpenRouter.

A few days later, everyone found out: It’s GPT-4.1, exclusively launched by OpenAI on OpenRouter.

Because OpenRouter possesses a killer asset: the largest multi-model usage dataset on the internet.

Every day, millions of developers call different models here. OpenRouter knows:

• Which model performs best for which task

• Which provider is most stable

• Which time period is cheapest

These data have created the industry’s most authoritative LLM ranking. According to Menlo Ventures, even Andrej Karpathy (former Tesla AI director and OpenAI co-founder) has publicly recommended it.

Once the data flywheel starts turning, it’s hard for latecomers to catch up.

Andrej Karpathy mentions OpenRouter LLM rankings on X

How does OpenRouter make money?

OpenRouter’s business model is relatively simple: You spend $100 on models, they take $5.

They charge based on the pricing set by model providers. They earn “toll fees,” not “margin.”

This model aligns with Western intermediary business practices:

2. Maintaining neutrality: If OpenRouter has its own models, would you trust its rankings?

3. Growing naturally with the market: The larger the AI market, the more revenue they generate

4. Network effects: More users → better data → more valuable rankings → more users

Alex’s words: “We want developers to not feel vendor lock-in. We want them to feel like they have choice and they can use the best intelligence, even if they didn’t before.” (We don’t want developers to be locked into vendors. We want them to have options and access to the best intelligence at any time.)

Financial data (disclosed)

8 people, nearly $100 million GMV annually.

This person’s efficiency is among the top in similar startups.

Big market, small space

After highlighting the strengths, it’s necessary to discuss some issues with this model:

OpenRouter’s core advantages are “data” and “community.” The flywheel has started turning (more users → better data → more valuable rankings), but this model also means its ecosystem’s prosperity is closely tied to the number of developers.

This business’s prosperity depends on more small and medium developers emerging, as they lack the time for aggregation development and the scale to negotiate prices with AI vendors, so they need an intermediary.

For large companies, it might have some testing value, but once scaled, they will likely bypass it.

In fact, even medium-sized projects with larger usage will want to avoid it, such as the open-source alternative LiteLLM, which is free and self-deployable.

Cost-sensitive developers will ask: “Why should I give you 5%?”

If competition intensifies, this fee could drop to 3%, or even 2%.

Whether it can sustain the current high valuation of 100x remains uncertain.

Of course, it is still in early stages and will continue to grow rapidly. Its ceiling is a question to consider in future analysis.

One-minute overview of OpenRouter

Q1: What is OpenRouter?

OpenRouter is a large language model API aggregation platform. Through a single API interface, developers can access over 300 models (including GPT-4, Claude, Llama, etc.) without integrating each provider’s API separately.

Q2: How is OpenRouter different from LiteLLM?

Both provide LLM API aggregation, but with different models. OpenRouter is a managed SaaS charging a 5% fee; LiteLLM is open-source, deployable locally, and free. OpenRouter’s advantage is its public model ranking and broader provider coverage.

Q3: Who is the founder of OpenRouter?

Alex Atallah, Stanford Computer Science graduate, former co-founder and CTO of OpenSea (the world’s largest NFT marketplace). He left OpenSea in 2022 and founded OpenRouter in 2023. His personal net worth once exceeded $2 billion.

Q4: How much funding has OpenRouter raised?

As of June 2025, OpenRouter completed a total of $40 million in funding (seed + Series A), led by a16z and Menlo Ventures, with Sequoia participating, valuing around $500 million.

Q5: Why does OpenAI test new models on OpenRouter?

According to OpenRouter, OpenAI has used its platform to anonymously test new models to gather unbiased developer feedback. This indicates that the OpenRouter community has a certain influence in the industry.