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

Author: Diyaofan

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 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 role 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, one 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-founder 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 chaotic NFT inventory… AI today is similar.)

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

In a podcast, Alex said something that left a deep impression: 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 was: OpenAI is far ahead, and other models are just followers.

But Alex was right.

Today, there are over a thousand open-source models. Claude, Gemini, Llama, Mistral, DeepSeek… new players emerge 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 is a visible trend in the AI market: “Inference” will replace “training” as the main driver.

The difference between inference 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 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 a unique advantage: a leaderboard.

After processing over 100 trillion tokens, they know:

• Which model is best at coding

• Which model offers the best value

• Which model suddenly excels at specific tasks

This leaderboard 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 learned: 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.

Millions of developers call different models here daily. OpenRouter knows:

• Which model performs best for which task

• Which provider is most stable

• Which time period is cheapest

This data fuels 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 this data flywheel starts, it’s hard for later entrants to catch up.

Andrej Karpathy’s post on X about OpenRouter’s LLM rankings

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 “markup.”

This model aligns with Western intermediary business practices:

2. Neutral stance: If OpenRouter owns models themselves, would you trust their leaderboard?

3. Market-driven growth: The bigger the AI market, the more they earn

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

Alex’s words: “We want developers not to feel vendor lock-in. We want them to feel like they have a choice and 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 AI, anytime.)

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 advantages, it’s necessary to acknowledge some issues with this model:

OpenRouter’s core strengths are “data” and “community.” The flywheel has started turning (more users → better data → more valuable leaderboard), but this model also means its ecosystem’s prosperity depends heavily on the number of small and medium developers.

This business’s prosperity relies on more and more small and medium developers, who lack the time for aggregation development and the scale to negotiate prices with AI vendors, thus needing 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 tend to avoid it, such as the open-source alternative LiteLLM, which is free and self-deployable.

Cost-sensitive developers might ask: “Why give you 5%?”

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

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

Of course, it’s still early days, and rapid growth will continue. Its ceiling is a question to consider in analysis.

One-minute overview of OpenRouter

Q1: What is OpenRouter?

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

Q2: How does OpenRouter differ 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 leaderboard and broader provider coverage.

Q3: Who is the founder of OpenRouter?

Alex Atallah, Stanford CS 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 the platform to anonymously test new models to gather unbiased developer feedback. This indicates that the OpenRouter community has some influence in the industry.