OpenRouter, the leading large-model API aggregator, is reported to be seeking a buyer and talking to several technology giants. The twist: just two months earlier it closed a 113 million dollar Series B led by Google’s CapitalG, with Nvidia, Snowflake, Databricks and ServiceNow alongside, at a 1.3 billion dollar post-money valuation, processing over 100 trillion tokens a month with annualised revenue above 50 million dollars.
In most scripts, the undisputed king of multi-model distribution should sprint to Nasdaq. Instead it is pushing itself onto the merger market at the peak of its business and reputation. Meanwhile China’s own OpenRouter, SiliconFlow, has filed for a Hong Kong listing at a loss. Why do the two take opposite paths?
What OpenRouter got right
In 2023 the model market moved from OpenAI’s single dominance to an open-source explosion, and no single model could win on cost, latency, reasoning, code and context at once. Enterprises drifted into mixed-model architectures, and wiring APIs for hundreds of models, managing dozens of supplier bills and handling rate limits and outages became a nightmare. OpenRouter appeared with one standard endpoint reaching more than 400 models, the traffic matchmaker of the token era.
It routes requests in milliseconds when a provider throttles or fails, picks the best-price endpoint by prompt complexity, and consolidates billing and permissions. By mid-2026 it served over 8 million developers and enterprises, handling 25 trillion tokens a week, effectively the scheduling and settlement hub of North American model infrastructure.
Two hidden faults
The business logic is blunt: a roughly 5 per cent fee on top of provider price. But two faults sit beneath. First, scale revenue is not profit; OpenRouter’s true annualised revenue was only about 50 million dollars in early 2026, below the threshold even mid-market US SaaS firms such as Snowflake or Databricks cleared before their IPOs. Second, its pricing power is fragile; push the fee up and large clients build their own routers or sign direct cloud deals.
Worse than low margin is the pipeline risk. OpenRouter sits squeezed between upstream model giants and downstream cloud vendors. If OpenAI ships built-in routing, or clouds deep-bundle models, traffic is intercepted at the source. Amazon Bedrock, Google Vertex and Azure AI Studio already offer multi-model access, unified billing and smart scheduling, and treat routing as a free lure to keep compute and data inside their clouds.
Why giants still want it
On public markets a pure intermediary rarely supports a multibillion-dollar listing. But on a strategist’s board, OpenRouter is a prize. Whoever buys it gains cross-model behaviour data no single product backend shows: which model is chosen in which scenario, how fast traffic shifts after a price change, which startup’s API is exploding and worth buying early.
Controlling OpenRouter is controlling the dispatch valve of global AI traffic. Its shareholders, Google, Nvidia, Snowflake and the rest, are exactly the likely buyers. For Snowflake it is a neutral routing layer; for clouds and Nvidia it steers 8 million developers and 100 trillion tokens toward their own infrastructure.
Why China’s twin chose an IPO
The contrast reflects three gaps between the two AI ecosystems. On business model, US infrastructure is standardised on Nvidia and CUDA, so OpenRouter is light-asset plumbing easy for AWS or Google to disrupt; China runs a fragmented mix of Huawei Ascend, Biren, Moore Threads and Cambricon, where the same open model performs differently per chip, so SiliconFlow must also do cross-chip adaptation and inference optimisation, a harder moat. On competition, US model power sits with a few closed giants controlling traffic, while China’s open-source models need third-party deployment. And on delivery, Chinese banks, state firms and manufacturers pay for private deployment, a higher-margin line absent in OpenRouter’s model.
OpenRouter proved, with sharp timing, the strategic value of the routing layer in a multi-model world. China’s OpenRouters, rooted in heterogenous chips and private delivery, show a different survival path. Both, say insiders, are simply the best answer to their own soil.
Editor’s note: translated and adapted for RobotBelt from LeiPhone. Read the original report here.