OpenRouter closed a $113 million Series B on May 26, led by Alphabet’s independent growth fund CapitalG, with participation from NVIDIA’s NVentures, ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, and Databricks Ventures. TechCrunch reported the post-money valuation at approximately $1.3 billion, more than double the $547 million it reached after its $40 million Series A in June 2025.
The funding validates multi-model routing as critical AI infrastructure. OpenRouter now serves over 8 million users and processes 25 trillion tokens per week, a 5x increase from six months ago, according to the company’s press release. CEO Alex Atallah called multi-model routing “a permanent infrastructure requirement,” telling Yahoo Finance: “The era of picking a single model is over.”
But the platform’s own traffic data surfaces a question its investors may not have fully priced in.
Where the Tokens Are Going
Chinese-developed open-weight models held roughly 61% of token volume among OpenRouter’s top 10 models during the week of February 24, 2026, according to TechTimes. By April, combined Chinese-provider traffic still accounted for approximately 51% of all platform tokens, per analysis from Digital Applied. The shift took eighteen months: Chinese models went from under 2% of OpenRouter traffic in late 2024 to consistent majority leadership.
The concentration is structural, not random. OpenRouter COO Chris Clark told TechTimes that Chinese open-weight models are “disproportionately heavy in agentic flows run by U.S. developers.” Programming workloads grew from roughly 11% of total OpenRouter token volume to more than 50% through 2025, and agentic workflows now account for more than half of all output tokens on the platform, according to AICost.org.
The economics explain the pattern. MiniMax M2.5 charges roughly $0.30 per million input tokens and $1.20 per million output tokens. Claude Opus 4.6 charges approximately $5 per million input and $25 per million output. For agentic workloads that invoke a model thousands of times per session, that gap compounds from a rounding error into a budget line.
Benchmarks Are Closing, but Caveats Remain
The capability gap between Chinese open-weight models and Western proprietary models narrowed substantially in Q2 2026. Moonshot AI’s Kimi K2.6 and Xiaomi’s MiMo-V2.5-Pro both scored 54 on Artificial Analysis’s Intelligence Index, placing them 3 to 6 points below GPT-5.5 (60), Claude Opus 4.7 (57), and Gemini 3.1 Pro Preview (57). On SWE-Bench Pro, Kimi K2.6 scored 58.6%, ahead of GPT-5.5 at 57.7%, per The Batch.
Those figures come from Moonshot AI’s own benchmarking, and independent third-party verification had not been published as of early May 2026. The Stanford HAI 2026 AI Index put the US-China performance gap at 2.7% as of March 2026, while noting invalid question rates on major benchmarks range from 2% to 42%. Kili Technology found a 37% gap between lab benchmark scores and real-world enterprise deployment performance.
DeepSeek V4 Pro posted a 94% hallucination rate on the AA-Omniscience benchmark, according to Artificial Analysis. When it does not know the answer, it almost always responds anyway. That matters for agentic workflows where an agent acts on its own output.
The Legal Risk That Pricing Cannot Offset
China’s National Intelligence Law, enacted in 2017, requires all Chinese companies to “support, assist, and cooperate” with government national security investigations and intelligence collection. This applies to Moonshot AI, MiniMax, DeepSeek, Zhipu AI, Alibaba, and Xiaomi regardless of where their model weights are hosted or whether the company has incorporated a Western subsidiary, according to analysis from the American Enterprise Institute.
The House Committee on Homeland Security and the House Select Committee on China announced a joint investigation on April 29 into national security risks posed by Chinese AI models, naming Moonshot AI, MiniMax, Alibaba, and DeepSeek specifically. The investigation also targeted Anysphere (maker of the Cursor coding tool) and Airbnb for building on Chinese AI infrastructure.
For developers routing agentic coding workloads through these models, the exposure is specific: any prompt sent to a Chinese-provider API endpoint falls within the scope of a law that provides no opt-out. The obligation exists continuously and does not require a demonstrated government request before it creates legal exposure. Enterprise customers contracting directly with Anthropic, OpenAI, or through Azure and Google Cloud have largely not replicated the OpenRouter pattern. The broader 100-trillion-token study placed proprietary Western models at roughly 70% of total global API share.
The Infrastructure Bet
The $113 million raise positions OpenRouter as the routing layer for a multi-model world where developers pick models by workload, not brand loyalty. That thesis is working: $1.3 billion valuation, 8 million users, 25 trillion weekly tokens.
The governance question is whether the platform’s own data, showing Chinese models dominating the highest-volume workload category, represents a feature of the multi-model future or a vulnerability. For developers building agents that process proprietary code, customer data, or internal business logic, the answer depends on which law governs the model on the other end of the API call.