A $113 million funding round answers two long-standing, ambiguous questions in the AI industry: what is the unit of settlement for this industrial revolution, and who will build its infrastructure?
The $1.3 billion valuation doesn't represent a single company, but rather the value of a position within the token economy.
On May 26, 2026, OpenRouter completed a $113 million Series B funding round. What it does is simple—allowing developers and enterprises to access over 400 AI models through a single API, with unified billing and intelligent routing. It doesn't build its own models, doesn't sell computing power, and earns its profits through commissions.
On May 26, 2026, OpenRouter completed a $113 million Series B funding round.
With 8 million users, its scale has quintupled in six months. But look at who's investing: Google's CapitalG led the round, NVIDIA Ventures followed, and strategic investment departments from ServiceNow, MongoDB, Snowflake, and Databricks all entered the fray. a16z and Menlo Ventures both increased their investments. One is a model maker, the other a chipmaker, both appearing on the investor list of a token distribution platform—and Google's move is far more than just financial investment. In the same month, Palo Alto Networks acquired its competitor Portkey, valuing it at $120-140 million. Within the same sector, both independent financing and M&A exits are emerging simultaneously. This isn't about pricing a company. It's about pricing a position—the intermediary layer that allows tokens to circulate. What will this intermediary layer ultimately become? A router, an electricity grid, or a Taobao for the token economy? CapitalG Partner Mo Jomaa: "Every platform shift creates an infrastructure gap: the internet era had Cloudflare, digital payments had Stripe, and the data era had Databricks. OpenRouter is the infrastructure gap for the AI inference era." Mo Jomaa is right. But a more pressing question is: why specifically tokens? I. Token: The Settlement Unit in the AI Era Tokens are not the fuel of AI. They are the measure of AI. Every industrial revolution requires something to be standardized, quantified, priced, and traded before the infrastructure surrounding it truly begins to be built. Before coal was sold by the ton, it was just ore; before electricity was sold by the kilowatt-hour, it was just a laboratory phenomenon; before bandwidth was sold by the gigabyte, it was just an academic network. The key to transforming resources into economic units lies not in their use, but in the establishment of a system of measurement—only with a system of measurement can a market exist. For the past two years, the measurement of AI has been vague: computing power, GPUs, model capabilities, data—all are correct, but none are precise enough. Those are production tools, not units of account. In March 2026, at the GTC conference, Jensen Huang provided the answer. He no longer called data centers data centers; he called them "Token Factories"—the raw materials are electricity and data, and the basic economic unit produced is a token. The factories of the last industrial revolution produced electricity; this round of AI factories produces intelligence. The greatness of electricity lies not in discovering a new energy source, but in creating a universal interface—it transforms primary energy sources like coal, hydropower, and oil, which are not interchangeable, into a standardized form that can be transmitted, stored, controlled, and plug-and-play. AI is doing the same thing with intelligence. A lawyer's analytical abilities, a programmer's coding skills, a doctor's diagnostic abilities—this kind of intelligence used to exist only in the individual brain, unable to be stored, taken away when leaving a job, costly and unstable to access, and impossible to trade across borders. Tokens, for the first time, extract these from the human brain, turning them into standardized APIs that can be accessed 24/7. Electricity standardized energy, and tokens standardized intelligence. But there's a fundamental difference between tokens and electricity. Electricity plays only one role: a carrier of energy. It transmits energy, contains no information, doesn't carry judgment, and doesn't constitute a product. Tokens are different—they are simultaneously the fuel for AI (inference consumption), the product (model output), the unit of account (business model), and a strategic resource (the focus of geopolitical competition)—a four-in-one entity. Almost no other commodity in modern economic history simultaneously plays all four roles. This assessment has already been confirmed at the policy level. In the same month, Liu Liehong, director of the National Bureau of Data Science and Technology, named the token "Ciyuan" in Chinese, calling it a "settlement unit" connecting technology supply and business demand. The data he disclosed confirmed this acceleration: China's daily Ciyuan usage jumped from 100 billion in early 2024 to 140 trillion in March 2026, a more than 1,000-fold increase in two years. On May 22—four days before OpenRouter's funding announcement—the National Bureau of Data Science and Technology held a "Ciyuan Economy Symposium," explicitly incorporating the promotion of Ciyuan economic development into its work system. When a settlement unit is priced by the market and named by the state, the infrastructure race surrounding it has already begun. II. What OpenRouter Did The person who built the largest marketplace in the NFT industry saw the same structural opportunity in the AI token market—the methodology was exactly the same. OpenRouter founder Alex Atallah was the co-founder and CTO of OpenSea, founded in 2017 with Devin Finzer and left in 2022. The essence of OpenSea was traffic entry + standardized interface + commission—buyers didn't need to know which blockchain the NFT was minted on; the platform handled everything. Atallah transferred this framework to AI: the shelf space changed from JPEG to LLM, and the pricing changed from ETH to tokens. The platform he built is called OpenRouter, officially positioned as an "AI model exchange"—allowing developers and enterprises to access over 400 models from vendors such as Anthropic, Google, OpenAI, xAI, and DeepSeek through an API, selecting models on demand, with intelligent routing and unified billing. Currently, it serves over 8 million users and processes 25 trillion tokens weekly—six months ago, this number was 5 trillion.

(An unfiltered conversation with Alex Atallah, CEO of OpenRouter)
Why is it growing so fast? In its post-funding blog, Menlo Ventures dissected a core insight: Large-scale AI adoption is fundamentally a multi-model problem. You don't need a Ferrari to go grocery shopping, but you also can't ride a bicycle on the highway—small, inexpensive models are sufficient for document summarization, flagship models are needed for deep inference, and multimodal computing requires a completely different approach. A 2026 Deloitte survey showed that 67% of enterprises consume more than 1 billion tokens per month, and F5's research showed that the average enterprise runs 7 different AI models simultaneously. Furthermore, the same model running on different vendors exhibits varying latency, price, and even output quality. Furthermore, most LLM APIs can't even achieve 90% uptime, and each vendor has its own separate billing and key management system—enterprises don't lack models. What they lack is a middleware layer that makes all models usable. Andrej Karpathy calls OpenRouter an AI "transfer switch": it doesn't produce electricity or manufacture appliances, but it determines where the electricity comes from and how it's distributed. However, its value goes beyond routing. What OpenRouter essentially does is equip each machine with an independent motor: allowing enterprises to organize AI capabilities according to business needs rather than model vendors. This is not just about reducing costs and improving efficiency; it's a revolution in organizational methods. A set of production data recently released by Vercel AI Gateway confirms the validity of this logic. In April 2026, by spending amount, Anthropic accounted for 61%, Google for 21%, and OpenAI for 12%; however, by token quantity, Google accounted for 38%, Anthropic for 26%, and OpenAI for 13%. Cheap models consume volume, expensive models consume money—the same customer group chose completely different winners in two dimensions. This is the true meaning of "multi-model": not choosing the best one, but choosing the most suitable one for each scenario. III. Why is this position the most important entry point in the token economy? When no single model can handle all scenarios, the value of the middle layer emerges. This judgment is not merely an analyst's deduction.
3.1 Collective Judgment of Investors

(OpenRouter Raises $113M Series B)
The investor lineup in this round of financing is not just "a group of VCs who are optimistic about a company"—it is AI The collective alignment of upstream, midstream, and downstream players in the industry chain. CapitalG (Google/Alphabet) led the investment. Google has its own Gemini, but it's betting on a neutral model distribution platform. This isn't contradictory—Google has Gemini but can't win all customers; supporting a neutral distribution platform ensures that Google models are also distributed there. CapitalG partner Jane Alexander put it more directly: "OpenRouter has a unique positioning, serving as a data clearing center and unified intelligence layer for AI models." However, to interpret this merely as "Google wanting to distribute Gemini" underestimates this move. Let's look at Google's three moves in May together: CapitalG led a funding round for OpenRouter (routing layer – who processes the agent's code), donated the AP2 protocol to the FIDO Alliance (protocol layer – which wallet the agent uses for authorization), and released Universal Cart at the I/O conference (entry point layer – whose ecosystem the agent completes the purchase within). The protocol layer is open; Google doesn't need to win the protocol war. What it needs is to become the default option at both the routing and entry points. When the agent's model selection, authorization acquisition, and purchase completion all fall within Google's sphere of influence, it's not just controlling the protocol, but controlling the three crucial checkpoints through which traffic must pass. NVentures (Nvidia) also participated in the investment. Nvidia's interests are straightforward—more tokens consumed = more inference computing power used = more GPUs purchased. OpenRouter is an amplifier of inference power. This is the same logic as Nvidia's investment in CoreWeave: it doesn't matter whose model wins, it cares about the overall scale increasing. The strategic investment departments of ServiceNow, MongoDB, Snowflake, and Databricks have all entered the fray. These are all enterprise software infrastructure companies. Their simultaneous investment from their strategic investment departments indicates that enterprise customers are upgrading multi-model orchestration from a "technology choice" to a "standard infrastructure feature." a16z, Menlo Ventures, and Sequoia have all increased their investments. When Menlo invested a year ago, OpenRouter had 2.5 million developers and processed approximately 100 trillion tokens annually. Currently, there are 8 million developers and approximately 1.5 quadrillion tokens processed annually. Menlo's Deedy Das made a striking comparison: OpenRouter's token throughput is already equivalent to 15-30% of Google's, 20-40% of OpenAI's, and over 50% of Azure Foundry's. He also revealed that revenue doubled from the signing in February to the announcement in May. When model providers (Google), chip manufacturers (Nvidia), enterprise software companies (ServiceNow/Databricks/Snowflake/MongoDB), and top VCs (a16z/Sequoia/Menlo) all appear on a single cap table, it's not a sign of confidence in one company. This is the entire industry chain voting with real money. 3.2 Industry Data Evidence TechCrunch's assessment is: "OpenRouter's success means that AI models are becoming invisible, interchangeable engines. Enterprises are no longer locked into a single vendor as they were in the SaaS era. The multi-model future has arrived." The rise of Chinese models further solidifies this assessment. During the 2026 Spring Festival, Chinese models once accounted for 61% of OpenRouter platform token consumption. When American and Chinese models are available for global developers to purchase on demand on the same platform, multi-model is not a trend, but a reality. But the key here isn't "multi-model" itself—it's who controls the entry point between the model and the user. OpenRouter's rankings are already one of the most widely cited model adoption metrics in the global AI industry—investors, researchers, and media all keep an eye on this list to gauge trends. When the entire industry uses your data to make decisions, you're not just a router; you're the information hub of the entire ecosystem. This isn't an add-on feature; it's structural power: every route is a data collection, every traffic record is a real-time map of global AI demand. In the age of electrification, the ultimate power lies not in power plants, but in the power grid. In the internet age, it's not websites, but search engines and app stores. In the token economy, the laboratory for producing models is the power plant, the enterprises and agents using the models are the terminal appliances, and the routing layer sitting in the middle, deciding "which request goes to which model"—is the power grid. Whoever controls this entry point controls the map of the token economy. IV. The Future of the Token Economy: What Can E-commerce Teach Us? Headless merchants have emerged. But no one has answered where they should be. Noah Levine of a16z defined this concept in March: no storefront, no account system, no sales team, just one server, a set of API endpoints, and a pay-per-use price. In its first week, Stripe's Machine Payments Protocol saw over 60 headless services listed, with 894 agents completing 31,000 transactions. Visa's Cuy Sheffield and Levine discussed the same trend on a podcast—pay-per-token is replacing subscription models, and agents are replacing consumers. The definition is there, the protocol is there, and transactions have occurred. But where are these headless merchants? 60 services could be in one directory. 6,000? 600,000? In theory, an agent can crawl the entire web to find API endpoints, but in practice, it needs a place with rankings, reliable data, and unified billing for discovery and comparison. This place has a name: marketplace. OpenRouter today is the early form of this marketplace—except that currently, it only offers one type of headless merchant: LLM providers. If we follow OpenSea's path from a single category (headshot PFP) to all digital asset categories, headless merchants will eventually expand their categories from LLM to image generation, data retrieval, document processing, and payment verification—this is a prediction, but not one without precedent.

If we use the development path of e-commerce to make predictions, the evolution of token marketplaces will most likely go through three stages:
The first stage: the marketplace establishes standards.
The first stage: the marketplace establishes standards.
The first stage: the marketplace establishes standards.
Just as Taobao defined its product catalog, search ranking, credit rating, and unified payment in its early days, OpenRouter is using massive amounts of transactions to define "what an LLM API should look like"—naming conventions, pricing formats, usability metrics, and context parameters. These weren't designed by a committee; they were developed through extensive transaction practice. Standards cannot be cold-started. The second phase: Deepening the ecosystem. This is currently OpenRouter's thinnest layer. Referring to e-commerce: Taobao's moat isn't its search function, but its store design, paid advertising, positive review system, and Alipay's Huabei. OpenRouter still lacks a recommendation engine (evolving from routing to "guess what you need"), scenario-based evaluation (not just ranking by total call volume, but scoring by scenario (legal/code/translation), financial tools (token pre-purchase, budget management, consumer credit), and a third-party ecosystem (plugins, fine-tune hosting, prompt marketplace). These combined constitute its competitive advantage. The third stage: A2A emerges from the marketplace. When standards are mature enough and trust data is abundant enough, leading customers will "graduate"—directly connecting with headless merchants, no longer going through the marketplace. Just like large brands graduating from Taobao to open independent Shopify websites, and large enterprises using LiteLLM to build their own routers. But the marketplace won't disappear—it will become a discovery layer and a trust layer, with small and medium-sized customers and long-tail services remaining on the platform forever. Historically, this has always been the order: first, centralized exchanges (NYSE) establish clearing standards, then electronic payment and dark pools emerge; first, SWIFT and Visa establish payment networks, then P2P and on-chain settlement; first, Amazon and Taobao establish e-commerce standards, then Shopify allows merchants to operate independently. Centralization builds standards and trust first, then decentralization inherits those standards and trust. This is the real gamble OpenRouter faces. A 5% take rate is a significant cost for large customers—for enterprises with monthly inference spending exceeding approximately $37,000, building their own using the open-source LiteLLM is more cost-effective. Vercel AI Gateway has already achieved zero-token markup, and Cloudflare has made AI Gateway a free built-in feature. Trends.vc's assessment is quite sober: "Once a well-funded player commits to passthrough pricing, competitors will follow suit. Profits will shift from routing itself to caching, governance, and integration depth." The ultimate answer in the e-commerce industry is: Taobao won the debate of "why not build your own website to sell," not because its search function was superior, but because the buyers were there, the reviews were there, and the trust was there. Whether OpenRouter can reach that point depends on whether it can build an ecosystem so robust that people can't live without it before take rates are compressed. The $1.3 billion valuation isn't about today's 5% commission—it's about whether this evolution from marketplace to operating system can be completed. Whether OpenRouter can become the ultimate winner is an open question. But this round of financing marks an irreversible turning point—standards are being defined by transaction volume, and trust is being built by the platform. The question is no longer "whether to build token infrastructure," but rather who builds it and how that infrastructure will be inherited. As for whether this is the Taobao era for token economics—the answer depends on who builds a robust ecosystem first. The window of opportunity has opened.