Google and NVIDIA bet on OpenRouter: The Taobao moment of token economy?

A $113 million funding round answered 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.

$1.13 billion valuation is not for a company; it’s the value of a position within the token economy.

On May 26, 2026, OpenRouter completed a Series B funding round of $113 million. What they do isn’t complicated—through an API, they connect developers and enterprises to over 400 AI models, with unified billing and intelligent routing. They don’t build models themselves, don’t sell computing power, and earn middleman commissions. With 8 million users, their scale has quintupled in half a year.

But look at who’s investing: CapitalG from Google led the round, NVentures from NVIDIA followed, strategic investments from ServiceNow, MongoDB, Snowflake, Databricks all joined in, and all of a16z and Menlo Ventures increased their stakes.

**A model builder and a chip manufacturer, appearing together on an investor list for a token distribution platform—and Google’s move here is far more than just financial investment.**In the same month, Palo Alto Networks acquired its competitor Portkey, valued at $120–140 million. In the same track, both independent funding and M&A exits are happening simultaneously.

This isn’t about valuing a single company. It’s about valuing a position—the middle layer that makes token liquidity possible. What will this middle layer grow into? A router, a power grid, or a Taobao for the token economy?

CapitalG partner Mo Jomaa: “Every platform shift creates a gap in infrastructure: Cloudflare in the internet age, Stripe in digital payments, Databricks in data. OpenRouter is the infrastructure gap in the AI inference era.”

Mo Jomaa is right about the position. But what’s more worth asking is: why specifically tokens?

1. Token: The unit of settlement in the AI era

Tokens are not the fuel of AI. They are the measure of AI.

Every industrial revolution waits until something is standardized, measured, priced, and traded before the infrastructure around it truly begins to build. 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 GB, it was just academic networking.

The key to turning resources into economic units isn’t just their use, but their measurement—once measured, markets can form. Over the past two years, AI’s measure has been fuzzy: compute power, GPUs, model capabilities, data—all are relevant but not precise enough. These are production tools, not settlement units.

In March 2026, Jensen Huang clarified the answer at GTC. He no longer calls data centers “data centers,” but “Token Factories”—raw materials are electricity and data, and the basic economic unit produced is the token. The last industrial revolution’s factories produced electricity; this new AI factory produces intelligence.

The greatness of electricity wasn’t just discovering a new energy source, but creating a universal interface—it transformed coal, hydro, oil—once incompatible primary energy sources—into a transmittable, storable, controllable, plug-and-play standardized form.

AI is doing the same for intelligence. Lawyers’ analytical skills, programmers’ coding abilities, doctors’ diagnostic skills—these have historically only existed in individual brains, unstorageable, lost when people leave, costly and unstable to call upon, and impossible to trade across borders.

Tokens are the first to extract these from human brains, turning them into standardized APIs available 24/7. Electricity standardized energy; tokens standardize intelligence.

But tokens and electricity differ fundamentally. Electricity plays only one role: energy carrier. It transmits energy but contains no information, no judgment, no product. Tokens are different—they are AI’s fuel (for reasoning), product (model output), billing unit (business model), and strategic resource (geopolitical focus)—all in one. Few commodities in economic history have played these four roles simultaneously.

This recognition has been confirmed at the policy level. Liu Lihong, director of China’s National Data Bureau, gave tokens a Chinese name, “词元” (word element), calling it a “settlement unit” connecting technological supply and commercial demand. Data shows acceleration: daily token calls in China jumped from 100B early 2024 to 140 trillion in March 2026—over a thousandfold increase in two years. On May 22, just four days before OpenRouter’s funding announcement, the National Data Bureau held a “Word Element Economy Symposium,” explicitly promoting the development of the word element economy.

When a settlement unit is priced by the market and named by the state, the infrastructure race around it has already begun.

2. What OpenRouter has done

Someone who built the largest NFT marketplace saw the same structural opportunity in the AI token market—the methodology is identical.

OpenRouter founder Alex Atallah is a co-founder and CTO of OpenSea, established in 2017 with Devin Finzer, and left in 2022. OpenSea’s essence is traffic entry + standardized interface + commission cut—buyers don’t need to know which chain their NFTs are minted on; the platform handles everything. Atallah applied this framework to AI: replacing JPEG shelves with LLMs, switching billing from ETH to tokens.

He built OpenRouter, officially positioned as “AI model exchange”—via a single API, it connects developers and enterprises to over 400 models from providers like Anthropic, Google, OpenAI, xAI, DeepSeek, with on-demand selection, intelligent routing, and unified billing. It now serves over 8 million users, processing 25 trillion tokens weekly—half a year ago, that was 5 trillion.

(An unfiltered conversation with Alex Atallah, CEO of OpenRouter)

Why such rapid growth? Menlo Ventures, in a blog post after leading the round, analyzed a core insight:

Massive AI usage fundamentally boils down to a multi-model problem. You don’t need a Ferrari to buy groceries, but you can’t ride a bicycle on the highway—simple tasks like document summarization can use cheap small models, deep reasoning requires flagship models, and multimodal tasks need a different approach.

Deloitte’s 2026 survey shows 67% of enterprises consume over 1 billion tokens monthly; F5’s research indicates each enterprise runs an average of 7 different AI models simultaneously. The same model runs on different providers with varying latency, price, and output quality. Plus, most LLM APIs can’t even guarantee 90% uptime, and each provider manages billing and keys separately—enterprises don’t lack models; they lack an intermediate layer that makes all models usable.

Andrej Karpathy calls OpenRouter an AI “transfer switch”: it doesn’t produce electricity or devices, but it determines where the electricity comes from and how it’s distributed. Its value isn’t just routing.

OpenRouter’s core is essentially installing independent motors on each machine: enabling enterprises to organize AI capabilities based on business needs rather than model vendors. This isn’t just cost-saving or efficiency-boosting—it’s a transformation of organizational structure.

Recent data from Vercel AI Gateway confirms this logic. In April 2026, based on consumption, Anthropic accounted for 61%, Google 21%, OpenAI 12%; but by token volume, Google led with 38%, Anthropic 26%, OpenAI 13%. Cheaper models handle volume, expensive models handle money—within the same customer base, different winners emerge depending on the dimension. This is the real meaning of “multi-model”: not choosing the best one, but selecting the most suitable for each scenario.

3. Why this position is the most important gateway in the token economy

When no single model can cover all scenarios, the value of the middle layer appears. This isn’t just analyst speculation.

3.1 Collective investor judgment

(OpenRouter raises $113M Series B)

The investor lineup isn’t just “a group of VCs backing a company”—it’s a collective stance across the AI industry’s upstream, midstream, and downstream.

CapitalG (Google/Alphabet) led. Google has Gemini, but it’s betting on a neutral model distribution platform. That’s not contradictory—Google can’t win all customers with Gemini alone, supporting a neutral distribution platform ensures Google’s models are also distributed there.

CapitalG partner Jane Alexander put it plainly: “OpenRouter has a unique positioning—it can become the data clearinghouse and unified intelligence layer for AI models.”

But if you only interpret this as “Google wants to distribute Gemini,” you underestimate this move.

Looking at Google’s three actions in May: CapitalG led the OpenRouter (routing layer—whose brain is the agent), the AP2 protocol was donated to the FIDO Alliance (protocol layer—what wallet agents use for authorization), and the Universal Cart was launched at I/O (entry layer—where agents complete purchases within ecosystems). The protocol layer is open; Google doesn’t need to dominate it. It needs to be the default at the routing and entry points.

When the three key steps—model selection, authorization, and purchase—are influenced by Google, control isn’t just over the protocol but over the flow of traffic through these three gateways.

NVentures (NVIDIA) also followed. NVIDIA’s interest is straightforward—more token consumption means more inference compute used, which means more GPUs bought. OpenRouter amplifies inference volume. This is similar to NVIDIA’s investment in CoreWeave: it doesn’t care which model wins, only that the overall market grows.

Strategic investments from ServiceNow, MongoDB, Snowflake, Databricks. These are enterprise infrastructure companies. Their participation indicates that multi-model orchestration is shifting from a “technology choice” to an “infrastructure standard” for enterprise clients.

a16z, Menlo Ventures, Sequoia all increased stakes. When Menlo invested a year ago, OpenRouter had 2.5 million developers and processed about 100 trillion tokens annually. Now, it has 8 million developers and handles roughly 1.5 quadrillion tokens annually. Menlo’s Deedy Das made a striking comparison: OpenRouter’s token throughput now accounts for roughly 15–30% of Google’s, 20–40% of OpenAI’s, over 50% of Azure Foundry’s. He also revealed that from signing in February to announcing in May, revenue doubled.

When model providers (Google), chip makers (NVIDIA), enterprise software firms (ServiceNow, Databricks, Snowflake, MongoDB), and top VCs (a16z, Sequoia, Menlo) are all on one cap table, it’s not just optimism about a single company. It’s the industry voting with real money.

3.2 Industry data validation

TechCrunch’s assessment: “OpenRouter’s success means AI models are becoming invisible, interchangeable engines. Enterprises no longer want to be locked into a single vendor like in the SaaS era. Multi-model is the future—and it’s already here.”

The rise of Chinese models further solidifies this view. During the 2026 Spring Festival, Chinese models accounted for 61% of token consumption on OpenRouter. When US and Chinese models are available on the same platform for global developers to choose on demand, multi-model isn’t just a trend—it’s reality.

But the key isn’t “multi-model” itself—it's who controls the entry point between models and users.

OpenRouter’s leaderboard is now one of the most widely referenced metrics for model adoption in the global AI industry—investors, researchers, media all watch this list for signals. When your data influences industry decisions, you’re not just a router—you’re an information hub for the entire ecosystem. This isn’t an auxiliary feature; it’s a structural power: every routing is data collection, every traffic record a real-time map of global AI demand.

In the electrification era, ultimate power belonged to the power grid, not the power plant. In the internet age, it’s not websites but search engines and app stores. In the token economy, the lab that produces models is the power plant; enterprises and agents using models are the terminals; and the routing layer—deciding “which request goes to which model”—is the power grid.

Whoever controls this gateway holds the map of the token economy.

4. The future of the token economy: what e-commerce can teach us

Headless merchants have already appeared. But no one has answered where they should appear.

a16z’s Noah Levine defined this concept in March: no storefront, no account system, no sales team—only a server, a set of API endpoints, and a pay-per-use pricing model.

In the first week of Stripe’s Machine Payments Protocol launch, over 60 headless services went live, 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 subscriptions, agents are replacing consumers. The concept is set, the protocol is in place, and transactions are happening.

But where do these headless merchants appear?

Sixty services can be listed in one directory. Six thousand? Six hundred thousand? Agents could theoretically crawl the entire web for API endpoints, but in practice, they need a trusted, ranked marketplace with reliable data and unified billing to discover and compare options. This place is called a marketplace.

Today, OpenRouter is an early form of this marketplace—though its shelves currently only feature one type of headless merchant: LLM providers. If we look at how OpenSea expanded from a single category (profile pictures) to all digital assets, the same path applies: headless merchants will eventually extend from LLMs to image generation, data retrieval, document processing, payment verification—and beyond. It’s a forecast, but not an untested one.

Using e-commerce development as a model, the evolution of a token marketplace likely passes through three stages:

Stage 1: Establish standards. Just as early Taobao defined product categories, search rankings, credit ratings, and unified payments, OpenRouter is defining “what an LLM API should look like”—naming conventions, pricing formats, usability metrics, context parameters. These aren’t designed by a committee; they emerge from massive transaction practice. Standards can’t be cold-started.

Stage 2: Enrich the ecosystem. This is the thinnest layer of OpenRouter today. Like e-commerce: Taobao’s moat isn’t just search, but store decoration, promoted listings, reviews, and credit systems. OpenRouter still lacks recommendation engines (“what you might need”), scenario-based evaluation (not just total call volume, but legal, coding, translation-specific scores), financial tools (token pre-purchase, budget management, credit), and third-party ecosystems (plugins, fine-tuning hosting, prompt marketplaces). These together form the real moat.

Stage 3: A2A (agent-to-agent) emerges from the marketplace. When standards are mature and data is trusted enough, leading clients will “graduate”—directly connecting with headless merchants, bypassing the marketplace. Like big brands moving from Taobao to Shopify stores, or enterprises building their own lightweight LLM routing. But the marketplace won’t disappear—it will become a discovery and trust layer, serving small clients and long-tail services.

Historically, this sequence repeats: centralized exchanges (NYSE) set clearing standards before electronic and dark pools; SWIFT and Visa built payment networks before P2P and on-chain settlement; Amazon and Taobao established e-commerce standards before Shopify enabled independent merchants.

Centralization first establishes standards and trust; decentralization inherits those standards and trust.

This is the real gamble OpenRouter faces. A 5% take rate is a significant cost for large clients—monthly inference costs exceeding about $37,000 make it cheaper for enterprises to build their own with open-source LiteLLM. Vercel’s AI Gateway already offers zero token markup, and Cloudflare has integrated AI Gateway as a free built-in feature.

Trends.vc’s assessment is cautious: “Once a well-funded player commits to passthrough pricing, competitors will follow. Profits will shift from routing itself to caching, governance, and deep integrations.”

The ultimate lesson from e-commerce: Taobao succeeded not because its search was the best, but because buyers, reviews, and trust were all there.

Whether OpenRouter can reach that stage depends on whether it can deepen its ecosystem before the take rate is squeezed—whether the evolution from marketplace to operating system can be completed. The $1.3 billion valuation isn’t just for today’s 5% commission—it’s about whether this transition can happen.

5. Conclusion

Whether OpenRouter becomes the ultimate winner remains an open question. But this funding round marks an irreversible milestone—standards are being defined by transaction volume, trust is accumulating on platforms. The question is no longer “should we build token infrastructure,” but “who will build the infrastructure that will be inherited?”

As for whether this is the Taobao moment of the token economy—the answer depends on who can deepen the ecosystem first. The time window is now open.**