The 10M Weekly Active Users Signal You're Probably Ignoring
A blockchain news site reports that OpenAI's Codex (coding agent) and ChatGPT Work (office assistant) have reached 10 million weekly active users (WAUs), hitting their final milestone of unlocking usage limits. The story is buried in a crypto feed, and most will dismiss it as just another piece of OpenAI PR fluff. But I see something else.
Every 100 million new users, they unlocked another cap. From 3 million to 10 million WAUs, in a quarter. That's a 233% growth in active agent users in 90 days. The crypto market is obsessed with memecoins and L2 TVL; it's completely missing the biggest narrative shift happening in real-time.
Decoding the Milestone Mechanics
First, let's understand what this 10M WAUs actually means on a technical and business level.
OpenAI didn't just announce a number. They tied it to a product strategy: unlock usage limits as you grow. This is a sophisticated game of behavioral economics. It turns a limitation into a reward. Each million users is a "boss fight" that, once beaten, unlocks more power for the existing players. This creates a self-reinforcing flywheel of growth and retention.
The Real Story is the Scale. 10 million weekly active users on a work-specific AI agent. This isn't a chatbot you use for fun. These are users who are actively delegating real work—writing code for Codex, composing documents and analyzing data for ChatGPT Work. This is the first time we have a concrete, massive-scale data point proving that AI agents are not just a demo, they are a production-grade tool that paying customers rely on.
The Core: Why This Matters More Than Any L2 TVL
Here's the KYC check. As a sector analyst, I've audited dozens of protocols. The question everyone asks is: "Where will the real revenue and activity flow in the next cycle?"
The answer is right here. A single centralized entity is generating more valuable, high-frequency data volume than probably all major L2s combined. This isn't just traffic; it's economic throughput. Every interaction on ChatGPT Work is a unit of labor that's been automated. Every line of Codex-generated code is a unit of value created.
*The contrarian angle? The market is hyper-focused on decentralized compute for AI training (Render, Akash, io.net), but the real monetizable, sustained demand is for inference—the moment of delivery. This event proves that the inference layer is already running at massive, centralized scale. The question is whether this demand ever spills onto decentralized hardware.*
But here's what's even more interesting for us: the cost of serving these 10 million agents. Each user probably consumes 1,000+ tokens per week. That's 10 billion tokens a week in inference. The raw compute cost is staggering. It's a clear signal that anyone who can provide high-quality, low-cost inference compute will have a massive market. The current winners are hyperscalers (Azure for OpenAI). The question is whether DePIN can compete on reliability and latency.
The Contrarian: The Decentralized Narrative is Wrong
Most crypto-native analyses would look at this and say: "Centralized agents win, DePIN is dead." That's recency bias at its worst. Let me give you the Counter-Intuitive discovery.
This 10M milestone validates the DePIN thesis, not invalidates it.
First, Demand is Real and Massive. No more theoretical discussions about "what if people use AI agents." Here's the proof: they are using them, at scale. This creates a massive total addressable market for the infrastructure to support them.
Second, Centralized Bottlenecks are Exposed. Running a service for 10M WAUs on a single provider is a single point of failure. Imagine a GPU shortage, a geopolitical event affecting Azure, or an internal OpenAI compute prioritization that slows down third-party access. The more dependent the world becomes on these agents, the more the demand for decentralized, resilient alternatives grows.
Third, The Data Flywheel is Real... for Now. OpenAI is building an insurmountable data moat. But if you're a crypto founder building an alternative, don't try to build a better chatbot. Build a better inference market. The value is not in the application layer for most; it's in the verifiable, permissionless compute layer that can serve these agents when the centralized pipes clog.
Takeaway: The Next Narrative
The narrative isn't "AI Agent." The narrative is "The Infrastructure for AI Work." These 10M users are not just a PR win for OpenAI. They are a stress test for the internet's compute backbone. And they are a 100-million-volt signal to every DePIN protocol that if you can solve for low-cost, high-reliability inference for agents, you will own the next cycle.
So the real question is not whether these agents are winning. They are. The question is: who will win the fight to power them? The hyperscalers, or a new, decentralized substrate?
As a Dublin-based analyst who bet early on ETH and DeFi, I see the same pattern. First comes the centralized experiment (ICO mania → Bitcoin then DeFi Summer on a centralized server). Then comes the demand for trustless, censorship-resistant rails. This OpenAI milestone is the first chapter of that story for Agents.
The story doesn't end with a single centralized product. It begins with the realization that to keep the world's work running smoothly, you need a more robust, more liquid, and more decentralized compute market. The players who understand this—and are building for inference, not training—will be the next Uniswaps.