10 million weekly active users. That’s the number doing the rounds for OpenAI’s Codex and ChatGPT Work agents. A blockchain news outlet reports the milestone triggered a reset of usage caps—a promise made when the user base hit 3 million. The surface story is about victory laps. But peel back the wrapper, and you find something else: a case study in inference economics, exponential demand, and the fragility of centralized agent platforms.
Context: The Agent Overlay
Codex is OpenAI’s programming agent. ChatGPT Work is the office automation counterpart—drafting emails, summarizing meetings, editing documents. Both are agentized wrappers around the underlying GPT-4o model. The key policy: for every additional 1 million weekly active users, OpenAI resets the usage limits for all subscribers. It’s a textbook growth hack—rewarding users for recruiting others and for staying active. The result: a 5x jump from 2 million to 10 million WAUs in a single quarter.
But here’s where the math gets interesting.
Core: The Inference Burden
Let’s run the numbers. Assume each active user generates, conservatively, 10,000 output tokens per week. That’s one medium-sized code generation task per day or a few email drafts. At 10 million WAUs, that’s 100 billion tokens per week. To process that many tokens, you need about 20,000 H100 GPUs running at full capacity 24/7—assuming optimized inference at roughly 5,000 tokens per GPU-second. That’s a capital expenditure of nearly $600 million just for the hardware, never mind power, cooling, and staff.
Based on my experience auditing Layer2 sequencer economics, I’ve seen similar growth curves before. The difference is that sequencers can batch transactions. Agent inference is less compressible—each user’s context window is unique. The cost per token doesn’t drop linearly with scale. It drops only with architectural improvements like speculative decoding or model distillation.
OpenAI’s inference infrastructure is clearly world-class. But 10 million WAUs means they are spending millions per week on compute. The resetting of usage limits is a clever move: it encourages users to stay engaged, but it also defers immediate revenue. Instead of charging per token or per task, they reset limits—essentially giving away more free compute to retain users. This is a classic “grow at all costs” play. It worked for Uber. It worked for WeWork. For AI, it might work, but the unit economics are still opaque.
The Hidden Metrics
The article from the blockchain outlet didn’t mention the conversion rate. How many of those 10 million WAUs are paying? If it’s 5%, that’s 500,000 paying users. At $20/month average, that’s $10 million monthly revenue—not bad, but barely covers the inference cost. If it’s 20% paying, the picture improves, but still, the margin is thin when compute is the primary cost.
More importantly, the resetting of limits is a double-edged sword. It signals that user growth is outpacing the ability to monetize efficiently. Why else would you give away more free usage? It’s a sign that the product’s value proposition is strong enough to attract users, but the business model hasn’t yet stabilized.
Contrarian: The Security Blind Spot
Now let’s talk about what no one in the hype cycle wants to discuss: agent security. 10 million agents executing code, editing documents, and interacting with external APIs. Each one is a potential attack surface.
I’ve spent years dissecting smart contract vulnerabilities. The same patterns appear here: privilege escalation, prompt injection, unsafe deserialization. In DeFi, one faulty oracle can drain a liquidity pool. In agent land, one malicious instruction can exfiltrate an entire codebase. Openai’s safety measures are robust, but the attack surface grows with every new user. The resetting of limits doesn’t reset the security posture.
Consider the 2017 ICO boom. Projects used TVL milestones to attract liquidity. The result: inflated numbers, rug pulls, and a crash. Impermanent loss is real. Do your math. The same principle applies here. The milestone mechanism is a psychological tool, not a technical verification of product quality.
2017 vibes. Proceed with skepticism. The agent ecosystem is booming, but the underlying infrastructure—inference, security, cost models—is still maturing. The 10 million number is a testament to product-market fit. But it’s also a warning: rapid growth amplifies systemic risk.
Takeaway
Entropy wins. Always check the fees. The hidden fees here are not monetary—they are the opportunity cost of dependency, the risk of centralized failure, and the unsustainability of growth funded by compute subsidies. The next milestone may not be user count, but a cost crisis. Ask yourself: when the usage limits stop resetting, what happens to the 10 million users who have built their workflows around a single API?
The architecture of trust matters. And right now, the architecture is proprietary, opaque, and fragile. Proceed accordingly.