10 million weekly active users. That is the headline. Two OpenAI products – Codex, the coding agent, and ChatGPT Work, the office agent – allegedly hit that milestone. The number comes from a single source: a blockchain news site citing an entity called "Dongcha Beating." No official confirmation. No GitHub commit hash. Just a claim wrapped in a growth narrative.
Let's be clear: I don't trust the number. Not because I doubt OpenAI's reach, but because I've seen this play before. In 2020, people claimed transaction volume on new DeFi protocols as proof of adoption. I spent 72 hours reverse-engineering the Uniswap V2 launch code before I executed a single trade. The ledger showed liquidity, but the truth was in the order flow. Here, the only ledger is a press release dressed as news.
Code does not lie, but liquidity does. The liquidity here is attention. And attention can be manufactured as easily as a synthetic stablecoin.
But even as a hypothetical, 10M weekly active users is a signal worth dissecting. If true, it redefines the AI landscape. If false, it reveals a market desperate for a narrative. Either way, the structural implications are real.
Context: The Milestone Mechanism
OpenAI, according to the report, set a public milestone: for every 100,000 new monthly active users, they would reset usage limits on Codex and ChatGPT Work. The starting point was 3M monthly actives. The goal was 10M weekly actives. They claim to have hit it.
This is not a standard growth hack. It is a commitment device. By tying a visible metric (usage limits) to user growth, OpenAI created a feedback loop: more users → more unlocked limits → better product → more users. It is a textbook flywheel, but executed with the leverage of a centralized platform.
The moon is a myth; the ledger is the only truth. The ledger here is the reported number. Without transparent on-chain attestation or audited data, we are left with trust. And trust is a weak foundation for any investment thesis.
From my engineering standpoint, the mechanism itself is sound. It incentivizes existing users to recruit others. The viral coefficient could be high. But the reported 10M figure is a point estimate. Anyone who has run A/B tests knows: one data point is noise.
Core: What 10M Agents Mean for the Stack
Assume the number is real. What does 10M weekly active users imply?
Inference cost is the new gas fee. Each user generating 1,000 tokens per session (conservative for coding or office tasks) means 10 trillion tokens per week. That requires hundreds of thousands of H100 GPUs running at near-full utilization. The inference cost alone could be in the millions per day. OpenAI's margins depend on optimization tricks I've used myself: speculative decoding, kv-cache sharding, and adaptive batching.
But here's the hidden insight: Agent usage is not linear. A coding agent like Codex might generate 10,000 tokens per session. A single user debugging a complex repo could consume more compute than a hundred casual users. The variance is extreme. The reported 10M figure likely hides a power-law distribution. The top 1% of users may account for 30% of total compute.
Trust the math, ignore the memes. The math says this: to support 10M active agents, you need a redundant, multi-region inference fleet. The cost structure resembles a cloud provider more than a software company. OpenAI is essentially running a private cloud for AI agents.
From my own experience building a copy-trading bot for Bitcoin ETFs in 2024, I learned that latency is everything. I coded a Rust-based execution engine to capture 0.5% spreads. That engine handled 10,000 transactions per day. Scaling to 10M agents is a different order of magnitude. It requires automating failure detection, load balancing across GPU clusters, and dynamic resource allocation. OpenAI must have a world-class infrastructure team.
But infrastructure is not a moat. Any well-funded competitor can buy GPUs. The real moat is data. Every agent interaction is a training example. Codex's outputs become input for the next model iteration. ChatGPT Work's office tasks generate a massive corpus of human-computer interaction data. This is the flywheel that can't be replicated quickly.
Survival is the first profit metric. The first question: can OpenAI sustain the cost? If the 10M figure is true, they might be burning cash at an alarming rate. The milestone mechanism (reset usage limits) encourages more usage, which increases cost. It's a gamble that the data value will outpace the infrastructure spend.
Contrarian: The Blind Spots
Everyone is celebrating the user number. But I see three blind spots.
First, the source is unreliable. "Dongcha Beating" appears to be a translation of a Chinese media outlet. The article's agenda is unclear. Is it a leak from OpenAI's marketing team? A speculative report from a journalist? The lack of verifiable details (specific date, user breakdown by region, product split) makes this indistinguishable from a bull case narrative created by a trader.
Second, user growth does not equal value extraction. A million users using a free tier or cheap subscription is not the same as a million paying customers. OpenAI's real revenue engine is the API and enterprise contracts. Codex and ChatGPT Work may be bleeding cash as loss leaders to build habit. The 10M figure could be a vanity metric.
Third, the data asymmetry favors OpenAI. They know exactly how many users, their retention rates, and their compute costs. External investors and developers are flying blind. This is a classic principal-agent problem. I saw it during the Terra/Luna collapse: the team knew the reserve mechanism was breaking, but they publicly projected confidence. The asymmetry killed thousands of portfolios.
Chaos is just data you haven't parsed yet. The chaos here is the uncertainty around the 10M number. Until OpenAI provides auditable proof – like on-chain verifiable usage logs – I treat it as a hypothesis, not a fact.
Takeaway: The Only Question That Matters
If the number is real, the market for AI agents is as hot as the crypto market in 2021. If it is false, the hype curve will correct sharply. Either way, the smart money is not on the headline. It's on the infrastructure that supports any agent ecosystem: compute, networking, and data pipelines.
My community – the "Verified Hands" – requires members to show trading logs before they join. I apply the same standard here. Show me the GitHub commit that logs 10M unique agent sessions. Show me the transaction hash that proves the usage limit reset. Until then, I remain a cynic with a spreadsheet.
Speed kills, but patience compounds. Wait for the data. Then act.
