10 million weekly active users on Codex and ChatGPT Work. That is not a user base. That is a liquidity event. But the market is staring at the wrong ledger.
OpenAI just confirmed a quiet milestone: its agentic products – a coding agent and a work agent – hit 10M weekly active users, driven by a deliberate growth strategy of resetting usage caps every 100K users. The narrative is that users love the product. The underlying reality is that agentic workflow validation has just injected the largest demand shock for compute since the ChatGPT launch.
As a crypto investment bank analyst who spent 2017 auditing ICO contracts for reentrancy flaws, I learned one thing: leverage hides where liquidity flows. Right now, capital is flowing into centralized AI inference, but the token market is still pricing compute tokens as speculative beta. The protocol isn't the product; the liquidity is. And the liquidity is shifting from chat inference to agent inference – a 10x to 100x increase in token throughput per user session.
Let me break down the macro mechanics.
The Agent Compute Multiplier
ChatGPT chat averages roughly 500 tokens per query. A coding agent like Codex – writing, debugging, refactoring – consumes easily 5,000-10,000 tokens per task. An office agent like ChatGPT Work, summarising documents, drafting emails, and scheduling meetings, lands somewhere in between. Conservatively, an agent user consumes 10x the compute of a chat user.
10M weekly active agent users × 10x token demand = 100M weekly active chat-equivalent users. That is a 40% jump over OpenAI's last reported chat MAU base, and it is all incremental. The market has not priced in the structural demand shift because the token count is not visible on a balance sheet. But on-chain, the compute token supply is fixed. The marginal cost of inference is dropping, but the marginal demand is surging.
Based on my 2020 DeFi liquidity trap analysis – where I modelled the divergence between APY and real value accrual in Yearn vaults – I see the same pattern here. Token prices for compute platforms like Render (RNDR) and Akash (AKT) are still anchored to speculative narratives, not to the actual compute utilisation rate. The real yield is coming from the GPU cycles being burned by these agents.
The Decoupling Thesis No One Believes Yet
The contrarian view is obvious: if centralized AI works this well, why do we need decentralised compute? OpenAI has solved the productisation problem. It is scaling reliably. The contrarian angle is that the very success of centralised agents creates a systemic fragility that only crypto can hedge.
10M weekly users uploading private code and proprietary business data to a single company's servers is a concentration risk that regulators will eventually force a solution for. The EU AI Act already imposes strict data governance requirements for high-risk systems. The real decoupling thesis is not that decentralised compute will beat centralised on price or speed – it won't, not yet. The thesis is that decentralised compute becomes the insurance policy for agentic infrastructure. When the next major breach or outage hits, the demand for verifiable, uncensorable compute will spike. Crypto tokens are best positioned to capture that marginal dollar.
Cycle Positioning
This is a long-duration bet. The bull market is still in its early innings for AI compute tokens. The 10M user milestone validates the demand, but the price discovery happens when institutional capital rotates from centralised cloud providers to decentralised alternatives – likely after a black swan event. The smart money is already positioning RWAs (real world assets) to back compute. I am watching the tokenomics of projects that tie emissions to actual GPU utilisation rather than vanity metrics.
Narratives are just leveraged beta. The real alpha comes from understanding that agent inference is a new asset class. The token that captures the marginal dollar of agent compute will outperform everything else this cycle.
Leverage doesn't care about your conviction. But the data does. 10M weekly active users of agents is not a milestone. It is a warning. Position accordingly.