The number is beautiful: 9 million active users. The growth rate is even more beautiful: 1 million new users every 33 hours. Sam Altman is busy warning about service interruptions. The team is replenishing quotas for the fourth consecutive day. The system is struggling to breathe, and the market is applauding.
I see a different number: the cost per inference on a large language model optimized for code generation. That number is ugly. And when you multiply it by 9 million, the illusion of free growth collapses. The code compiles, but the reality bankrupts.
The Context: When Demand Exceeds Physics
OpenAI's Codex and ChatGPT Work are the headline products of the current AI gold rush. Codex generates code. ChatGPT Work handles enterprise collaboration. Both are powered by GPT-4 variants, which are expensive to run. The recent surge—from 6 million to 9 million users in three days—is a product-market fit signal that any startup would kill for. But it also signals something else: the infrastructure is hitting a hard ceiling.
Altman's warning about "brief outages" is not a humble CEO's caution. It is a data point. It tells me that the marginal cost of serving the next 100,000 users is exceeding the marginal revenue. The quotas are not a feature; they are a rationing mechanism.
Core Teardown: The Unit Economics That Don't Close
Let me walk through the math. I spent years modeling DeFi liquidity pools and impermanent loss. The same first-principles thinking applies here.
The cost assumption: Each Codex query is not trivial. Code completion requires generating tokens at a low temperature, often with long context windows. Even with optimized inference, a single query on GPT-4 class models costs roughly $0.01 to $0.03 in compute. Codex is likely on a smaller model, but still — let’s be conservative and assume $0.005 per query.
The usage assumption: If 9 million users average 10 queries per day, that is 90 million queries. Daily cost: $450,000. Monthly: $13.5 million. That is only for inference. Does not include research, personnel, or the massive investment in training the next generation of models.
The revenue assumption: We don't know the free-to-paid ratio. But consider: ChatGPT Plus costs $20/month. If even 10% of Codex users are Plus subscribers (900,000 users), that’s $18 million per month. Sounds okay, until you realize that free users are burning compute too. And the heavy users—the developers who run hundreds of queries per day—are likely on the free tier or on Team plans that cost $25/user/month but include higher usage limits. The unit economics are razor thin.
The growth trap: The faster users join, the faster the cost spirals. OpenAI is forced to either raise prices, impose throttling, or lose money on every new user. The current quota system is a band-aid. The real solution requires either a massive reduction in inference cost (through model distillation, quantization, or custom silicon) or a radical pricing model that extracts more value from high-usage users.
Based on my audit experience with DeFi protocols, I recognize the pattern: exponential user growth often hides exponential cost growth. The liquidity mining yields were subsidized. Here, the compute is subsidized. Once the subsidies stop, the real users vanish.
The Contrarian Blind Spot: What the Bulls Got Right
I am not dismissing the achievement. The demand is real. Developers are abandoning other tools for Codex. The product-market fit is undeniable. The bulls are correct that this validates AI's role in software engineering at scale. The network effects—more users generating code, more training data for the model—are genuine.
But the bulls ignore the bottleneck. They see the 9 million and extrapolate to 100 million. They assume that cloud capacity scales infinitely. They assume that inference costs will drop at the same rate as training costs. They assume that OpenAI will capture value through eventual monetization. These are all assumptions, not laws of physics.
The transaction is permanent; the mistake is not. The mistake is believing that user growth alone defines winning. In infrastructure-heavy businesses, the winner is the one who can sustain the cost curve longer than the competitor. Right now, OpenAI is running faster than its legs can carry it.
Takeaway: The Accountability Call
OpenAI has two paths. The first is to aggressively optimize inference—shrink the model, improve hardware utilization, and push for custom chips. The second is to raise prices dramatically, which will decelerate growth but improve unit economics. The third, unspoken path is to let the system break, watch users churn, and rebuild with a more sustainable architecture.
The market will not tolerate a third straight quarter of "growing beyond infrastructure." The VCs who funded the compute will want returns. The customers who rely on Codex for daily work will not accept outages.
I do not trust the audit; I trust the exploit. The exploit here is the assumption that growth can continue without solving the cost equation. When the next model release comes and the inference bill doubles, the illusion will have a price tag. Truth has none.
OpenAI should stop celebrating user numbers and start publishing cost-per-user metrics. Until then, the growth story is a fairy tale with an ugly epilogue.
