Hook: The Number That Doesn’t Add Up... Yet
One hundred million dollars in funding, a fresh round of hype, and a single data point: OpenAI’s Codex and ChatGPT Work have allegedly hit 10 million weekly active users. The source is a blockchain news aggregator, citing something called "Dongcha Beating." Zero technical details. Zero context. Just a headline and a number. For someone who spends their days verifying zero-knowledge proofs and auditing smart contracts, this immediately raises red flags. Not because the number is impossible, but because the claim is presented without any cryptographic or empirical backing. It’s like seeing a DeFi protocol boast a Total Value Locked (TVL) of $1 billion without a single audited smart contract to verify it. The number is the hook, but the real story is far more technical and far less certain. Zero knowledge isn't magic; it's math you can verify. And this data point, right now, is a claim without a proof.
Context: The Agent Shift
To understand the weight of this claim, we need to understand the landscape. For the past year, the AI narrative has shifted from “how smart is the model?” to “how useful is the agent?”. OpenAI’s Codex (a programming agent) and ChatGPT Work (an office agent) represent the front line of this shift. These aren't just chat interfaces anymore. They are designed to act: to write code, to edit documents, to send emails, to interact with APIs. This is a productization of AI intelligence, and it requires a fundamentally different infrastructure than a simple chatbot.
A standard ChatGPT session might generate a few hundred tokens per request. An agent session, on the other hand, might generate thousands of tokens, call multiple external tools, and maintain a complex state over several minutes. This is a computationally intensive, high-latency, and high-cost operation. The claim of 10 million weekly active users on such a product is not just a marketing metric; it is a massive engineering statement. It implies that OpenAI has solved a set of non-trivial infrastructure problems: scaling inference for complex, multi-step agent workflows, managing state at an enormous scale, and doing so without bankrupting their compute budget.
From my experience auditing the Gnosis Safe multisig wallet in 2018, I learned that a project's underlying engineering reality is often hidden behind a narrative. The code tells the truth, not the press release. Here, we have no code. So I must rely on first principles: the math of the infrastructure.
Core: The Engineering Math of 10M Agents
Let’s build a model. Assume a conservative average of 500 tokens per agent interaction (a realistic minimum for a code generation or document editing task). A weekly active user might have, say, 10 interactions per week. This is a very conservative estimate for a professional tool. That gives us 100 million unique agent interactions per week. At 500 tokens per interaction, that’s 50 billion tokens processed weekly. That is the raw compute demand.
Divide that over seven days, you get ~7.14 billion tokens per day. The most efficient modern inference hardware, using quantization and speculative decoding, might achieve around 50 tokens per second per H100 GPU for a model like GPT-4 class. To process 7.14 billion tokens per day, you need at least 1,650 H100 GPUs running continuously. But that’s a theoretical maximum. Real-world utilization, including overhead, batch processing inefficiencies, and the need for redundancy, pushes that number significantly higher. A more realistic estimate is 5,000 to 10,000 H100s just for the inference tier for these agent products.
This number aligns with my own experience with zero-knowledge proof generation, which is also computationally heavy. The AMM model hides its truth in the invariant; the cost of scaling a DeFi protocol is hidden in the gas calculations. Similarly, the cost of scaling an AI agent is hidden in the GPU count. The 10 million weekly active users claim, if true, implies a massive and continuous hardware commitment from OpenAI and Microsoft. It suggests that their inference pipeline is not just optimized, but has reached an industrial maturity that few companies can replicate.
Here’s the technical check I performed mentally: In my 2020 analysis of Uniswap V2, I built a Python simulation to test their slippage model. The code revealed that their invariant was mathematically sound but economically fragile under specific liquidity conditions. I don't have a Python simulation for OpenAI’s infrastructure, but I can reverse-engineer the economic fragility of the claim. If OpenAI has achieved this scale, it implies a unit cost per token that is aggressive. This could be via extreme model compression, efficient attention mechanisms (like MQA/GQA), or a custom ASIC. Any of these would be a proprietary advantage, reinforcing their moat.
Contrarian Angle: The Security Blind Spot of Scale
Here is where my skepticism as a security analyst kicks in. Scaling an agent product to 10 million users introduces a class of security vulnerabilities that are fundamentally different from scaling a chat bot. When I reverse-engineered the Axie Infinity smart contracts in 2021, I found a bug in the breeding fee calculation that allowed infinite token generation. It was a logic error, not a syntax error. The error was in the economic model, not the code itself.
For AI agents, the equivalent is the prompt injection or an agent-level logic error. An agent that can read an email, edit a document, and send a message has a massive attack surface. A single prompt injection could instruct the agent to act maliciously on the user’s behalf. This isn’t a theoretical risk; it’s a cryptographic certainty. The more complex the agent’s workflow, the more opportunity for an attacker to embed a malicious sub-instruction. This is the security blind spot of the “agent era”: we are scaling a system where trust is assumed, but the underlying verification is difficult.
Furthermore, the data from these actions (code repositories, company strategies) is a goldmine for attackers. A supply chain attack on an agent platform could be catastrophic. The fact that OpenAI has achieved this scale without a major public security incident is either a testament to their security engineering (which is likely very good) or a sign that the exploit surface hasn’t been fully tested. I don't trust the hype; I trust the invariant. And the security invariant of an agent is only as strong as its weakest logical link.
Takeaway: A Forward-Looking Judgment
The 10 million weekly active users is a data point, not a conclusion. It is a signal that the AI industry has reached a new phase of product maturity. For investors, the math implies a massive demand for compute hardware, validating NVIDIA’s thesis. For competitors, it creates a moat built on data and infrastructure. For users, it means the agent era has arrived, but that arrival comes with a security tax. My advice: treat this data point with the same critical rigor as a new DeFi protocol’s TVL. Ask for the proof, not just the announcement. Based on my audit experience, I will not trade the narrative; I will wait for the confirmable state. The code, or in this case, the detailed technical architecture and security audits, will tell me if the 10 million is a real victory or a cleverly designed bug.