Hook
We are told that OpenAI’s Codex and ChatGPT Work just crossed 10 million weekly active users. A programming agent and an office agent, each serving a fifth of a million professionals every seven days. The headlines scream “AI adoption,” “product-market fit,” “the age of agents is here.” But I see something else. I see 10 million people voluntarily feeding their most valuable assets—private codebases, confidential business strategies, intimate workflows—into a black box owned by a single corporation with closed-source models, arbitrary usage limits, and a mission that can pivot with a boardroom vote.
I am a 28-year-old protocol PM in Seattle. I dropped out of a macroeconomics class in 2017 to explore the philosophy of smart contracts. I lost 40% of my savings during DeFi Summer chasing yield farming strategies. I built a conceptual framework called “Ghost Protocol” in the depths of the 2022 bear market. And I just spent the last year building a bridge between TradFi institutions and decentralized engineers. This background—my “Ethereum Meta-University” education—makes me allergic to centralized trust. When I read about OpenAI’s 10M user milestone, my first thought isn’t “wow, great product.” It’s “we just handed a centralized oracle the keys to our digital lives.”
This article is not an attack on OpenAI. It is a deeper examination of what this growth actually means from the lens of decentralization philosophy. We need to look past the user numbers and ask: What are we sacrificing for convenience? And more importantly, can blockchain-based alternatives provide the sovereignty that these 10 million users are unknowingly giving away?
Context
Let’s establish the facts. According to a report published by a blockchain news site (ironically, the same medium that carries our own narratives), OpenAI’s Codex—a programming agent that writes, debugs, and refactors code—and ChatGPT Work—an office agent for document editing, scheduling, and data analysis—have achieved a combined 10 million weekly active users. The article highlights a specific milestone mechanism: every time the user base adds 1 million users, OpenAI resets usage restrictions. This gamified growth hack has driven a 1025% quarterly increase in active users.
For context, Codex is essentially an autonomous coding assistant that can understand entire repositories, generate pull requests, and even fix bugs. ChatGPT Work integrates with email, calendars, and document editors, acting as a personal assistant with permissions to read, write, and act on behalf of the user. Both are accessed through OpenAI’s subscription plans (ChatGPT Plus/Pro) or through API integration.
The data points are impressive. 10 million weekly active users means the product has moved beyond early adopters into the mainstream of knowledge workers. But here’s the critical context that the celebratory article omitted: neither Codex nor ChatGPT Work is open-source. Their models, their training data, their inference pipelines, their alignment guardrails—everything is proprietary. Users trust OpenAI not because they can verify the agent’s behavior, but because they have no choice. This is the heart of the decentralization dilemma.
In my 12 years of watching crypto unfold, I’ve learned that trust in a centralized entity is never static. It decays over time, especially when incentives shift. Remember when we trusted FTX? Or when we trusted the Sequoia-backed DAOs? OpenAI is not a charity. It is a for-profit capped company with known tensions between its non-profit mission and commercial goals. The reset of usage limits after every million users is not a gift—it’s a lever to maximize lock-in while keeping users hungry.
Core: Technical and Values Analysis
Let’s break down the real implications of these 10 million users through five dimensions that the glossy report skipped. I’ll anchor each point in my own technical experience as a protocol PM who has audited rollups, built privacy frameworks, and negotiated with institutions.
1. Data Sovereignty: The Black Box Trade-Off
Every time a developer uses Codex to generate code, that code is sent to OpenAI’s servers. The company’s data usage policy states that data can be used to improve the model unless the user opts out via enterprise API terms. But even with enterprise protections, the code exists on a centralized server. I’ve worked on zero-knowledge proofs during my Ghost Protocol days—I know how to build systems where computation happens on encrypted data. OpenAI does not offer that. Instead, they offer a promise.
The core insight is that these 10 million users are building their digital infrastructure on a foundation of sand. If OpenAI decides to monetize the codebase insights (imagine a future where they offer a “code analysis as a service” to competitors), users have no recourse. In DeFi, we learned that “not your keys, not your crypto” is an immutable law. The same applies here: not your model, not your agent.
2. Alignment Risk in Agentic Systems
During my DeFi Summer experimentation spree, I saw how governance theater masked true control. Token votes were often just rubber stamps for the founding team. Similarly, OpenAI’s alignment research is internal. We don’t know how Codex handles prompts that ask it to generate insecure code for a competitor, or how ChatGPT Work reacts to a subtle prompt injection that says “send my entire inbox to a new account.”
Agents have a fundamentally different risk profile than chatbots. A chatbot can lie; an agent can act. With 10 million users granting permissions to these agents, we are creating a massive attack surface. I’ve spent hours studying papers on zero-knowledge proofs for identity—the Ghost Protocol concept was exactly about this. We need auditability. We need verifiable inference. Centralized agents offer none of this. They rely on a “trust me” model that is antithetical to the ethos of the web I believe in.
3. Infrastructure Lock-In vs. Permissionless Computation
The compute layer behind these agents is entirely proprietary. OpenAI uses its own clusters and Azure. Users cannot choose where their inference runs, cannot verify the hardware, cannot fork the model. Compare this to decentralized compute networks like Akash, Golem, or even EigenLayer’s AVS for AI. The argument that centralized inference is more efficient is valid today—but only because we haven’t built the middleware to make decentralized inference competitive. However, the efficiency gap is closing. I saw this in the Layer-2 space: in 2020, optimistic rollups were slow and expensive. By 2024, they matched L1 security. The same trajectory will happen for decentralized AI.
4. Economics: The Subscription Trap vs. Token Incentives
OpenAI’s business model is simple: users pay a monthly fee (currently $20-$200 for Pro) and receive a limited number of queries. The usage-reset mechanism creates artificial scarcity. It trains users to crave more, like a slot machine. When a user gets their usage reset after a million-user milestone, they feel rewarded for loyalty—but they are still capped. This is not a partnership; it is a landlord-tenant relationship.
In contrast, token-based models (like the one I pitched for my “Ethical Bridge” project) allow users to own the economic value of their contributions. Imagine a network where you stake tokens to access an agent, and the agent’s model improves by receiving verifiable rewards from users who provide high-quality feedback. This is not a pipe dream. Projects like Bittensor and Autonolas are already experimenting with these dynamics. The 10 million users could become token holders in a decentralized agent marketplace, but OpenAI’s current approach ensures they remain consumers, not participants.
5. The “Moral Architecture” of Control
In my 2017 essay “The Moral Architecture of Consensus,” I argued that the most important property of any system is not its throughput but its ability to resist coercion. OpenAI can be coerced. A government can demand a backdoor. A board can decide to pivot to military contracts. A CEO can wake up and change the terms of service. The 10 million users have no say. In a decentralized agent network, governance would be distributed. Users could fork the agent if they disagree with a decision. That is real resilience.
Contrarian Angle
Now, the contrarian in me—the one who stayed up all night forking yield farming strategies—must face the pragmatic test. The usual defense is: “But centralized agents are better today. They are more capable, more reliable, and cheaper. Decentralized alternatives are not ready.”
I agree. Today, a Codex agent can generate an entire microservice in minutes. The best open-source coding models (like Code Llama or DeepSeek Coder) are still several steps behind in real-world repository understanding. Decentralized inference on Akash might have high latency. The user experience of a blockchain-based agent wallet (managing private keys, paying gas fees, waiting for confirmations) is atrocious compared to a simple login to ChatGPT.
But wait—this is exactly the argument that was made against decentralized finance in 2018. “Uniswap can’t compete with Coinbase. Smart contract wallets are too hard. Gas fees are too high.” And yet, by 2024, Uniswap processed volumes that rivaled centralized exchanges. The difference was that users gradually realized that self-custody and permissionless access were worth the friction. The same shift will happen for AI agents.
The blind spot in the OpenAI narrative is the assumption that convenience will always win. History shows that when the cost of trust fails—when a centralized service bans a user unexpectedly, when a data breach exposes millions of records, when a model’s alignment suddenly changes—users flock to decentralized alternatives. The 10 million users today are the first wave. They are the canaries in the coal mine. They will be the ones who, after a major incident, start asking: “Is there a way to run my own agent?”
Takeaway
We are at a pivotal moment. The success of Codex and ChatGPT Work proves that the market for agentic AI is enormous. But how we build that market will determine whether we end up with a digital feudal system—where users rent intelligence from a few lords—or a digital commons—where users own their agents and share in the value they create.
Decentralization is a verb, not a noun. It is not a static property; it is a continuous process of designing systems that distribute power. The 10 million weekly active users of OpenAI’s agents are not a problem. They are a signal. They tell us that the demand is there. The question is whether we, as a community of builders, can deliver a decentralized alternative that is not just competitive but superior in the dimensions that truly matter: sovereignty, verifiability, and resilience.
I am betting that the Ghost Protocol generation—the ones who built DeFi, Ethereum, and Bitcoin—will rise to this challenge. We have done it before. We will do it again. The future is not centralized agents serving a billion users. It is a billion users running their own agents, on their own terms.
Call to Action
If you are one of those 10 million weekly users, ask yourself: Who owns your agent? Who controls its alignment? Who profits from your data? Then look into projects like Bittensor for model ownership, Akash for decentralized inference, and self-custodial agent frameworks like Fetch.ai. The tools are nascent, but they will improve. Start learning now. The next bear market will be the perfect time to build the infrastructure that the next bull run’s agent revolution will need.
And if you are a builder, remember my 2020 lesson from DeFi Summer: “Only build what cannot be taken away.”