10 million weekly active users. That is the headline from OpenAI’s latest product milestone: Codex and ChatGPT Work, their dedicated AI programming and office agents, have crossed that threshold. The company celebrated by resetting usage limits—a clever growth hack that turned user gains into product unlocks.
But the math doesn’t lie. Behind every token generated by these agents lies a hidden cost: massive, centralized compute clusters owned by a single entity. For those of us who audit decentralized protocols for a living, this number is not just a success story. It is a stress test for blockchain’s claim to host the future of AI.
The Context: What These Agents Actually Do
Codex is an AI programming agent—it writes, debugs, and refactors code within an IDE environment. ChatGPT Work is a general-purpose office agent that reads emails, edits documents, and manages calendars. Both are powered by OpenAI’s latest models, likely GPT-4o or a derivative. The 10M weekly active user figure represents a 5x increase over a single quarter.
OpenAI’s strategy was simple: tie product improvements to user growth. They promised to reset usage limits every time the user base crossed a new milestone (3M, 5M, 7M, 10M). This created a viral loop—existing users recruited new ones to unlock higher caps. The result? A textbook example of product-market fit.
But here is where most blockchain commentators stop. They applaud the growth and move on. They ignore the infrastructure under the hood.
Core Analysis: The Compute Burden and the Decentralization Gap
10 million agents producing 1 trillion tokens per week. That is my conservative estimate based on average coding and office task loads. Each token requires inference on a GPU cluster. At current efficiency levels (optimistic), this means 50,000 to 100,000 H100 GPUs running at near-full capacity—a fleet that costs upward of $2 billion to build and millions per day to operate.
OpenAI can afford that. Microsoft provides the Azure credit and the hardware. But what about the 99% of projects that claim to build “decentralized AI” on blockchain? None of them can serve even 1% of that load. Not Akash. Not Render. Not Bittensor. Not even the most optimistic estimates of decentralized compute capacity come close.
Based on my audit experience in 2022, I evaluated a decentralized GPU marketplace that claimed to offer competitive inference pricing. The protocol had a critical flaw: the proof-of-use mechanism could be gamed, and the actual latency for real-time tasks exceeded 30 seconds. The math didn’t lie—decentralized compute networks are orders of magnitude slower and more expensive than centralized clouds for high-throughput AI inference.
Security is not a feature; it is the foundation. Right now, the foundation of agent infrastructure is wholly centralized.
Contrarian Angle: The Real Bottleneck Is Not Compute, It Is Trust
Most blockchain advocates argue that decentralized compute is the solution. I disagree. The bottleneck for AI agents is not the raw number of GPUs—it is the trust model. Users are uploading proprietary code and sensitive business data to OpenAI’s servers. Even if decentralized compute were fast enough, no blockchain today provides a practical, privacy-preserving execution environment for real-time agent tasks.
Fully homomorphic encryption (FHE) is too slow. Zero-knowledge proofs (ZK) can verify model outputs but not the process itself. Trusted execution environments (TEEs) introduce hardware-level centralization risks. The result: the “AI on blockchain” narrative is a three-year-old story that no one wants to admit has a fundamental scalability wall.
But here is the contrarian opportunity: agent behavior verification. While you cannot execute the full agent workload on-chain, you can record and verify critical actions using ZK proofs. For example, an agent that manages a DeFi portfolio could submit a proof that its trades follow a given strategy without revealing the strategy. This is where blockchain adds true value—not in running the agent, but in auditing it.
Complexity hides the truth; simplicity reveals it. The simplest truth is that the AI agent boom exposes the centralization of AI infrastructure. Blockchain’s role should be to provide transparency and verifiability for those agents, not to replace the underlying compute.
Takeaway: The 10M Signal for Blockchain Builders
OpenAI just proved that AI agents can achieve mainstream adoption. For blockchain, this is both a warning and a roadmap. A bug fixed today saves a fortune tomorrow—but only if you are looking at the right problem.
The warning: if distributed ledger technology cannot provide the compute or trust infrastructure for the next billion agent users, the entire narrative of decentralized AI will collapse into a niche experiment. The roadmap: focus on attestation, not execution. Build protocols that allow agents to prove their actions without revealing private data. Use cryptographic primitives like ZK-SNARKs or MPC to give users auditability over black-box models.
Trust the code, verify the trust. The 10 million users are using agents that are opaque, centralized, and governed by a single company’s terms. The blockchain community has a narrow window to build the tools that make those agents accountable. Otherwise, the future of AI is just a more efficient version of Web2—with a few tokenized GPUs on the side.