Hook
On the morning of September 10, the announcement arrived the way most infrastructure news arrives now โ quietly, in a written post, without a livestream, without a countdown clock. Dee Goens, who co-founded Zora alongside Jacob Horne in 2020, would step into the chief executive's chair. Horne, who had occupied that chair for more than six years, would depart to begin what he called a new chapter, promising only that he would not stray too far from crypto. And buried beneath the two names, in the same paragraph, sat the number that actually mattered: Zora's team, once dozens deep, now numbers fewer than ten people, and the organization is reorienting itself around a more AI-driven operational model.
I read that post three times, and not because the succession startled me. Six years is an epoch in this industry; founders who outlive their own thesis almost never survive the transition, and Horne stayed closer to Zora's original promise than most. What startled me was the arithmetic. Fewer than ten people is not a team โ it is a quorum. It is barely enough humans to run a multisig ceremony without someone being on holiday. And yet Zora is not a corpse. It is a live network with a recognizable culture, a functioning rollup, a treasury, and users who mint on it every single day. The distance between what Zora is and how many people keep it breathing is the most important technical fact of this news cycle, and almost nobody is reading it that way.
Context
To understand why nine people is a story and not a footnote, you have to remember what Zora was supposed to be.
The company launched in 2020 as a marketplace with an unusually philosophical pitch: anything could be minted, and minting should cost almost nothing. Not art, necessarily โ the founders were openly bored by the gatekeeping of the curated drop, the hand-selected artist, the six-figure primary sale. Their thesis was that the interesting object was the onchain token itself, a permanent, programmable, publicly auditable record of a moment, and that if you drove the marginal cost of creating one close enough to zero, culture would find its own way there. The Zorb โ that flat, slightly menacing gradient sphere โ became the visual shorthand. It was never a logo so much as a blank currency of participation, a shape you could pour meaning into.
Then 2021 arrived, and the whole industry mistook a speculative mania for product-market fit. Zora did fine in that period; nearly everyone did. But the mania broke in 2022, and something subtler happened that most people never noticed: mints did not go to zero, they went to the wrong places. Volume concentrated in a handful of blue-chip collections and the occasional airdrop farm, while the long tail โ the actual argument for cheap minting โ thinned to a murmur. I have watched this pattern before, from a different chair. From the chaos of 2017, we forged a compass: the lesson of the ICO era was that a token sale is not a business, and that a protocol which measures itself in issuance rather than usage will always be measuring the wrong thing.
In June 2023, Zora made the move that defines its present situation. It launched Zora Network, an OP Stack rollup, converting itself from a marketplace into a chain โ a place where minting could happen natively, cheaply, and under its own fee schedule, with a protocol reward fixed at 0.000777 ETH per mint flowing back to the network. The technical rationale was sound and, in fact, correct: general-purpose layer-one blockspace is the wrong venue for high-frequency, low-value cultural transactions, and a dedicated rollup could offer sub-cent minting with coherent creator payouts. The strategic rationale was more complicated. Every marketplace that becomes a chain inherits the obligations of a chain โ sequencer uptime, bridge security, upgrade coordination, incident response โ and none of the token liquidity that traditionally pays for them.
Then came Dencun. In March 2024, EIP-4844 shipped blobs to Ethereum, and overnight the cost of posting rollup data collapsed by roughly two orders of magnitude. Every L2 on the stack received a subsidy from the base layer, and for a while fees across the ecosystem drifted toward zero. Zora's unit economics improved without anyone having to do anything at all. It was, in retrospect, the most generous gift the network ever received โ and the most misleading, because it was never priced to last.
That is the context into which Dee Goens steps. A protocol with a beautiful thesis, a real product, a cheap chain, a collapsed mint market, and a headcount that would embarrass a seed-stage startup. And a plan โ the plan everyone now has โ to let machines carry the weight.
Core
The Arithmetic of a Nine-Person Protocol
What does it actually take to run an OP Stack rollup? I have audited enough of this infrastructure to give a rough accounting, and the list is unforgiving: a sequencer operator and an on-call rotation; a verification or at minimum a monitoring path for the challenge mechanism; bridge contracts and their key ceremonies; an upgrade process with a security council or its functional equivalent; a node fleet; a data pipeline for analytics; a front end that does not fall over on a mint day; a contracts team; and a communications function capable of posting the correct thing within the first ten minutes of an incident. That is, generously, twelve to twenty people if you respect on-call rotations and refuse to accept a single point of human failure.
Zora intends to do it with fewer than ten, plus machines. The interesting question is not whether that is brave. The interesting question is which of those functions are genuinely automatable and which are not.
Sequencer operations, node monitoring, alert triage, log correlation, documentation, first-line support, and code review triage are all substantially automatable today โ not perfectly, but well enough that a smaller team can hold the line. Governance communications, incident statements, treasury policy, and anything touching key material are not. They require a person with a name, a face, and something to lose. I want to be precise here, because the industry is currently sloppy about the distinction: an AI can operate a system, but it cannot be accountable for one. Accountability requires the capacity to be blamed, and blame requires a subject. A model is not a subject. It is a very expensive mirror.
So the shape of the modern protocol team is becoming clear, and it is not simply "smaller." It is bifurcated. A thin layer of humans who own keys, judgment, and legal exposure; a wide layer of agents that own throughput. Nine people is not nine engineers doing the work of thirty. It is nine people doing the work that cannot be delegated, plus a fleet doing the work that can.
What "AI-Driven Operations" Means at the Contract Level
Here is where my current research becomes unavoidable, because "AI-driven" is not a vibe โ it is a set of cryptographic claims about who or what authorized an action, and those claims either verify or they do not.
When an organization announces that it is becoming AI-driven, three distinct engineering questions hide inside that sentence. Does the AI produce artifacts โ copy, code, documentation, support responses โ that a human reviews before they reach the public? That is the low-stakes case, and it is where most of the current wave of layoffs is landing. Does the AI participate in decision-making, generating proposals, risk assessments, or parameter recommendations that humans then execute? That is the middle case, and it is where the last two years of tooling have quietly gone. And โ the case nobody raises in a press post โ does the AI hold authority, meaning keys, roles, upgrade rights, or the ability to move funds without a human signature?
The third case is where my team's work on the Human-Centric AI Ledger lives, and I will describe the mechanism honestly, including its weaknesses, because the honesty is the point. The idea is an attestation chain. A model is identified by a hash of its weights and configuration. A prompt and its retrieved context are hashed into a reproducible input root. The inference executes inside a hardware enclave, or โ where the workload permits โ is accompanied by a zero-knowledge proof of correct execution, and the output is signed by an attestation key whose certificate chains back to a vendor root. The signed receipt is then anchored onchain, so that any later dispute about what the machine decided can be resolved against a record no operator can quietly rewrite.
Every layer of that stack has a crack in it. Hardware enclaves depend on trusting Intel, AMD, or Arm, and the history of side-channel research โ Foreshadow, Plundervolt, the long tail of SGX embarrassments โ tells us that trust should be narrow and time-bounded, never existential. Zero-knowledge proofs of inference are mathematically stronger but currently two to four orders of magnitude too expensive for anything at production scale; the proofs cheap enough to run verify models too trivial to matter. And the deterministic-input assumption is quietly false for most deployed systems, where sampling temperature, retrieval state, and tool-call ordering make "the same input" a fiction.
Still, the crack is not the point. The point is that the alternative โ an AI-driven protocol with no attestation at all โ is strictly worse, because it produces a system in which decisions are made, funds move, and nobody can reconstruct why. If Zora, or any chain, intends to run its operations with agents, then the audit trail is not a nice-to-have feature. It is the constitutional document of the organization. Trust is not a metric; it is a memory we share โ and a memory that lives only inside a model's context window is not a memory at all. It is an evaporation.
The Blob Subsidy and the Real Reason for the Pivot
I want to advance one claim in this piece I have not seen made elsewhere, and I will state it plainly: the shift to AI-driven operations is not primarily a philosophy. It is a margin strategy, and it is a hedge against the repricing of blob space.
Here is the mechanism. Post-Dencun, Ethereum gives rollups a separate fee market for data availability. Blobs are targeted at three per block, with a maximum of six, each blob 128 kilobytes โ so the target is roughly 384 KB per block, about 1.9 megabytes per minute, on the order of 2.8 gigabytes per day for the entire rollup ecosystem combined. For most of the last two years, because blob demand sat below target, the blob base fee stayed pinned at its minimum of one wei. Rollups were, in effect, paying nothing to be secured by Ethereum's data layer, and passing that nothing on to their users in the form of sub-cent fees.
That is not a market. That is a subsidy wearing a market's clothes. And a subsidy is only stable while nobody else wants the resource. We have already seen what happens when someone does: the blob fee spikes of spring 2024, when a single inscription-style minting spree briefly pushed the blob base fee up by orders of magnitude and reminded everyone that three blobs per block is an extremely tight budget for the world's entire scaling roadmap. The space saturated within minutes. It saturated because of exactly the kind of low-value, high-volume minting that Zora's own thesis celebrates. That is the tension: the cheap mint is the product, and the cheap mint is the thing that will eventually make the mint expensive.
I have written before that blob data will saturate within two years and rollup gas fees will double again, and I stand by it. The demand side is not merely L2 growth โ it is Base, Arbitrum, Optimism, Scroll, Linea, Zora, dozens of app-chains settling to L2s, and whatever new consumer of blobs emerges next quarter. The supply side is governed by a fee market with a target Ethereum's own researchers have been cautious about raising, precisely because raising it weakens the bandwidth requirements for solo stakers. When the fee market does its job, the cost lands on the rollups. The rollups will pass it to users, cut costs, or both.
A nine-person team with an agentic operations layer is exactly what "cut costs" looks like when you cannot cut the chain. The AI pivot is legible as ideology, and it is being sold as ideology. Underneath, it is a response to a cost curve that a small protocol cannot control and can no longer ignore.
The Fragmentation Story, and Why I Do Not Buy It
There is a version of Zora's situation the market has already written, and it goes like this: Zora is one among dozens of rollups, gas is cheap everywhere, users are spread thin, liquidity and attention and creators are fragmented across chains, and the fix is infrastructure โ intents, solvers, unified liquidity layers, chain abstraction, an interoperability protocol that lets you pretend the fragmentation is not happening.
I have spent a great deal of time inside that narrative, and I have come to believe that most of it is manufactured. Liquidity fragmentation is not a problem the market discovered; it is a problem infrastructure vendors need to exist so they can sell the solution to it. In 2017 the equivalent narrative was "interoperability," and it produced a dozen bridges, three of which were catastrophically exploited, and none of which meaningfully changed how anyone used a network. The real problem on a chain like Zora is not that liquidity lives somewhere else. It is that attention is finite and the supply of places to put it has become infinite. Another bridge does not fix that. It adds another place.
The Inscription Lesson, and Where Cheap Blockspace Goes
I want to close the analysis with an analogy I know will irritate some readers, because it is the clearest one I have.
In early 2023 a standard called BRC-20 appeared on Bitcoin, allowing people to inscribe token-like data into witness space, and for a few months it consumed an enormous share of Bitcoin's blockspace and pushed fees to levels ordinary users had not seen in years. Runes followed in 2024, formalizing and slightly rationalizing the practice. My view then and now is unchanged: using the most robust, most deliberately constrained, most conservatively engineered settlement layer in human history to haul around image pointers and ticker registries is like using a Rolls-Royce to haul cargo. It insults the car, and it does not carry much. The technique is clever. The application is a category error.
Zora's version of the same impulse is different in kind, and the difference matters. Its minting lives on a rollup whose entire reason to exist is cheap, high-frequency, low-value cultural transactions. You do not buy a Rolls-Royce to run a delivery route; you build the delivery route, and then you route the deliveries. But the analogy still carries a warning for the people inside Zora, and it is this: cheap blockspace will be consumed by whatever is cheapest to produce, always, without exception. If the culture around a protocol does not hold a line โ if it does not insist that a mint means something, that a creator is not a bot, that a collection is not a farming script wearing a costume โ then the cheapest producer will find the space and fill it. The blob spike of 2024 was, in part, precisely this. The subsidy invites the parasite.
Culture is the only defense, and culture is the one thing an AI-driven team of nine cannot automate. That is not a rhetorical flourish. It is a resource constraint, and it belongs on the balance sheet next to the runway.
Contrarian
Now the part where I try to talk myself out of my own pessimism, because the reflex to read layoffs as decline is the cheapest read available, and it is frequently wrong.
Here is the uncomfortable counterargument: the AI-driven, nine-person Zora may be exactly right, and the reductions may be a correction rather than a collapse. The 2021 vintage of crypto companies was overcapitalized and overstaffed by construction โ venture money arrived faster than product-market fit, and headcount became a signal of legitimacy rather than a cost of production. Teams of forty burned runway to build things that nine people and a fleet of agents can now build better, in less time, with fewer coordination losses. If Zora is being honest that the mint market is smaller than the story, then shrinking is not failure. It is integrity. There is a version of this reorganization that is not a retreat at all, but a protocol that has finally priced reality.
And yet. There is a blind spot in that reasoning, and it is the one I keep circling, because I have lived through enough cycles to see it recur: organizations do not primarily run on labor. They run on memory. The reason a project survives its own transitions is that somewhere inside it there are people who remember the decision that was made in 2022, and why, and what it cost, and what nearly broke. That memory is not in the documentation โ documentation records conclusions, never the argument that produced them. It is not in the model either, and I say that as someone who has spent two years of his life building cryptographic provenance for machine decisions. A model can tell you what the treasury did. It cannot tell you which of the nine people in the room was shouting.
Trust is not a metric; it is a memory we share โ and memory is labor-intensive. When a team goes from forty to nine, you do not lose thirty-one units of productivity. You lose thirty-one perspectives on your own history, and you usually do not discover which ones you needed until the incident. I have watched protocols learn this at three in the morning, mid-exploit, with an on-call rotation of exactly one.
There is one more thing, and it is the ugliest part of the pattern. In a bull market, "AI-driven operations" is very often the graceful exit in disguise. Automation is the perfect explanation for a smaller org chart because nobody can argue with it, nobody can audit it, and it does not require a founder to say the thing that is actually true โ that the market for what this company built is smaller than the raise implied, that the token has not delivered its promise, and that the people who made the thing are being let go while the protocol keeps running. I am not accusing Zora of this. I am saying that the industry has now produced enough of these announcements that the words have stopped carrying information, and that we should have the discipline to say so out loud. A reorganization announced with optimism is not evidence of optimism. It is evidence of a press release.
Takeaway
So: what is the minimum number of humans required to keep a protocol honest? It is not zero, and it is not a figure you can reach by adding agents. It is the number of people who can still be held responsible when the machine, executing a policy nobody quite remembers writing, moves funds that nobody quite remembers approving. Zora has chosen to find out in public, with a beautiful idea, a cheap chain, and a quorum.
I will be watching the attestation, not the announcement. If Zora publishes a verifiable record of what its agents do โ signed, anchored, auditable by anyone with a node โ then this week will look, in hindsight, like the moment a protocol grew up. If it does not, then nine people is not a team, and it is not a quorum. It is a countdown.
An audit is not a gate; it is a witness. The protocols that last are the ones that write it down.