The Balance of Power Argument Applies to Advertising
On August 10, Mark Zuckerberg argued that the safest path for superintelligence is to distribute it broadly rather than concentrate it in a few institutions. We agree with the core of it, and the argument holds one layer down, in the infrastructure that will decide how AI surfaces make money.
Advertising is not the first way AI surfaces make money. Subscriptions and metered token consumption came first and are larger today. But those are bilateral: a user or an enterprise pays the model provider directly, and no third party holds a position. Advertising is the first mechanism requiring a market, where publishers, advertisers, agencies, and intermediaries all hold positions, and it is what funds the free tier where most of the users are. That is why advertising is where the balance of power question bites first.
What the argument actually says
The piece rests on three claims: individual empowerment produces prosperity, invention rather than automation is the purpose of the technology, and a balance of power across many holders beats a single well intentioned holder. The third carries the weight. Alignment cannot resolve genuinely conflicting values, so the practical answer is distribution: many agents, many principals, checking each other. The summary line is blunt. “There is no such thing as a singular benevolent superintelligence.” That is a claim about market structure dressed as a claim about AI safety, and it is almost word for word what the sell side has said about walled gardens for fifteen years.
The essay drew immediate criticism aimed at the messenger rather than the argument: the case for distributed power is being made by a company built on closed models, aggregated personal data, and ad revenue. Fair. It is also the oldest dynamic in our industry, where every company argues for transparency in the auctions it does not control and discretion in the ones it does, which is why the argument should be judged on architecture rather than intent. The question worth asking about any agentic ad system is not whether its operator means well. It is whether the structure would still protect the counterparty if the operator stopped meaning well.
The centralization risk already has a verdict
The default path is a closed loop where one company owns the buying tools, the ad server, the exchange between them, and the measurement that grades the result. Advertising already ran that experiment, and there is a verdict. In April 2025 a federal court found that Google illegally monopolized the publisher ad server and ad exchange markets and unlawfully tied the two products together. Publishers had been describing that structure for a decade before a court agreed with them. The remedy is still pending.
That shape is now being rebuilt one layer higher. A surface that owns the model, the interface, the ad decisioning, the demand, and the measurement is the same structure with better latency. When one entity holds every layer, no counterparty has standing to ask why a bid lost, and the answer arrives years later from a courtroom instead of from a log file. The agentic version is tighter still, because the surface is also the interface and the user never sees a page.
Alignment in advertising means knowing whose agent it is
The essay defines alignment as an agent serving its principal's goals rather than its builder's, which converts into a testable question: whose agent is this, and who does it answer to when the interests diverge? A buying agent operating inside the seller's platform is not aligned with the buyer.
This is why Adgentek builds the two sides as separate systems. The Agentic Ad Server is sell side decisioning. It works for the publisher or the AI surface, and its mandate is that inventory's value. Adgentek ORCA, Orchestrated, Real-time, Collaborative Agents, is the buy side counterpart, dispatching agents for advertisers and agencies, negotiating against sell side agents rather than through them.
The obvious objection is that both sit inside one company and roll up to one P&L. That is the same shape the industry has fought over for a decade, and no assurance about intent resolves it. The structural answer is substitutability. ORCA transacts against third party seller agents over AdCP and does not require Adgentek supply. The Agentic Ad Server takes demand that has nothing to do with ORCA. Neither side has a private interface to the other. The interface is a public specification any counterparty can read. If a publisher never takes an ORCA bid and an advertiser never buys Adgentek inventory, both still work. That is the test, and it is the one we would apply to anyone else.
| Layer | Centralized model | Distributed model |
|---|---|---|
| Decisioning | Surface owner runs the auction | Publisher controls decisioning |
| Buying agent | Operated by the seller | Operated by the advertiser or agency |
| Format | Set unilaterally by the surface | Set by an open specification |
| Intent signal | Retained by the surface | Returned to the parties who earned it |
| Take rate | Not visible to the counterparty | On the record and negotiable |
Distribution requires a protocol, not a platform
Distributed capability only works if two parties can transact without first joining the same platform, and that is what a protocol is for. Adgentek is a founding member of AgenticAdvertising.org, the organization that stewards the Ad Context Protocol. AdCP is an open specification governed by a consortium, not an Adgentek product. A protocol owned by one vendor is a platform with better marketing.
Meta shipped against the essay the same day, open weighting a 30 billion parameter agentic model under Apache 2.0. The same test applies here. An open protocol is only open if the reference implementations are.
Agent to agent transaction already works. In January 2026, FreeWheel completed an MCP based media buy with sell side agents from NBCUniversal and a buy side agent operated by the agency RPA. The agency cited fee reduction, which is what happens when the layers between buyer and seller become negotiable. None of this replaces real time bidding, which will carry most spend for years. What changes is that a second path opens where terms are set by a specification rather than by whoever owns the pipe.
Invention over automation is the right frame for ad agents
Agents in advertising today mostly automate work that already exists, which has a low ceiling. The bigger opportunity is formats that could not exist before the surface became conversational. Spark is our version: the ad is a conversation the user can interrogate rather than a static link appended to an answer. It delivers 3x to 8x higher engagement than display, averages 3.2 interactions per session, and produces intent signal a static link cannot. That is conversational advertising as invention rather than automation.
What this means for publishers and the long tail
The question that matters for publishers is whether agentic monetization becomes available to everyone or only to the fifty largest sellers. The essay's lawyer thought experiment maps cleanly: if one advertiser has an agentic buying stack it wins auctions it should lose, and if every advertiser has one the market gets more efficient. The same runs on the sell side: if only the largest publishers can stand up agent facing infrastructure, everyone else gets disintermediated by surfaces that already have scale.
There are more than 800 million weekly AI assistant users. Most publishers have no way to reach them, because their content is read by agents and summarized with no monetization path attached. Closing that gap has to be something a mid sized publisher can do in a week.
Two more objections worth taking seriously
The first is that distribution produces fragmentation rather than balance, and fragmented markets get consolidated by whoever can afford the pieces. Fair reading of ad tech history. OpenRTB did not prevent concentration; the largest players implemented it and outspent everyone else. Distribution is necessary for balance, not sufficient.
The second is sharper, and the one we hear most: if buyer agents and seller agents each take a fee, the industry has swapped one set of intermediaries for another. The protocol alone does not solve that. What it changes is whether the fee is legible. Inside a closed loop, the take rate is a number the counterparty cannot see and therefore cannot negotiate. Over a published specification, terms are on the record and a buyer can route around a price it does not like. Legible fees compress. Fees inside a closed loop do not.
Getting started
Adgentek builds the infrastructure for the distributed version of this market: the Agentic Ad Server for sell side decisioning, AdsMCP as its MCP integration path, Spark as the conversational format, and ORCA on the buy side. All of it is built to AdCP 3.1, so neither end has to join a platform to transact.
If you run an AI surface, a publisher business, or a brand that wants to advertise in AI chatbots without handing the auction to the surface, we should talk.
Common questions
What does the balance of power argument mean for AI advertising?
It means the structure of who controls agentic ad infrastructure matters more than how any single system is tuned. If one company owns the model, the surface, the auction, and the measurement, publishers and advertisers have no counterparty with standing to challenge an outcome. Distributing decisioning across independent buy side and sell side agents, connected by an open protocol, keeps leverage with the parties that own the audience and the budget.
Does agentic advertising just replace DSP and SSP fees with agent fees?
That is a real risk, and the protocol alone does not solve it. What an open protocol changes is whether the fee is legible. Inside a closed loop, the take rate is a number the counterparty cannot see and therefore cannot negotiate. When agents transact over a published specification, terms are on the record and a buyer can route around a price it does not like.
How does Adgentek keep buy side and sell side agents aligned to different principals?
Adgentek builds them as separate systems with separate mandates. The Agentic Ad Server is sell side decisioning that works for the publisher or AI surface. Adgentek ORCA is the buy side counterpart that dispatches agents for advertisers and agencies. The two transact over AdCP rather than inside a shared platform, so neither side is negotiating against a counterparty that also operates its own agent.
