Consolidation & Aggression

Full analysis

Amazon is sending engineers into agencies to install Bedrock Agent Core as the one screen that runs every other advertising AI, and it's doing it before Google can. That's the claim Adam Heimlich made on this episode, and it reframes the fight that matters for the next two years. Not who has the best buying algorithm. Who owns the layer that sits on top of all of them.

Two threads run through the whole conversation, and they're the same thread. On the buy side, whoever controls the "master agent" (the top-level AI that tells the other AIs what to do) controls the account. On the publisher side, the prebid management shakeout shows what happens to a market where everyone runs the same open-source code and the only thing left to compete on is price. Both are stories about leverage collapsing to whoever owns something the others can't route around.

This is hard to undo for anyone who picks a master-agent horse early. Rip-and-replace of an orchestration layer wired into your agency workflow is not a Tuesday-morning decision. The prebid piece, by contrast, is easy to undo and mostly already decided.

The Market Analyst. Heimlich's framing puts The Trade Desk in the worst seat. The whole pitch of a neutral demand-side platform (the tool agencies use to buy ads across the web) was that neutrality is a feature. In an agent war, neutrality is a bug. If the only real leverage is inventory nobody else can sell, then Google (YouTube, Search), Meta (Facebook, Instagram), and Amazon (retail, Prime) each hold a hostage. The Trade Desk holds none. That's why Heimlich floats a Trade Desk tie-up with Anthropic or Microsoft: buy your way into a layer you can't build. For an informed outsider: the companies that own must-buy ad space get to set the terms, and the middlemen who don't will pay rent or disappear.

The Skeptic. Slow down. Amazon "sending engineers to agencies" is a sales motion, not a market. Where's the spend? Heimlich's own $5 billion ads-in-chat number is unsourced and, by the fact-check, speculation. And Adam Heimlich and Gareth Glaser already name the reason this may not work: Google and Meta have every incentive to refuse to connect to an Amazon-run master agent, exactly as Facebook gated its API for years. A master agent that can't touch YouTube, Search, or Facebook manages a rounding error. Amazon lacks the one thing that forces interoperability, which is inventory the agencies must buy through it. The precedent Heimlich and Glaser cite cuts against the thesis Heimlich is selling.

The Operator. Think about the agency execution reality. You do not run Coke and Pepsi through one integrated buying brain today because holding companies physically wall those teams off, and Mark Ritson's piece nails the double standard: Google and Meta already pool both brands in one algorithm and nobody blinks. So when a master agent shows up promising to orchestrate everything, the first question from any agency's legal and client-services team is: whose data touches whose. Gareth Glaser's answer, per-advertiser models using only that advertiser's data, is the clean version. But installing that inside a live agency, mid-year, across dozens of conflicting accounts, breaks on governance long before it breaks on technology. For a non-specialist: the plumbing is the easy part. Keeping rival clients' data apart is the part that gets someone fired.

The Customer / End User (the publisher). The prebid story is where a real operator gets hurt this quarter. Gareth Glaser says the prebid battle is over, open-source Prebid.js won, and five management companies collapse to two or three. When everyone runs identical code, the only lever left is the revenue guarantee, and that's why Playwire CEO Jason Dubin is calling out what looks like a Raptive case study on LinkedIn. Here's the trap for a mid-tail publisher: a guarantee juiced by temporary private-marketplace deals looks great for two quarters, then reverts. You switched vendors, you ate the migration, and your baseline is back where it started. The guarantee sold you a number, not a business.

The CFO. Follow the money on both fronts and it's the same lesson: don't pay to rent leverage you'll never own. On the buy side, if your agency wires itself into one vendor's orchestration layer, your switching cost becomes that vendor's pricing power. On the publisher side, chasing the biggest guarantee is negative-margin thinking once you count the migration cost and the revert. The APS retreat rumor Gareth Glaser floats is the evidence that even Amazon does the math: a server-to-server product losing to free open-source code is a line item you eventually kill because the juice isn't worth the squeeze.

Where the smart people split. Adam Heimlich thinks the master agent is the whole game and Amazon striking first matters. The skeptical reading of the same episode says the walled gardens will simply refuse to connect, and an orchestration layer that can't reach the big three inventory pools is a demo, not a business. That disagreement is the decision. The second split is narrower but matters more for publishers: is a revenue guarantee a real edge or a financing trick? Gareth Glaser's own description, temporary tactics that revert, answers it.

What it hinges on. One fact settles most of it: will Google and Meta let a rival's master agent operate their inventory? If yes, neutral orchestration is a real market and The Trade Desk's exposure is overstated. If no, and the Facebook API precedent says no, then the master-agent layer belongs to whoever owns must-buy inventory, and neutral players pay rent or partner up. Before anyone wires a workflow into Bedrock Agent Core, the thing to confirm is whether it can actually execute a YouTube or Facebook buy without bouncing you back into Google's or Meta's own screen. Everything else is a sales deck.

Prediction: By May 31, 2027, one of the six largest agency holding companies (Publicis, Omnicom, WPP, Dentsu, Havas, Stagwell) will publicly announce that it uses Amazon's Bedrock Agent Core, the Amazon Web Services product Amazon pitches as the single control screen directing other advertising AI tools, to plan or buy client media on platforms Amazon does not own.

Confidence: Medium

Why: Amazon is putting its own engineers inside agencies to install Bedrock Agent Core, and it is moving before Google offers an equivalent. Ripping out an orchestration layer once it is wired into agency workflow is not a quick decision, so Amazon is subsidizing installation now to hold pricing power later, and a holding company gets an AI story it can sell to clients by saying yes in public. The case against is real: Google and Meta have no reason to let a rival's system run buys inside YouTube, Search, Facebook and Instagram, echoing the years Facebook kept its developer interface gated, and agencies must keep rival clients' data walled off, which is where mid-year installs usually break. But Amazon does not need either walled garden's blessing to land this announcement; it needs one holding company willing to be the reference account, and it is paying for exactly that.

Revisit by 2027-05-31: We're right if Right if, on or before May 31, 2027, a press release or earnings presentation from Publicis, Omnicom, WPP, Dentsu, Havas, or Stagwell states that the company is using Amazon's Bedrock Agent Core to plan or buy client media on media platforms Amazon does not own. We're wrong if Wrong if no such statement from any of those six companies exists as of May 31, 2027.

The prebid consolidation is the safer read and worth stating plainly: the five-to-two-or-three collapse Gareth Glaser describes is already in motion, and revenue guarantees are the accelerant. But the money and power move on the buy side, which is why that's the call.