Netflix: Why Don't You Mount Us?

Full analysis

Adam Heimlich and Gareth Glaser spent an episode arguing that the whole buying chain, DSP to SSP to ad server, is on its way out, and that the end state is a brand's custom bidding algorithm living inside the publisher's own ad server. Netflix is their poster child. The question for operators: is this a real architectural shift worth building toward, or a pitch for Heimlich's own company dressed in a bigger idea?

What's being decided: whether sell-side decisioning collapses the middle of the supply chain, and whether publishers with real control over their inventory (premium streamers, walled gardens) start hosting advertiser logic directly. Easy to undo if you wait. Hard to undo if you re-architect your stack around a bet that doesn't arrive by 2028. Nothing sets a hard deadline here except Heimlich's own 2028 call.

The Market Analyst

Follow the money in this episode and it points one way: infrastructure, not intermediaries. Kevel raised $52 million by abandoning open-web ad serving and becoming plumbing for retail media networks, a segment growing 20 to 30% a year. Koddi is doing the same. That is the market voting for ad-serving infrastructure that data owners control, which is exactly the layer Heimlich says custom bidders will mount onto. The thin-layer thesis and the Kevel raise are the same trade from different ends. For a generalist: the valuable position is becoming the rails a retailer or a streamer runs their own ad business on, not another box in the middle taking a cut.

The Skeptic

The CAPI claim is the part of the pitch that everything else rests on, and Heimlich owns Chalice, whose entire value proposition is that it models a brand's own conversions from scratch. Convenient that Meta supposedly can't. The fact-check is blunt: Meta does use uploaded conversion signals for lookalikes and bid multipliers, which Heimlich describes, then reframes as mere "quarantine." The architecture point is real. A single advertiser's 100,000 conversions barely move a 50-billion-event pool. But "they just tag it for attribution credit" overstates it to make custom modeling look like the only rigorous option. Discount accordingly. The 2028 "DSP buying from SSP has gotta stop" call is a man describing the world he needs to be true.

The Operator

Try to "mount" a custom bidder inside Netflix's ad server on a Tuesday and the first thing that breaks is: who owns the auction logic, who certifies it, and who eats it when a containerized model misprices inventory at scale. Publishers do not hand their decisioning environment to a third party's code lightly. The reason DSP-to-SSP persists is not stupidity, it is that nobody has to trust anybody. Everybody runs their own box. Collapse the middle and you collapse the mutual distrust that keeps the chain honest. The agency problem is the harder one Heimlich named himself: to win AI systems integration work, Publicis and Stagwell have to look neutral across tech vendors while profiting from recommending media platforms. You can't hold both for long.

The Customer / End User

Large advertisers should hear the convergence argument clearly, because it is the most useful thing in the episode. LLMs and all-in-one AI campaign tools optimize toward the average, and every big brand goal is an edge case. Heimlich is right that off-the-shelf AI planning commoditizes strategy. The irony is that his fix, custom bidding, and the industry's drift, agent-allocated budgets across Google, Meta, and Amazon, point in opposite directions. Sophisticated advertisers want differentiation. The tooling pushes sameness. Meanwhile the Meta scraping terms give those same advertisers' publishers a real reason to walk: use any Meta business service and you grant Meta scraping rights for AI training. Heimlich says every publisher he's spoken to plans to drop Meta rather than comply.

The CFO

The economics that make the thin-layer idea viable are the same ones making it a hard sell. Infrastructure costs have fallen and AI is cheap enough to run bidding logic sell-side. Fine. But the payback assumes publishers invest to host third-party algorithms, advertisers trust someone else's environment, and the savings from cutting DSP and SSP margins land somewhere you can book. None of that is in a pilot today. The clean trade with visible payback is Kevel's: own the serving infrastructure for a category growing 20 to 30% a year. The speculative trade is betting your roadmap on premium publishers rebuilding their auction stack by 2028.

The tensions

Three real disagreements. First, the Skeptic versus the Analyst on CAPI: is Heimlich describing an architecture truth or selling his book? Both, which is why you take the mechanism and leave the conclusion. Second, the Customer versus the industry drift: advertisers need differentiation, the tools deliver convergence, and nobody has resolved which wins. Third, the Operator versus the Pitch on trust: the middle of the chain exists because nobody trusts anybody, and no amount of cheaper compute fixes that.

Synthesis

The episode hinges on one belief: that premium publishers will open their ad-serving environment to third-party bidding logic. Everything in the thin-layer pitch depends on it, and nothing in the episode shows a publisher actually doing it. The council leans toward the infrastructure read over the architecture read. The Kevel raise is money in the ground today; the Netflix "mount us" pitch is a 2028 hypothetical with a commercial sponsor. What's worth testing before building anything: does any premium streamer or retail media network actually let an outside algorithm run inside their auction, versus just buying infrastructure like Kevel's and keeping the logic in-house? The in-house answer is the likely one, and it does not need Chalice.

The convergence warning, though, stands on its own merit and costs nothing to believe. If you run a large advertiser, assume your off-the-shelf AI planning tools are quietly making your media plan look like your competitor's.

Prediction: Kevel will close or announce at least one new named retail media network or large data-owner customer running its ad-serving infrastructure before Chalice (or any custom-bidder) announces a live custom bidding integration embedded inside a premium streaming publisher's own ad server, measured by public announcements through 2027-07-31.

Confidence: Medium. The infrastructure trade has funding and a working playbook; the embedded-bidder thesis has a pitch and a 2028 date.

Why: Kevel just raised $52 million specifically to serve retail media networks, a segment growing 20 to 30% a year, and already powers Reddit's internal ad platform, so it has both capital and a proven motion for signing data owners. The thin-layer idea of mounting a brand's bidding algorithm inside a publisher's ad server requires a premium publisher to hand decisioning control to third-party code, which the Operator read says breaks on trust and certification before it breaks on cost. The opposite outcome, a streamer embedding an outside custom bidder before any new Kevel customer surfaces, would require publishers to reverse their reluctance to run foreign algorithms in their auction, and nothing in the episode shows one doing it. Heimlich himself dates the architecture shift to 2028, which tells you it isn't landing in the next three quarters.

Revisit by 2027-07-31: We're right if Kevel publicly announces a new retail media or data-owner customer in this window while no custom bidder has announced a live, embedded integration inside a premium streaming publisher's own ad server. We're wrong if a custom-bidding company announces such a live embedded integration with a premium streamer first.