'Models Built For Me'

Full analysis

Ali Manning, COO and co-founder of Chalice AI, sat down with AdExchanger's Allison Schiff to argue something that ought to worry every DSP running an off-the-shelf bidder: the brand-specific model belongs inside the pipes, not on the DSP's terms. The claim underneath the whole conversation is that generic optimization, clicks, viewability, last-touch CPA, is the wrong target, and that a model trained on a brand's actual business goal can run at the SSP layer and route around the DSP entirely.

What's being decided (briefing frame): whether operators should treat containerized, custom bidding, a brand's own ML model running inside supply infrastructure with full bid-stream access, as a real architectural shift or a founder's pitch. Reversibility is Type 2 for most readers. You can pilot this without betting the company. The forcing function is the IAB Tech Lab's Agentic Real-Time Framework (ARTF), which gives the idea a standards home and a reason for DSPs to respond.

The Market Analyst. Follow the money and the layer it sits at. Andrew Casale's insight, that the full bid stream and URL-level signal live at the SSP, not the DSP, is a claim about where value accrues next. If brand logic runs inside Index Exchange's cloud, the SSP captures a function the DSP used to own. In plain terms: the middleman that decided what to bid on may get thinner. That favors SSPs (Index Exchange, PubMatic, Magnite) and pressures DSP margins. But note the tell: Manning also says "some DSPs are already exploring equivalent functionality." The incumbents don't need to lose. They need to open the box. The Trade Desk absorbing external models is the more likely equilibrium than The Trade Desk getting disintermediated.

The Skeptic. Manning profits from every sentence she said. The "Meta got 10 percent better without IDs" number is one she couldn't source, attributed to a co-founder's paper she half-remembers. The 80% in-store figure for One A Day is Chalice's own case-study material. The anti-curation line, "a new fancy way to say ad network," is convenient for a company selling the thing that isn't curation. For this to work, the containerized model has to beat the DSP's optimization by enough to justify running your own ML, and the only evidence offered is one Bayer case study with a favorable KPI. One win is a demo, not a category.

The Operator. Try to run this Tuesday morning. You need a brand data set worth training on, an offline-measurement partner like the panel Chalice used for One A Day, and an SSP willing to host your container. That is a lot of moving parts for a mid-size advertiser. Manning herself admits the gate: with 50 people, Chalice turns away any client that can't name a concrete success metric. That's honest, and it's also the whole game. Most brands can't. "We want sales and also viewability" is how real briefs read. This architecture works for the disciplined few and breaks for everyone else at scale.

The Customer / End User. The advertiser CMO and CFO are the point here, and Manning is right about one thing. New-to-brand purchases are a KPI a CFO recognizes. Click-through rate is not. The One A Day setup, bidding higher into Amazon DSP for modeled new-to-brand shoppers when 80% of sales are in-store, speaks the language of the person who signs the budget. That's the genuine pull. Buyers are tired of being sold click proxies dressed as business results. Whether they'll build the plumbing to fix it, or just keep complaining, is the open question.

The CFO. The economics cut two ways. Manning says the containerized setup surfaces undervalued inventory and lowers effective CPMs, which is real money if true. Against that: you're now paying for a custom-model vendor, an offline-measurement partner, and SSP-side compute, to replace optimization you already pay a DSP for. The payback only clears for advertisers spending enough that a few points of media efficiency dwarfs the added vendor stack. Chalice's own survival story tells you the scale problem: they got profitable when one customer went roughly 10x, not by signing many small ones.

The tensions.

The real disagreement is between the Market Analyst and the Skeptic on where this lands. Does brand logic at the SSP genuinely restack the value chain, or do DSPs simply add model-injection and neutralize the threat by the next renewal cycle? Both can point to the same sentence in the episode.

The second tension is the Operator against the Customer. The demand is real, CFOs want business KPIs, but the operational bar Chalice sets, refusing clients who can't name a metric, means the addressable market is narrow by design. The thing buyers want and the thing most buyers can execute are not the same thing.

Synthesis. This hinges on two beliefs. First, that full-bid-stream, URL-level prediction at the SSP beats DSP-sampled optimization by enough to matter. Plausible, but evidenced by a single case study and a vendor with every reason to say so. Second, that DSPs won't just copy the feature. On that, the episode gives the answer away: they're already exploring it. The council leans toward "real shift in where models can run, absorbed by incumbents rather than fatal to them." The winners are SSPs, who get a new reason to hold bid-stream data close, and disciplined large advertisers with clean offline measurement. The losers are pure "outcomes" arbitrage vendors guaranteeing CTR and CPA, whom Manning is right to call out.

Before committing, an operator should verify the one thing the episode can't: get a second containerized pilot's numbers, ideally your own, and compare effective CPM and a real business KPI against your current DSP setup. One Bayer result doesn't clear the bar.

Prediction: By the IAB Tech Lab's ARTF progress update at the next major industry milestone (Programmatic IO or an IAB Tech Lab release by mid-2027), at least one major DSP will publicly announce support for external or containerized model injection rather than ceding the function to SSPs.

Confidence: Medium. The episode says DSPs are already exploring it, and incumbents copy before they concede.

Why: Manning states directly that some DSPs are already building equivalent functionality, which means the capability is in motion inside the incumbents, not just at Chalice and Index Exchange. The pattern in ad tech is that platforms with distribution absorb threatening features rather than watch budget route around them, and a DSP that lets a brand run its own model keeps the seat even if it loses the optimization monopoly. The opposite outcome, DSPs standing pat while SSPs quietly capture the bidding-logic layer, is less likely because it would mean the largest buy-side platforms ignore a standards effort they can see forming and let their core function erode without a response.

Revisit by 2027-06-30: We're right if a top-tier DSP publicly supports external model injection or an ARTF-style container. We're wrong if no major DSP has announced such support and the capability remains confined to SSP-hosted pilots.