Ad tech is trifurcating. Stop benchmarking every DSP against The Trade Desk.

Three ad-tech companies got repriced in the same week, and the market told three different stories about them.

The Trade Desk is down about 44% on the year and just hired an outsider CFO — Nate Olmstead, who ran finance at a book publisher — to re-narrate the margin story to the Street. Liftoff Mobile — itself a 2021 merger of two Blackstone mobile ad networks, the original Liftoff and Vungle — went public at $23 and closed its first week at $26.88, the first real ad-tech IPO since MNTN, with buyers paying up for a profitable, performance-native name. And OpenAI now sells both CPM and CPC ads inside ChatGPT, a live performance-ad product on the biggest consumer-AI surface there is.

One sector, one week, three diverging futures.

The easy read — ad tech is having a bad year, the strong DSPs pull away from the weak ones — is a year out of date. These three businesses share a category label and little else: different P&Ls, different buyers, different bets on where the next decade of budget sits. The market is repricing them as the separate things they are. The independent ad tech world is trifurcating and the walled gardens are putting their big bets on only one leg of the stool.

Three poles are pulling apart, each a different business: the independent open-web DSP, the closed-loop performance network, and the AI platform building an ad business.

1. The independent open-web DSP #

The Trade Desk is the big name here, and it’s facing a structural question, not a bad quarter. The whole business is priced on the open internet’s shared identity layer holding — the assumption that you can stitch a person across the open web well enough to target them and measure the result. That layer is cracking. Email-based IDs don’t hold up for CTV, which is exactly where the budget is moving, and the slow erosion of third-party identifiers across the rest of the open web hasn’t stopped.

Watch where The Trade Desk is hedging, not the stock chart. It started in late April, joining Goldman Sachs and Bain in an investment in Hightouch; then on June 1 it partnered with Hightouch to pipe its own impression logs straight into advertisers’ cloud data warehouses, resolving identity against records the brand already owns rather than in the bid stream TTD spent a decade building. When the canonical open-web DSP starts moving identity into the advertiser’s warehouse, read it as the company pointing to the ground it still trusts. The outsider CFO says the same thing: you hand the margin story to someone with no ad-tech background exactly when the job is to re-sell it to investors who’ve stopped buying the growth story.

The structural risk for this pole is the cleanest of the three. The model rests on the open identity graph holding, and that’s the assumption in doubt.

2. The closed-loop performance network #

Liftoff’s IPO is the latest evidence. It raised about $437M, and the market paid a premium for a company that’s profitable, performance-native, and paid on outcomes instead of impressions. AppLovin is the larger public version of the same animal, down 17% on the year, which only sounds rough until you set it beside The Trade Desk’s minus 44%. These names are getting marked down less than half as hard as the open-web DSP. That spread is the split showing up in the tape: the market doesn’t think these are the same business having the same year.

Moloco is the one still private — performance AI applied to CTV, closed-loop down to the install, increasingly running on its own first-party inventory. (I wrote about its performance-CTV launch as the cleanest “we don’t just have AI, we have the closed loop” pitch in the market; that’s the shape of this pole.)

The structural risk here is distribution. These businesses win by getting bigger or by getting absorbed. There’s no comfortable middle where a mid-sized performance network sits still and stays independent.

3. The AI platform building an ad business #

This pole didn’t exist a year ago, and this month it has a live product in market. OpenAI turned on CPC ads in ChatGPT, with CPA ads on its near-term roadmap. Criteo just cut its minimum buy for ChatGPT ads to $10,000, a long-tail-SMB number. Projections run from about $2.4 billion in ad revenue this year to $102 billion by 2030, a target close to half of Google’s $224 billion search business.

The economics sit somewhere the other two poles can’t reach. An AI platform’s ad business runs on compute, not on owned or brokered inventory. Its sales motion is long-tail and self-serve, closer to early Google than to a CTV upfront. Its regulatory exposure is its own — the FTC already put a $930K price on a fabricated AI-targeting claim, and a frontier lab selling ad outcomes will draw exactly that kind of attention.

It’s also the one pole the walled gardens can’t sit out: Google and Meta hold the same model-and-distribution hand, and the assistant surface is new enough that incumbency doesn’t automatically carry. OpenAI is just first and loudest.

The structural risk: nobody has shown that a frontier-model lab can run an ad sales operation at scale. The thing that changed this month is that the question went from hypothetical to empirical. It’s running now, on a near-zero distribution cost and a captive surface neither of the other two poles can touch.

What this changes for buyers #

If you’re writing a DSP RFP in 2026, stop scoring it against “what The Trade Desk costs, minus ten percent.” Most RFPs I see are still built as though every DSP is TTD-shaped — open-web, identity-graph-dependent, priced on the same curve. The useful question now is which of the three poles a vendor is competing on, and whether their answer to that pole’s structural risk holds up. A performance-native network, an open-web DSP, and an AI platform’s ad product are three different businesses that happen to clear through similar-looking line items. Pricing them against one benchmark is how you end up comparing an apple to an orange to a fruit roll up.

The next eighteen months won’t pick a winner across the three. They’ll set the order of magnitude of consolidation inside each one. The independent open-web pole probably absorbs at least one major name. The performance-native pole probably gets one more IPO that resets the comp set for everybody else. And the AI-platform pole probably spends the period watching the FTC and the SEC draw the lines around what it’s even allowed to sell.

This is the same direction of travel I traced when Viant bought TVision: the neutral middle gets pulled into the platforms on either side of it. Trifurcation is the wider-angle version of that move, across the whole category instead of one corner of measurement.

The position I’ll stake, and the one worth arguing with: by the end of 2027 the trifurcation is locked in, and anyone still pitching a single “DSP strategy” will sound like they’re describing a board that’s already been cleared and reset.