The board · 7 ranked
The leaderboard.
Private adtech companies, judged by three models from three different labs. Split by which side of the supply chain they sit on, because an SSP and a measurement vendor are not doing the same job and should not share a table.
What is on it
- Ranked
- 7 companies
- Weight
- 3 middleweight · 4 lightweight
- Age
- 4–21 years
- Scores
- 10–22.3 of 30
- Panel
- NVIDIA, Zhipu AI and Google DeepMind
The board.
They own or represent the inventory.
- 1
BlockthroughAdblock Revenue RecoverylightweightdependentwhitelistedvulnerablePanel in agreementNot convinced16/30 - 2
EzoicPublisher Monetizationmiddleweightadvertiser-ledsprawlingfragilePanel split — 5 points between highest and lowestNot convinced15.6/30 - 3
AdPushupPublisher MonetizationmiddleweightacquiredcoastingwrapperPanel in agreementNot convinced14.6/30
Nothing on this side matches that filter yet.
They spend somebody's media budget.
They sell nobody's inventory and spend nobody's budget.
- 1
The Media TrustFraud & Traffic Qualitylightweightbattle-testedgrizzledvigilantPanel split — 5 points between highest and lowestWorth a meeting22.3/30 - 2
ConfiantFraud & Traffic Qualitylightweightthreat-intel-ledsentryvigilantPanel in agreementWorth a meeting22/30
Nothing on this side matches that filter yet.
Rare air 24+ · Worth a meeting 19+ · Not convinced 12+ · Brutal <12 · Ticks — each panelist's own total; scattered ticks are a split panel
Judged by three language models reading public pages, and written up by a fourth. An opinion, not research. Want a row gone? Email hello@andorlabs.ca and it goes.
The panelists.
Three models from three different labs — NVIDIA, Zhipu AI and Google DeepMind — each handed the same pages and asked the same three questions. Each holds one fixed lens across all three, so the engineer answers the outlook question as an engineer. They cannot see each other's answers — the calls run at the same moment, which is what makes agreement between them worth something.
- NVIDIANemo
Staff Engineer, Seat 1 · I judge by what breaks at 3am and who gets paged.
Nine years on the exchange side, most of it in the part of the stack nobody demos. Holds that a product is whatever survives Black Friday, and that everything else is a landing page. Reads the careers page before the homepage.
Nemotron 3 Ultra — Mixture-of-experts, 550B total parameters with roughly 55B active per token — both numbers are published in the model's own name. Trained by the company that makes the accelerators everyone else rents.
- Zhipu AIAtlas
Partner, Seat 2 · I judge by what this looks like at 10x revenue and whether anyone is left to buy it.
Partner at a fund you have heard of and cannot quite name. Passed on three companies that later mattered and has made peace with exactly one of them. Will happily tell you a great product is a bad business, which is the most useful thing anyone on this panel does.
GLM 5.3 — Open weights with a published architecture, though the exact size of this tier is undisclosed. Built by a lab that spun out of Tsinghua and ships more than it announces.
- Google DeepMindJuno
Operator-in-Residence, Seat 3 · I judge by whether this survives the renewal conversation eighteen months in.
Three exits, two of which are not up for discussion. Has sat through roughly four hundred QBRs and can tell you the exact moment a renewal died in each one. Holds that most category-defining products are one procurement cycle from being a line item someone forgets to cancel.
Gemini 3.5 Flash Lite — The smaller of Google's fast tiers. Parameter count undisclosed, architecture undisclosed, and the word 'Lite' is doing all the disclosure there is.
Vega, Clerk of the Panel, writes the summary and scores nothing. Never averages. When the panel splits, the split is the finding. Played by GPT-5.6 Luna (OpenAI). Parameter count undisclosed. Present solely to turn nine numbers into a paragraph, and disqualified from voting on the grounds that it has read everyone else's answers.
The three questions.
- Innovationhard no → extremelyWas this first, or only?
- Hard to buildhard no → extremelyIs this hard to replicate?
- Future outlookhard no → extremelyWill this still matter in three years?
Nine answers in all, averaged into three scores out of ten. The total is out of thirty.