David Steinberg moved 400 engineers off profit for three years to rebuild Zeta Global around data and AI, buying BoomTrain and DISQUS.
Zeta Global, now worth around $8bn, has more than 800 branded clients and four straight years of 30% top-line, 50% EBITDA and 75% free cash flow growth.
Steinberg says Zeta's Athena model runs 65% more efficiently and 12% more accurately than frontier models on its own data.
Zeta bet the company on owning its own data
Zeta Global is the product of a founder's bet that the AI shift would be won by whoever owned the data and the customer relationship. Steinberg rebuilt the company around that idea from 2017, giving up profit for two or three years while 400 engineers built a new data and AI platform.
Takeaways
Proprietary data is the durable moat. Steinberg cites Alex Karp's line that data sovereignty protects a company's alpha. Feeding proprietary data into frontier models gives away the differentiator. Zeta's data cloud holds 550m opted-in individuals and serves more than 800 global brands.
The rebuild delivered four years of compounding growth. From 2021 to 2025 Zeta compounded top line above 30% for four straight years, with free cash flow growth above 75%. Clients get 600% to 700% back on marketing spend against a best-of-breed 250%.
Small models beat frontier models on owned data. Athena, Zeta's inference model built with Fireworks AI and NVIDIA, runs 65% more efficiently and 12% more accurately than frontier models on Zeta's own data. Steinberg predicts a split between frontier models and small targeted inference models.
Clients pulled Zeta beyond marketing into BI. After clients were told to market skirts to women aged 18 to 35 while inventory sat elsewhere, Zeta built its Business Intelligence engine, first for inventory then hotel location planning. In Q2 2026, 89.6% of new code was automated.
Every technology wave brings the same death predictions. Steinberg says the cycle runs every 18 to 24 months. Companies with deep customer relationships or distribution that adopted the internet became juggernauts. He calls ChatGPT the moment AI moved from science fiction to boardroom conversation.
Zeta buys only tuck-ins that add unique data. Every Zeta acquisition has been a tuck-in, because transformative M&A transforms both companies for the worse. Targets must be accretive the day they close and bring technology or a dataset Zeta could not build as fast.
The board meeting that rebuilt Zeta
In 2017 Zeta ran its marketing platform in silos and expected about $50m in EBITDA. A client report the India team said would take a week, plus Steve Gerber's tip about BoomTrain, pushed Steinberg to rebuild.
He told the private equity backed board Zeta would make no profit for two or three years while 400 engineers built the new platform.
He wanted BoomTrain and DISQUS but could not afford both. An analyst's opposition in the most contentious board meeting in company history led him to buy both, and he still votes more than 50% of shares.
BoomTrain co-founders Chris Memberg and Nish Gore now run technology and data at Zeta. A Harvard case study written by Columbia with Harvard was published on that meeting.
Tap a name to go there · tap a line to see why they are linked