The Vendor Problem

OpenAI launched an advertising platform.

Pilot in January. Self-serve Ads Manager in May. Sponsored answer cards, product spotlights, contextual sidebar placements. Partners include Dentsu, Omnicom, Publicis, WPP. Early advertisers: Best Buy, AT&T, Target, Adobe, Expedia, Ford, HelloFresh. Revenue target: $2.5 billion in 2026. Trajectory: $100 billion by 2030.

This matters to People Inc because D/Cipher — their most important proprietary asset — runs on OpenAI.

Four Financial Flows

No technology vendor relationship in media has ever looked like this. Google was a platform dependency — you sold ads through their pipes. OpenAI is something new: the company that powers your ad intelligence product while building its own ad intelligence product that competes for the same advertiser budgets.

The Signal Asymmetry

D/Cipher and OpenAI Ads both sell contextual targeting. Both promise privacy-first, cookieless ad matching. Both pitch to the same agency holdcos — Dentsu, Omnicom, Publicis, WPP are OpenAI launch partners AND People Inc's core buy-side relationships.

But they read different signals.

D/Cipher's taxonomy is deeper. ~100M scored connections built over years of editorial classification, NLP analysis, and OpenAI's own semantic enrichment. OpenAI's contextual targeting is broader — ~200 topic clusters, but reading direct user intent from conversations rather than inferring intent from what someone is reading. An advertiser buying "Hamburg restaurants" on D/Cipher reaches someone reading a Hamburg travel article. The same advertiser buying on OpenAI reaches someone asking where to eat in Hamburg.

Direct intent commands a premium. Every advertiser knows this from search. The question is whether content-inferred intent can defend its share when conversational intent is available at scale.

The Dependency

In September 2024, People Inc moved D/Cipher's language understanding layer entirely to OpenAI. The batch re-analysis of the entire content library. The semantic enrichment that finds connections human taxonomists miss — Hamburg museums to Hamburg restaurants. The parsing of non-text content: video, images, audio. 70% of OpenAI's connections matched existing NLP. The 30% that were "meaningfully different" are the competitive edge.

That 30% runs on your competitor's infrastructure.

Jon Roberts, the Chief Innovation Officer, has described the OpenAI relationship as "very, very positive" and said the company has "a seat at the table" with OpenAI's product team. He also said People Inc "deliberately does not have productive relationships" with OpenAI's competitors. That was a defensible position when OpenAI was a technology vendor. It is a dangerous position when OpenAI is an advertising competitor.

What Changed

OpenAI is losing more than $5 billion annually. Compute costs scale with free-tier usage. 800 million weekly users, 95% on the free tier. Ads are the only monetization model that turns free users into revenue without converting them to subscribers.

The advertising business isn't optional for OpenAI. It's structural. The $2.5 billion 2026 target, the $100 billion 2030 target — these aren't experiments. This is an AI company that needs advertising to survive the gap between inference costs and subscription revenue. Perplexity tried and killed its ad program in two months. OpenAI has the user base to make it work.

The competitive dynamic is already visible. People Inc's content appears in ChatGPT under the $16M licensing deal. OpenAI serves ads alongside that content. The advertiser who once could only reach that content's audience through D/Cipher can now reach a broader audience discussing the same topics through OpenAI — using People Inc's own content as the context, with OpenAI pocketing the ad revenue.

The Multi-LLM Signal

There is one signal that suggests People Inc sees the problem. A growth marketing technology engineering role posted in May 2026 mentions "experimenting with AI-assisted and agentic workflows using tools such as n8n, ChatGPT, Claude, Lovable, Clawdbot, and custom scripts."

Claude. In a People Inc job posting. Despite Roberts saying the company deliberately avoids productive relationships with OpenAI's competitors.

The gap between public narrative (deep OpenAI exclusivity) and hiring reality (multi-provider experimentation) suggests internal recognition that single-vendor dependency is untenable. Growth marketing is a low-risk place to start diversifying. The question is whether multi-LLM experimentation reaches D/Cipher — where the vendor lock-in actually matters.

Three Paths

The engineering response to a vendor-competitor problem has three levels, each progressively harder.

Path 1: Abstract the inference layer

Wrap OpenAI calls behind a model-agnostic API. Test Anthropic Claude, Google Gemini, and open-source models (Llama, Mistral) against the same taxonomy inputs. Measure quality of semantic connections. If the 30% differential holds across providers, the intelligence is in the taxonomy, not the model. If it doesn't, you need to understand exactly which model capabilities drive which connections before you can switch.

This is engineering work, not strategic work. Two engineers, two sprints, a benchmark suite. The cost of not knowing is higher than the cost of finding out.

Path 2: Internalize the embeddings

D/Cipher's pre-computed taxonomy is the defensible asset. The OpenAI dependency is in generating and scoring those connections. If the embedding computation can move to a self-hosted model — or even a fine-tuned open-source model trained on D/Cipher's existing taxonomy data — the intelligence stays proprietary and the vendor relationship becomes optional.

This is harder. Fine-tuning requires the right training data (which People Inc has — years of human-validated taxonomy connections), ML engineering talent (which the four-geography org may not have in sufficient density), and a willingness to accept that an 85% quality model you control is better than a 100% quality model your competitor controls.

Path 3: Accept the dependency, mitigate the exposure

If D/Cipher genuinely performs better on OpenAI than alternatives — if the 30% differential is model-specific — then the engineering response is contractual and architectural. Negotiate data isolation guarantees. Ensure D/Cipher taxonomy data isn't training OpenAI's ad targeting. Build monitoring for the gap between what D/Cipher sends to OpenAI and what OpenAI's ad platform knows about the same content.

This is the weakest path. Contracts don't survive incentive misalignment at scale. If OpenAI's ad business reaches the $100B target, the value of understanding publisher content for ad targeting will exceed the value of any individual publisher contract.

What I'd Recommend

Path 1 immediately. Path 2 as a 6-month project. Path 3 only as a bridge while Paths 1 and 2 are underway.

The multi-LLM signal in the job posting is encouraging. But growth marketing experimentation with Claude is not the same as D/Cipher running on anything other than OpenAI. The critical question — can the taxonomy intelligence survive a model swap? — can only be answered by testing it.

D/Cipher is People Inc's most important asset. It's in 50%+ of direct digital revenue. Orders are 50%+ larger than non-D/Cipher orders. The entire "shrink-to-grow" strategy — cutting editorial headcount while growing ad revenue — depends on D/Cipher continuing to deliver premium CPMs. That asset currently runs on a competitor's infrastructure.

The Google antitrust saga taught publishers what happens when your infrastructure vendor controls your economics. People Inc filed an 88-page lawsuit describing exactly that dynamic. The OpenAI relationship isn't there yet. But the trajectory is clear. And unlike Google, where the dependency accumulated over a decade, the OpenAI dependency was a deliberate architectural choice made in 2024.

The question isn't whether OpenAI will compete with D/Cipher. It already is. The question is whether People Inc's engineering team recognized the vendor problem before the advertising team did.