OpenAI launched self-serve advertising in ChatGPT. When inference economics don't close, even the most advanced AI company converges on the oldest internet business model.
On May 5, OpenAI opened a self-serve Ads Manager for ChatGPT. Any business can now bid on cost-per-click placements inside the conversational interface that two hundred million people use weekly. No minimum spend. The most technically advanced artificial intelligence company in the world just adopted the oldest business model on the internet.
The trajectory was swift. In February, OpenAI launched a pilot with a two-hundred-thousand-dollar minimum and CPM pricing starting at sixty dollars. By April, the minimum dropped to fifty thousand dollars and the company added cost-per-click bidding at three to five dollars per click. By May, the minimum vanished entirely. In ten weeks, CPMs eroded from sixty dollars to twenty-five to fifty — fast enough to force the pivot from impression-based to click-based pricing before the product had left its first quarter.
The confirmed launch partners tell the story of where the ad dollars come from: Target, Best Buy, Williams-Sonoma, Ford, Adobe, Expedia. Six hundred advertisers signed on within six weeks, generating a hundred million dollars in annualized revenue. OpenAI projects two and a half billion in ad revenue for 2026, scaling to eleven billion by 2027, twenty-five billion by 2028, and a hundred billion by 2030 — the last figure assuming 2.75 billion weekly active users. Geographic expansion began May 7, adding the UK, Japan, South Korea, Brazil, and Mexico to the initial markets.
The structural question is why. OpenAI crossed twenty-five billion dollars in annualized revenue in early 2026. The company projects fourteen billion dollars in losses this year. Inference costs at scale remain higher than revenue can cover, and the gap widens with every new capability that increases compute per query. The Bleed Rate documented this arithmetic in May: revenue doubled while losses nearly tripled.
Google waited four years before introducing the pay-per-click auction that became its primary revenue engine. The company was founded in 1998, launched AdWords with impression-based pricing in 2000, and introduced the CPC auction model in 2002. OpenAI waited two and a half years — from ChatGPT's launch in November 2022 to the February 2025 pilot. The compression ratio is not primarily about impatience. Google's search product had near-zero marginal cost per query. OpenAI's does not. Every conversation costs inference compute, and that cost scales with capability. The advertising timeline compressed because the economic pressure arrived sooner.
The ads appear only on Free and Go tiers. Plus, Pro, Business, and Enterprise subscribers see no advertising. This creates a two-tier information architecture: users who pay receive answers optimized for accuracy, while users who do not pay receive answers shaped by an additional variable — advertiser willingness to pay for placement adjacent to the response. The company that built its reputation on giving the best available answer to any question now maintains two definitions of best, distinguished by whether the user is the customer or the product.
CPM erosion from sixty dollars to as low as twenty-five dollars in ten weeks suggests conversational AI ad inventory commoditizes faster than search did. In search, the user's intent is explicit — the query itself signals what the user wants to buy. In conversation, intent is diffuse, contextual, and often non-commercial. The targeting signal is weaker, which means the inventory is less valuable per impression, which means volume must compensate. Cost-per-acquisition pricing is reportedly in development for late 2026, the next step in the familiar descent from brand awareness to direct response to performance marketing.
The pattern is not unique to OpenAI. Every major consumer internet platform that achieved scale eventually converged on advertising as its primary or secondary revenue model. The exceptions — those that sustained themselves on subscriptions or transactions alone — are notable precisely because they are exceptions. The structural logic is consistent: when the cost of serving users exceeds what users will pay directly, the gap gets filled by selling access to those users. The question was never whether OpenAI would adopt advertising but when the arithmetic would force the decision.
Falsification: If OpenAI subscription revenue exceeds advertising revenue through 2027, the thesis that advertising is an inevitable complement to inference-heavy AI products fails. The alternative explanation — that advertising is a temporary bridge while inference costs decline — would be validated.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)