Skip to main content
MindStudio
Pricing
BlogAbout
My Workspace
OpenAI ads businessOpenAI ARR adsChatGPT ads revenue

How OpenAI's Ads Team Hit a $1 Billion Run Rate So Fast

OpenAI's ads business reportedly hit a billion-dollar run rate in under a year. Here's what that reveals about building with LLMs instead of ML pipelines.

Edited by Luis Chavez-Mattos, Director of Product RSS
How OpenAI's Ads Team Hit a $1 Billion Run Rate So Fast

How did OpenAI’s ads team hit a billion-dollar run rate so quickly?

OpenAI’s advertising business reportedly reached a billion-dollar annual run rate in under a year, a pace that stands out even by the already-compressed timelines of generative AI products. The detail surfaced in a conversation between Vijay Raji, OpenAI’s CTO of Applications, and tech journalist Alex Kantrowitz, as part of a broader discussion about how coding and product-building have changed inside the company since agentic tools took over much of the engineering workflow. The core idea: a small team built the ads business by leaning on large language models directly, instead of building and training the custom machine learning pipelines that ad businesses have traditionally required.

TL;DR

  • OpenAI’s ads business hit a billion-dollar run rate in under a year, a speed that would normally require large, specialized ad-tech teams and years of model development.
  • The team behind it used LLMs in place of traditional ML pipelines, skipping the usual cycle of building, training, and tuning bespoke models for targeting, ranking, and prediction.
  • This mirrors a company-wide shift Raji described, where coding has become cheap enough that ideas move from concept to working prototype almost immediately.
  • OpenAI’s engineers have seen roughly a 17x increase in token usage tied to AI-assisted coding over the past year, compared to about 2x at typical software companies.
  • That speed created downstream bottlenecks in code review, CI/CD, and infrastructure, serious enough that OpenAI built a replacement for Git to handle agents that work “online all the time” instead of offline.
  • The ads story is presented as one example of a broader pattern inside OpenAI: small teams using frontier models as the build layer, not just the product layer.
AI-built software for mission-critical enterprise applications.
911 Dispatch ConsoleHarbor County Emergency Services
LIVE
✓ Audit log✓ Role-based access✓ Self-hosted

Why does this matter for an ads business specifically?

Traditional digital advertising businesses are built on machine learning infrastructure. Predicting which ad a user will click, ranking inventory, scoring relevance, these are classic supervised learning problems that normally require dedicated data science teams, labeled training data, feature engineering, and a slow iteration loop between model training and deployment. Companies like Google and Meta have spent decades and employed thousands of engineers refining these pipelines.

What Raji describes is a shortcut: instead of building that pipeline from scratch, the team used large language models to do work that would previously have required purpose-built ML systems. An LLM can reason about relevance, generate and evaluate candidate matches, and adapt to new patterns without the months-long cycle of collecting data, training a model, validating it, and shipping it. That compresses the traditional ad-tech build timeline dramatically.

This is consistent with the broader theme Raji laid out in the interview: coding and system-building have gotten cheap enough that teams can skip steps that used to be mandatory. If an LLM can do the reasoning work that a custom model used to do, a small team can stand up a functioning ads product without first building a traditional ML stack.

How small was the team, and what does that imply?

The framing throughout the conversation is that this was accomplished by a lean team relative to what a billion-dollar run rate would normally require. That’s the real headline, not just the revenue number. Ad businesses at that scale have historically needed large organizations: sales, ad operations, machine learning engineers, data infrastructure teams. Hitting that run rate with a small team signals that LLMs are absorbing work that used to require specialized headcount.

This connects directly to a pattern Raji described happening elsewhere in OpenAI. He talked about how, as a leader, he’s shifted from only directing teams toward ideas to personally prototyping those ideas using coding agents like Codex. He gave the example of building a working voice-modality prototype for ChatGPT himself, going from idea to functioning app before ever bringing it to his team. If a CTO can prototype a product feature solo, a small team building a full ads business with LLM-driven infrastructure instead of custom ML models becomes far more plausible.

What problems did speed create elsewhere at OpenAI?

The ads story doesn’t exist in isolation. Raji’s broader account of 2026 at OpenAI is about speed outpacing infrastructure. Starting around February, a combination of more capable models and agentic coding tools (including Codex running multiple agents in parallel using work trees) triggered a sharp jump in how much code engineers were shipping. Raji cited internal data showing token usage tied to coding grew 17x over the past year at frontier labs, versus roughly 2x at a typical software company.

Self-host
Remy.
On-prem or in your private cloud. Remy and every app you build stay inside your walls.

That acceleration created second- and third-order problems. More code shipped faster meant code review became a bottleneck. Review backlogs meant CI/CD pipelines needed rethinking. And the sheer volume of agents checking code in and out of OpenAI’s mono repo exposed limits in Git itself, a system designed for engineers working offline and periodically syncing, not for agents operating continuously. OpenAI built an internal replacement system to handle sparse checkouts, caching, and throughput at a scale Git wasn’t designed for.

The ads team’s experience is a variant of the same story: once you remove the traditional bottleneck (in this case, building custom ML models), the limiting factor shifts somewhere else, whether that’s infrastructure, review processes, or organizational speed.

Is replacing ML pipelines with LLMs a durable strategy?

It depends on the problem. LLMs are general-purpose reasoning engines, not specialized predictors trained on a specific objective function with years of labeled data. For some tasks, a well-tuned traditional model will still outperform an LLM on cost, latency, or raw accuracy once a business reaches enormous scale. But for the initial build phase, especially when speed to market matters more than marginal optimization, using an LLM to do the work a custom model would normally do removes a massive amount of upfront engineering.

The pattern Raji described, “code wins arguments,” applies here too. Instead of a long planning cycle followed by months of model development, a team can stand up a working version of the product almost immediately and then refine it. That doesn’t mean the custom ML layer disappears forever. It likely means it gets added later, once the LLM-built version has proven the product and the team understands exactly what needs optimizing. The ads business reaching a billion-dollar run rate is evidence that the LLM-first approach is viable at least for getting a business off the ground quickly, even if specialized models eventually get layered in for efficiency.

What does this signal for other teams building AI products?

The lesson isn’t “replace all machine learning with prompts.” It’s narrower and more useful: before investing in a custom ML pipeline, it’s worth checking whether a frontier model can already do the job well enough to launch. OpenAI’s own ads team apparently made that bet and it paid off fast enough to reach a billion-dollar run rate in under a year.

The tradeoff is that speed exposes new bottlenecks elsewhere. OpenAI’s engineering organization hit review, testing, and infrastructure limits once coding got cheap. Any team that uses LLMs to skip traditional build phases should expect the constraint to move, not disappear. The next bottleneck might be infrastructure, it might be quality control, or it might be organizational: who signs off on what an LLM-built system is doing before it touches real users and real money.

Frequently Asked Questions

How did OpenAI’s ads team avoid building traditional ML pipelines?

Based on Raji’s account, the team used large language models to perform work that ad-tech businesses traditionally handle with custom-trained machine learning models, such as relevance scoring and matching. This let them skip the usual cycle of data collection, feature engineering, and model training.

How big is OpenAI’s ads team?

The transcript doesn’t give an exact headcount, but the framing throughout is that it’s a small team relative to the scale of revenue achieved, which is presented as the notable part of the story.

Does this mean LLMs are replacing machine learning engineers at OpenAI?

Build enterprise applications with AI that are
Secure.Role: Adjuster✕ Export PHI
Regulation-compliant.✓ HIPAA✓ SOC 2
Auditable.09:14M. Chen approved claim #4471
Scalable.12,480 users online

Not entirely. The interview suggests LLMs are replacing specific pipeline-building work during the early build phase of a product, while engineering teams are still needed for infrastructure, review, and scaling problems that emerge once the product is live.

Why did OpenAI need to replace Git?

Git was designed for developers who work offline and periodically sync changes. OpenAI’s coding agents operate continuously and check code in and out of the mono repo at a volume and frequency Git wasn’t built to support, so OpenAI built an internal system better suited to always-on agents.

What does the 17x token usage figure actually measure?

Raji described it as a proxy for how broadly and deeply engineers are using AI coding tools, not a direct measure of output or quality. It reflects growth in AI-assisted coding activity at frontier AI labs over the past year, compared to roughly 2x growth at typical software companies.

Editorial standards

Presented by MindStudio

No spam. Unsubscribe anytime.