What enterprise strategy and business performance did OpenAI CFO Sarah Friar describe at Goldman Sachs’s Communacopia + Technology Conference, including the 32% month-over-month growth in enterprise revenue versus 20% overall annualized-revenue growth, the roughly equal balance o

Friar’s message was that OpenAI is becoming an enterprise-led, vertically specialized AI platform while preserving a large consumer business—and that lower inference costs are central to widening adoption. The most clearly reported Codex figure is 25 million users. 1

• Enterprise momentum: Friar cited enterprise revenue growing 32% month over month, faster than the company’s 20% growth in overall annualized revenue. Enterprise revenue had reached rough parity with—and reportedly then surpassed—the consumer business ahead of OpenAI’s prior schedule. 56

• Cost-per-task strategy: She argued customers should assess the cost of completing a task, not merely token price: a stronger model may cost more per token but cost less overall if it requires fewer attempts. OpenAI cut GPT-5.6 Luna’s price by 80%; its official pricing lists $0.20 per million input tokens and $1.20 per million output tokens. 71

  • Friar said the lower price drove a roughly 10× usage increase and maintained that Luna can be cheaper to run via cloud providers than ostensibly cheaper open-source models such as Z.ai’s GLM-5.3, once reliability and total deployment costs are included. That is an OpenAI comparative claim, not an independently established universal cost result. 1

• Vertical expansion: The strategy extends beyond general chat and coding into chip design, life sciences, and financial services, where OpenAI is aiming to sell systems that perform high-value professional workflows rather than only generic model access. 1

• Custom silicon: The reported Jalapeño program with Broadcom is an inference-focused custom chip. Reporting around the project says it reached manufacturing tape-out in about nine months and was claimed to reduce inference cost by roughly 50% versus typical AI GPUs; those savings are reported/vendor claims rather than independently audited benchmarks. 23

• Codex: Codex, OpenAI’s coding product, had 25 million users according to the Reuters conference report. 1

• Advertising ambition: The proposed ChatGPT advertising model was framed as combining Google-style high-intent queries with Meta-style personalized context. User memory would be a key differentiator because it can make recommendations and ads more contextually relevant; Fidji Simo, a former Meta executive, was associated with developing that consumer-and-ads opportunity.

  • Separately, Reuters reported that ChatGPT Ads reached a $1 billion annualized revenue run rate after roughly seven months of operation. An annualized run rate is a projection based on the current pace, not revenue already booked over a full year. 3

The main caveat is that several operational figures—especially comparative deployment costs, the 10× usage effect, and Jalapeño’s 50% savings—are management or vendor claims, so they should be treated as strategic assertions rather than independently verified performance results.