OpenAI's advertising business reached $1 billion in seven months, which Sarah Friar calls the fastest-growing advertising platform ever and the company's fastest-growing product.
The week's story about OpenAI is that it backed a call to slow down frontier development and ruled out an initial public offering. Friar's account of the business underneath that is a company with four revenue lines, a new advertising product and $122 billion raised in a single quarter.
"And in the end, I'm a CFO, so it has to come back to an ROI."
Friar is OpenAI's chief financial officer, began her career writing equity research on security software, and says the company took a quarter of its engineers off production work three weeks ago to patch its own code.
The full interview is covered here so you can skip it. 24 minutes of audio, 18 minutes of reading.
Here are the 13 arguments that matter.
👤 Guest: Sarah Friar, CFO of OpenAI, who began her career writing equity research on security software
🎙️ Host: Jim Cramer, host of Mad Money and manager of the CNBC Investing Club's charitable trust
🧩 Other segments: Jensen Huang of Nvidia and Marc Benioff of Salesforce
📰 Published: 15 September 2026 on the Mad Money podcast (CNBC)
🟣 Apple Podcasts | ⏱️ 24 min | ✅ Time saved: 6 min
Key Takeaways
The advertising business hit $1 billion in seven months, which she calls the fastest-growing ad platform ever
About 90% of ChatGPT's consumers use it free, so advertising is what carries the margin on them
OpenAI took 25% of its engineers off production work to patch its own code alongside the model
It pointed its frontier model at its own software three weeks ago and fixed what the model found
A credit check that cost $200 now costs about $0.17 inside OpenAI's own finance team
The run rate last year was 2,800 checks, and she says the outcome is more work at the edge, not fewer people
She says she is not a doomer, and will still slow the frontier if the researchers say so
"But as the CFO, I need to then make business decisions around it"
Cutting the price of the cheapest model by 80% took demand up tenfold
The frontier model is what trains the small ones, which is her argument for buying more compute
Customers looking at Chinese open-weight models are asking about cost, tunability and data sovereignty
She says OpenAI's cheapest model already undercuts GLM 3.5 on Cloudflare
The supply-chain strategy is deliberate diversification across Nvidia, AMD, Broadcom and Cerebras
Her stated worry is what happens if TSMC cannot provide all the wafers
330 million people a week ask ChatGPT a question about their health
1. Pacing the Frontier
Cramer set up the segment by noting that over the weekend OpenAI and Anthropic, the two leading frontier laboratories, both backed calls to slow the pace of development on the most powerful models so that safety protocols can catch up — and that a deliberate slowdown might save OpenAI money while hurting the data-center buildout.
His opening question was blunt: was she scared of dying in four years from a rogue actor manipulating agents to misalign weapons of mass destruction
Her answer was no, with a condition attached. "No. I'm not. However, we do need to take safety seriously."
She accepted the slowdown as a live option rather than a hypothetical. If it means pacing the frontier and slowing down, she said, absolutely — "We're going to listen to our researchers and do that. But as the CFO, I need to then make business decisions around it."
Her description of where the company is now is that it is already doing this, pacing the frontier and taking safety and alignment seriously
Cramer noted that she had said an initial public offering is ill-advised now, and asked what she would cut back on if the company wanted to pace itself
2. The Intelligence We Have
Her answer to the spending question was that the existing capability is already enough to matter. "So let's talk about that pacing because actually what I'm excited about is even if we stop today, the amount of intelligence that's available in the world is massive."
The model she named is Astra, which she described as the frontier, most intelligent and most aligned model, and the fastest selling the company has seen
Her first example was clinical. She said Boston Children's Hospital had 40 undiagnosed, unresolved diseases that clinicians there have now diagnosed. "If you're a parent of a child with a disease like that, that is life changing."
Her second was an investment one. An asset manager, she said, has a central-bank analyst tool that can take all the inputs — she named Kevin Warsh's as an example — and go from a two-hour synthesis down to minutes
On what she would actually cut, she said the decision is always a return question, and that from where she sits there is so much opportunity to drive growth that she remains focused on getting more compute to keep the flywheel going
Cramer's response was to name more hospitals he wants on it — Brigham and Women's, Dana-Farber
3. Big Models Train Small Ones
Her argument for buying compute is that the frontier model is an input to the cheap ones. With more compute the company builds the best models, and the best model is able to train what she called the child models — the mini and the nano
The worked example is a price cut that raised revenue through volume. "So what we saw with Saul, it trained Luna, which is the cheapest model. We dropped the pricing there 80% and we took demand up ten x." She said Luna is the number one model on OpenRouter right now
Her explanation of why that works is a segmentation argument. "It's because what customers really want is the right intelligence for the right task at the right price." The lineup sits on what she called a Pareto frontier
The loop she described runs in both directions: invest in compute to create frontier models, to create the best products, to create the cash flow that pays for more compute
4. Ads Hit $1B in 7 Months
Cramer asked whether, given $10 billion, it is possible to make $30 billion per gigawatt.
Her answer was one word plus a qualification. "Theoretically, yes it is."
The reason she gave is the mix. OpenAI is a diversified revenue stream running from consumer through small business up to the biggest businesses and governments around the world, and those businesses carry very good gross margins, so the return on those gigawatts is high
The counterweight she named is the mission, which is not a paying customer. The stated mission is AGI for the benefit of all of humanity, not just people who pay — "And so we have about 90% of our consumers get ChatGPT for free."
Advertising is how that free base becomes margin. "So our ad business just hit $1 billion in seven months, the fastest growing ad platform ever, the fastest growing product for us."
She framed it as a margin story rather than a revenue one, on the grounds that high-margin add-ons are being brought into the overall business model
5. The Defender Moment
Cramer put the week's politics to her directly: the President called fears of AI taking over a hoax, Jensen Huang said the doom stories are made up, and a panel suggested Sam Altman and Dario Amodei have an ulterior motive. He asked her to shoot it down.
She declined the doomer label and kept the safety point. "So I would say, first of all, I'm a tech optimist, so I'm not a doomer." It is also important, she said, to align around safety and take it at the right pace for the frontier
The topic she wanted instead was cyber security. "We have a moment. We're calling it the defender moment."
The benchmark claim is the evidence she offered. "These models, Astra scored 100% on the cyber bench, which is effectively the benchmark that says it can find just about every vulnerability in lines of source code."
The instruction that follows is operational: take the model, point it at all of your software, and it will find the vulnerabilities
Finding them is only half of it. She said you then have to patch at machine speed, because you do not have the luxury of keeping a list of things that need fixing, and that Codex, the company's coding tool, is what does that
Her summary of the Hugging Face incident was a call to action. "Today, Hugging Face should be an absolute clarion cry." She added that the defenders need access to these tools
Her position on the argument as a whole is a middle one. "Keep the optimism, which Jensen has, but also be pragmatic about the fact that there are real risks here."
Asked whether misaligned agents can be trained out of unauthorized communication, or their message boards taken down, her answer was alignment rather than demolition. "You need alignment." Guardrails go around the models as they get more intelligent, so they do what the human in the loop wants
Friar noted that she started her career writing equity research on security, back when viruses were first appearing; Cramer said she must have been about 12
6. The Health Desert
Asked to assure viewers that the road ahead is not a doomsday, she went to a small-business event she had attended in Texas that week.
The person she described is a qualified doctor who started in Alaska and moved to west Texas, and who talks about health deserts
The number attached to it: "What she sees is that if you live in a health desert, your life expectancy is 30% less than someone who's got active access to health care."
The demand side is already there. She said 330 million people every week ask ChatGPT a question about their health
The supply side is the product. There is now a clinician tool so doctors can use AI quickly, and OpenAI works with hospital systems including HCA, Boston Children's and Stanford Children's
What the doctor is building on top of it is a business: telemedicine into places like west Texas
Her framing of the risk question. "There have always been risks to technology. We should take them seriously, but we shouldn't allow it to stop the progress that we're making."
Cramer's reading was that this creates west Texas jobs that would not otherwise exist — more jobs, not fewer
7. What Finance People Ask
Friar said the company's own economic research team has looked at the effect on the wider economy. "What we see is that people are actually starting to make their jobs more interesting and broader."
The data she cited is what people with a given job title actually ask for. "If you look at all of the calls into ChatGPT from people whose title is finance, only about a quarter of what they asked ChatGPT is finance" — another quarter is engineering
Her example of what that looks like inside her own team is tax staff learning to program so they can do work themselves instead of waiting for an IT person
The third slice surprised her. She said the same finance population spends 25% of its time asking marketing questions, and that she is less sure what is going on there
Her rationalization of it is a line she uses on her own team: she tells her finance team that "we're all in sales, whether you know it or not" — you may not be selling to an external customer, but you have to sell your point of view internally
The shift she describes is from reporting to acting: going from being a historian of the business to actually moving the business forward
8. Why Customers Look East
Cramer compared the fear of falling behind China to the fear of falling behind Russia between 1959 and 1963, which he called a false construct, and asked whether she believes this one.
She accepted the competitive framing. "I absolutely believe that we're in a very competitive landscape right now." Customers do come and talk about open source and open weights, and many of those models are Chinese
The first reason customers give is cost, which she treats as OpenAI's problem to solve. The answers she listed are frontier models training mini models, and OpenAI's own chip, which lets it run inference at a much cheaper price
Her claim is that the price gap has already closed. "The fact that we can be cheaper today, if you buy Luna, it's cheaper than taking a Chinese model like GLM 3.5 on Cloudflare."
The second reason is control. Customers want the ability to go deeper, to take off some of the guardrails and tune the model to their own needs — which is where she says verticalization gets interesting. OpenAI went vertical in chip design because it had to do it for itself, and now does that work with customers, including in healthcare
The third is jurisdictional. "The third thing is data sovereignty." Where is my data, is it safe, is my intellectual property safe — questions she says have existed since technology started, and that she does not think Chinese models answer well either
She said customers outside the United States need to be heard on it, and that she was flying to the UK that night, where people worry about losing access to a model or to their data
9. 25% of Engineers on Cyber
Cramer said the Hugging Face write-up reads as a blueprint for attackers, and asked whether swarms could shut down a utility or a chemical plant that has not done its cyber security.
Her answer put a clock on it. "I mean, this goes back to why we think the defender window is closing and why there is a limit of time, frankly, to really harden our infrastructure."
OpenAI ran the exercise on itself. "At OpenAI three weeks ago, we took Astra, pointed it at ourselves." The first thing she says they learned was massive empathy for customers
The second was a staffing decision. "Number two, we took 25% of all of our engineers off of production level tasks and said, your job now is to run alongside the model and fix as it finds"
The third was product. The company has built tooling it can take out into the world to help customers protect themselves from things like an agentic swarm
The fourth was documentation: playbooks
Her framing of why this is not a solo effort is a proverb. "If you want to go fast, go alone. If you want to go far, go together."
The collective action already taken is a public letter. "We put out the letter. Over 100 other companies signed on, and we think it's really important." The work has to be done as a larger cyber community
10. Agencies, Not a Pause
Cramer asked whether AI needs an international agency of the kind created for atomic weapons, noting that it did not stop anyone building H-bombs if they wanted to.
Her answer was that usable agencies already exist. She named a British evaluation body and an American one as two groups that are good at looking at models before they launch
She credited a recent episode at Anthropic with raising the alarm usefully — she said it created a blast out in the world telling everyone to pay attention
What she rejected in it was the remedy. It suggested, in her reading, that the way to control the issue was to keep the technology in a box and not let people see it or use it
Her counter is the company's own name. "We are OpenAI because we are more open in our approach. We think security and safety comes from a broader sharing in the world so that more eyes are on things." More eyes, in her view, produce a better result
The role she leaves for the agencies is trust. OpenAI looks to them to help it do this well and in a trusted way, so that people believe the outcome is one the world wants
11. $122B and No IPO
Cramer said that when people hear no IPO they assume it is terrible for Broadcom, and that he thinks the opposite — that OpenAI is making so much money its recurring revenue would compare favorably with Anthropic's and with the hyperscalers'.
She would not be drawn on the comparison but liked the position. She said she can build a really strong, durable business
On the listing itself she repeated a line she has used with Cramer before. "As you know, with IPO, I've been on many a show with you where I've said an IPO is just a milestone in the journey."
The financing that removes the urgency is the number she gave. "Look, we raised 122 billion in Q1 of this year to give ourselves maximum flexibility."
Asked how much of it has gone, her answer was that it has not. "We still have an incredible balance sheet, actually, of just cold, hard cash sitting there" — available to fund partners as OpenAI buys product from them
The revenue description she gave is the same four-part one. "It's a very diversified set of revenue streams. Consumer, small business, large enterprise, very good margins", particularly as products are built on top — cyber, Codex and ChatGPT at work
The date she pointed at for more is Dev Day on 29 September, where she said there will be a lot of new announcements. The categories she named were healthcare and personal finance in ChatGPT, and verticalization or an office-of-the-CFO product in the enterprise
Her own example of the enterprise conversation was a call that morning with the chief financial officer of a well-known car company, which she declined to name, about zero-day closes
12. $200 to $0.17
Cramer said he is not hearing people say this is how they cut 20% of the workforce.
She did not deny the efficiency. "It's. I mean, efficiency is key." There is no doubt, she said, that things OpenAI does with its own products inside finance limit the number of people it needs
Her qualification is which jobs those are — in many cases, ones she does not think of as great jobs
The worked example is procurement. OpenAI's run rate last year was 2,800 credit checks. "The credit check is a pretty mundane piece of work." Does a junior analyst want to be doing it, she asked, or can an intelligent model do it
The cost comparison is the number that carries the section. "And the cost difference, it went from $200 a check to about $0.17." "That is a great ROI for a CFO"
The other end of the range is work she wants models on for the opposite reason. Technical accounting is complex, and going to the Securities and Exchange Commission for preclearance means writing intense memos, which she called a great place to bring raw intelligence to bear
Her conclusion is a reallocation rather than a cut. "So it's not fewer people. I think it's actually enabling people to do more work at the edge, more work where intelligence is needed", with the rote work done by the model at better scale and, she said, with more accuracy
13. Diversifying the Chips
Cramer asked how the decisions get made — the AMD deal with warrants, a $100 billion Nvidia deal he said appears to have been scaled back to $30 billion, and the Broadcom work — and whether it is simply whoever is best at that moment.
Her framing is the same one she used for revenue. "And that is just good CFO risk mitigation." Just as the company has a strategy for diversifying revenue streams, it has one for diversifying the supply chain
The first reason is availability. You never know when someone's supply chain is going to get gummed up, so you need multiple providers
The second is fit, and she assigned each supplier a job. "Nvidia is still an incredible platform of accelerators for training." She said Astra used 100,000 GPUs in Stargate Texas
Broadcom's chip is for the other half of the workload. She described the Jalapeño part as very focused on inferencing because it is set up exactly for OpenAI's models, and therefore very efficient, and thanked Hock Tan by name
AMD's MI355 has become a viable alternative in many cases, and the company is also working with partners including Cerebras
What diversification buys, on her account, is customer outcomes and negotiating position: lower latency, better reliability, better pricing, and leverage with the supply chain
The tail risk she named is the foundry. "What happens if TSMC can't provide all the wafers?"
Her closing rule for all of it: "And in the end, I'm a CFO, so it has to come back to an ROI." The landscape is dynamic enough that judgments about where and when to invest shift with where the demand is coming from
Bonus Insights
Cramer's framing of the segment's stakes was that OpenAI has been spending fortunes on the data-center buildout, so a slowdown in frontier development might save the company money while hurting the buildout — and he wanted to know which
He praised OpenAI's published papers, saying they are written in English and people can read them
Cramer's parting request was the one he always makes. He told Friar that if there is demand somewhere to make money she is in the right place, and then: come public so we can all have a piece of it
Friar's bottom line is that OpenAI will pace the frontier if its researchers ask for it, and that the business underneath does not depend on the answer: four revenue lines, a billion-dollar advertising product built in seven months, and a supply chain deliberately split across four chip suppliers.
Products, Companies & Tools Mentioned
OpenAI (Her own company — four revenue lines from consumer to government, about 90% of ChatGPT consumers on the free tier, and $122 billion raised in the first quarter)
ChatGPT (The consumer product, where 330 million people a week ask a health question and where healthcare and personal finance features are being built)
Codex (OpenAI's coding tool, which she says is what lets a company patch found vulnerabilities at machine speed)
OpenRouter (Where she says Luna, the cheapest model, is currently the number one model)
Hugging Face (The security incident she calls a clarion cry for the whole industry)
Nvidia (Still, in her words, an incredible platform of accelerators for training — she says Astra used 100,000 GPUs in Stargate Texas)
Broadcom (Maker of the Jalapeño inference chip built for OpenAI's own models, which she calls very efficient for that job)
AMD (Its MI355 has become a viable alternative in many cases, under a deal Cramer says came with warrants)
Cerebras (Named as another supply partner in the diversification strategy)
TSMC (The foundry risk she names: what happens if it cannot provide all the wafers)
Anthropic (The other frontier laboratory that backed the slowdown call, and the source of the episode she says usefully raised the alarm)
Cloudflare (Where she compares OpenAI's cheapest model against the Chinese GLM 3.5 on price)
HCA, Boston Children's Hospital and Stanford Children's (The hospital systems OpenAI works with; Boston Children's clinicians diagnosed 40 previously unresolved diseases)
Books & Resources Mentioned
The cyber letter OpenAI put out (Signed by over 100 other companies, and her example of the work being done as a larger cyber community)
OpenAI's economic research on AI and the workforce (The source of her claim that only about a quarter of what people with finance job titles ask ChatGPT is finance)
OpenAI Dev Day, 29 September (The date she points at for new announcements across healthcare, personal finance and enterprise verticalization)
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: