OpenAI CFO Sarah Friar: IPO, AI Rivalries, New Device, and Spending $100B+ on Compute

Primary speaker: Sarah Friar, identified in the transcript as OpenAI’s CFO.

Other speakers: Four interviewers participate. The transcript explicitly references Jason, David, Chamath, and Freeberg/Friedberg, but the caption export does not reliably label speaker turns. Where identity is clear from direct address, it is noted. Otherwise this record uses Interviewer rather than guessing.

Source context: Long-form live/podcast-style interview focused on OpenAI’s fundraising and possible IPO, competition with Anthropic, consumer and enterprise strategy, compute scarcity, data-center economics, long-term capital allocation, the Jony Ive hardware project, AI infrastructure, memory/context, and advertising.

Transcript quality: The source is an automated-caption-style export with extensive duplicated and overlapping lines, occasional merged speaker turns, false starts, and apparent speech-recognition errors. The duplication has been removed without removing substantive repetition. Several names, figures, model designations, and phrases remain questionable and are listed in Section D rather than silently corrected.

Verification status: This document records what the participants said. It does not independently verify their claims, numbers, announcements, competitive assertions, or predictions.

An opening interviewer characterizes OpenAI’s latest financing as what he regards as the most successful fundraising round in history.

Friar says OpenAI is going to raise north of $120 billion and presents the financing in the context of a much larger belief about AI. She describes AI as an unusually important economic era and says the industry is only beginning to understand what AI could mean for global productivity, potentially leading to greater affluence and better lives.

She also gives a motivational formulation about preparation meeting opportunity and the need to seize that opportunity, although the exact sentence is slightly garbled in the transcript.

On joining the discussion, Friar jokes that she is a “long time listener, first time caller” and says she is excited to be with the group.

An interviewer says he had seen an article, possibly in The Wall Street Journal, arguing that AI companies may benefit from going public earlier. He says SpaceX is expected to go public and asks when OpenAI and Anthropic might follow, and whether there is now a race to reach the public markets first.

Friar rejects the idea that an IPO should be treated as the objective of running the company.

Her framework is:

• an IPO is a milestone, not a destination;
• it is fundamentally another mechanism for raising capital;
• management should not organize the company around the IPO itself;
• her role as CFO is to maximize the company’s financial optionality.

She says OpenAI had raised $122 billion in March, describing the purpose as giving the company maximum flexibility. She frames that flexibility as important not just for OpenAI but for the extraordinary technological period in which it is operating.

Asked whether that financing was larger than any previous private or public raise, Friar says yes.

She says it was larger “by orders of magnitude” and cites Saudi Aramco’s roughly $30 billion IPO as what she believes had been the largest IPO to date.

Friar says it is remarkable that there could potentially be three IPOs of enormous scale. She compares the situation with the 2000–2001 market period but notes that capital markets themselves are now larger.

She adds another market consideration: buybacks, M&A, and other mechanisms continually return substantial amounts of cash to shareholders, leaving significant capital available to redeploy.

Her conclusion is that being first to the IPO market should not be confused with ultimately winning.

She compares this with Google versus Yahoo and Lyft versus Uber: public attention may focus on who goes first, but durable value is what ultimately matters.

She says the press likes drama, while companies will ultimately be judged on whether they build large, sustainable, durable businesses. Fundraising is one necessary component of doing that.

During the interview, a host announces what he calls breaking news: Anthropic has confidentially filed an S-1.

He asks whether that makes OpenAI “third place” in the IPO-filing race.

Friar says it means very little at that point because a company still has to go through the SEC process, whose timing can be uncertain.

The discussion then shifts from IPO timing to competitive performance.

An interviewer asserts that Anthropic had previously been far behind OpenAI but had now, in his assessment, “blown past” OpenAI among developers and corporations and apparently in revenue.

He asks how OpenAI lost that lead.

This characterization is the interviewer’s claim, not an independently established fact and not something Friar explicitly accepts.

Friar answers by arguing that OpenAI is pursuing a different strategic model rather than directly accepting the interviewer’s competitive framing.

Friar describes OpenAI as building an underlying AI/intelligence layer or infrastructure foundation with many interfaces into the world.

She identifies several interfaces:

• ChatGPT for consumers;
• Codex;
• Frontier, described in the transcript as OpenAI’s enterprise offering;
• other channels capable of reaching businesses of different sizes.

She claims that more than 900 million people use ChatGPT weekly and says ChatGPT has effectively become both a noun and a verb and is many people’s first experience of AI.

Friar says OpenAI’s economic research team had shown her that its fastest-growing geography was Africa, which she notes was growing from a relatively small base.

She says the fastest-growing languages included Azerbaijani and Kazakh.

Friar says Codex had reached 5 million users “over the weekend,” after being close to zero in January.

She argues that OpenAI’s multi-interface strategy creates a compounding advantage because the services are ultimately supported by one underlying system:

1. more users generate more data and opportunities for personalization;
2. ChatGPT acts as a front door;
3. larger or improved models generate greater efficiency;
4. efficiency should reduce the cost of producing tokens;
5. falling token costs should improve gross margins;
6. better margins create more money to purchase compute;
7. access to compute is itself a major competitive advantage.

She says each AI company must run its own race while recognizing that the industry is also an ecosystem that must bring the broader public along.

An interviewer asks whether OpenAI spread itself too thin across projects, including Sora, hardware, and other initiatives, at the expense of enterprise execution.

Friar rejects the premise.

She says commentators repeatedly force businesses into false binaries — consumer company or enterprise company — while OpenAI intends to be both.

She says its revenue was becoming approximately 50/50 between consumer and enterprise.

As evidence of enterprise focus, she describes her own recent customer activity:

• meeting Thermo Fisher in Boston;
• meeting banks in New York;
• speaking with Travelers;
• speaking that morning with another technology company.

Her point is that enterprise AI adoption is occurring across industries rather than being confined to one vertical.

She praises Denise Dresser, described as OpenAI’s new head of revenue and in the position since December, calling her a “force of nature.”

Friar says the enterprise operation is “firing on all cylinders.”

At the same time, she says OpenAI does not want to abandon consumers.

She ties that choice to OpenAI’s stated mission of AGI benefiting humanity broadly rather than benefiting only people or organizations able to pay.

Friar says OpenAI intentionally gives users substantial free access because it wants people to experience intelligence for themselves.

Her argument is that once people understand what the system can do, usage grows sharply.

She describes what she calls a commitment curve:

• free users: roughly seven questions/turns per day;
• an initial paid tier: about 15, approximately double;
• Plus at $20: approximately 3× free-user usage;
• Pro: approximately 11× free-user usage.

The wording around the exact tier structure is somewhat messy in the transcript and should be verified.

Friar compares this adoption path with the evolution of mobile phones.

People initially used a flip phone simply to make calls. Modern smartphones gradually became indispensable for many different functions.

She believes AI/intelligence is following an analogous path: users initially see a limited use case, then continually discover additional ways to employ it.

An interviewer recalls a framework he attributes to Friar from roughly 18 months earlier: converting gigawatts of compute into revenue.

He says he recalls her arguing that roughly 1 gigawatt could correspond to about $10 billion per year in OpenAI revenue.

He suggests that a similar relationship might apply to Anthropic, Gemini, and other leading AI providers.

He notes that OpenAI had aggressively sought power, data centers, and powered land while others thought the scale seemed excessive, but says that strategy now appears more understandable given the shortage of AI infrastructure.

The $10 billion-per-gigawatt ratio is stated by the interviewer as his recollection. Friar does not clearly reconfirm that exact ratio in the transcript.

Friar says compute is currently an extremely scarce resource.

She describes OpenAI as facing a “vertical wall of demand” with insufficient tokens available to serve all potential usage.

She says she is grateful that she, Greg, and Sam pushed aggressively to secure compute earlier despite criticism at the time.

Her retrospective assessment is that the decision was correct because OpenAI still would not have enough compute in 2026.

Friar says bottlenecks exist throughout the compute supply chain and will move from one point to another over time.

She identifies constraints including:

• energy;
• land;
• power availability;
• regulatory environments and permitting needed for rapid construction;
• racks;
• chips;
• memory;
• skilled talent;
• education and science capacity;
• social and community trust.

As a Stanford trustee, she says she worries about whether the education system is producing enough people with the necessary capabilities and argues for sustained focus on science and education.

Friar explicitly treats trust as another component of the AI supply chain.

She says Sam was in Saline, Michigan, at that moment and expected to participate in a ceremony roughly two hours later involving a 1-gigawatt data center that she describes as part of an Oracle complex.

Friar argues that companies building infrastructure cannot simply tell communities from the top down what those communities need.

Drawing on her seven years at Nextdoor, she says local communities will insist on defining their own priorities.

For the Michigan project, she says OpenAI and its partners were telling the community:

1. local electricity customers would not bear the project's electricity-infrastructure cost;
2. the project would pay for its own power and infrastructure;
3. it would create approximately 2,500 union jobs, including electricians and HVAC work;
4. the project would generate approximately $1 billion in taxes for Michigan;
5. approximately $45 million would be invested in education and Codex credits.

These are company claims in the interview and should be independently checked before publication.

While discussing the education investment, Friar makes a broader workforce argument.

A short phrase immediately before the analogy is garbled — “anyone who's not like coming in fossil to their new job” — and should be treated as [transcript unclear].

Her clear point is that AI tools are becoming basic professional infrastructure.

She says her teenagers use Codex.

She compares Codex literacy with spreadsheet literacy: she would not hire a finance employee who did not know how to use Excel and says she probably would not hire a finance employee today who did not know how to use a tool such as Codex.

This connects the education spending around infrastructure projects with her expectation that AI skills will become standard workplace skills.

Friar says OpenAI has to make investments ahead of demand.

That creates two simultaneous requirements:

• securing enough compute and all associated infrastructure components;
• financing those commitments before all corresponding revenue arrives.

She explicitly connects this requirement back to IPOs and fundraising: access to capital matters because the infrastructure has to be committed well before it is fully utilized.

On the revenue side, she says OpenAI’s economics continue to improve.

She believes the company is becoming better at demonstrating genuine customer value.

She also describes a shift from something resembling cost-plus pricing toward pricing more closely connected with the value delivered to customers.

Token scarcity, she says, is compressing the timetable on some of those economic decisions.

Asked about industry-wide compute availability, Friar says additional compute is essentially unavailable in 2026: if someone knows where more can be acquired, she would like to know.

She mentions Elon Musk as having had an unusual amount of compute and comments approvingly on his ability to sell some of it.

She says 2027 also remains highly constrained.

Her broader point is that the current shortage is not a brief one-year phenomenon.

Friar distinguishes between training and inference infrastructure.

She says most model training still occurs in the United States for “USG reasons,” describing it as necessary for what amounts to a national asset to remain on U.S. soil.

For inference, she says OpenAI wants a global footprint.

Her reasoning is that agentic systems increasingly need to operate in real time, making distributed inference and low latency more important.

This becomes particularly important as AI becomes multimodal.

Friar says OpenAI had to make a difficult choice around Sora/video because the company did not have enough compute.

An interviewer notes that video consumes large amounts of compute, which Friar agrees is currently true.

However, she says video is far from finished as an AI medium.

She argues that AI is moving toward a more fundamentally multimodal interaction model.

Her analogy is that the previous technology generation trained humans to “talk with our thumbs.”

She calls that habit “a disease,” describing people walking around looking downward and teenagers sitting together while messaging each other instead of speaking.

She recounts seeing her son message someone sitting next to him and telling them simply to talk.

Her conclusion is: “Multimodality is here.”

She says she personally talks to Codex every day.

Natural spoken interaction with AI systems, she argues, will require substantially more real-time compute because conversational delays would make the interaction unnatural.

Friar says OpenAI deliberately offers generous token access because it wants people to understand what AI can do.

She notes that free users may not receive the latest model but says OpenAI still wants to place useful intelligence in their hands.

She illustrates this through an autobiographical memory.

Growing up in a small community in Northern Ireland during the Troubles, she remembers Encyclopaedia Britannica arriving at the family’s front door.

She says the experience felt as though “the clouds parted.”

She wants broad AI access to provide a similar feeling of sudden access to knowledge and capability.

An interviewer asks about reports involving Jony/Johnny Ive, describing speculation about a puck and earpieces.

Friar jokes that if she confirms it is an earpiece, Jony will “come and steal” her teenage son.

She does not confirm the interviewer’s device description.

Instead, she says OpenAI is moving into a new “consumer substrate” that she cannot yet describe.

She says:

• it would be unveiled by the end of the year;
• consumers should be able to purchase it early the following year;
• she has personally seen it and tried it.

Friar says she normally talks with her hands and jokes that she is therefore sitting on them to avoid revealing anything inadvertently.

Asked whether trying it produced an iPhone-like paradigm shift, Friar says what Jony Ive and his team do especially well is “bringing humanity to devices.”

She struggles to define the quality precisely but says a person feels it when using the device.

She describes it as:

• very natural;
• very “lovable.”

Another interviewer suggests intimacy and seamlessness — not needing to take a phone out — based on what he has heard from people who tried the device.

Friar contrasts good design with technology that feels mechanistic. Great design, she says, makes the technology itself disappear.

She closes the thought with the idea that simplicity is difficult to achieve.

Asked to explain OpenAI’s capital-allocation model, Friar says high-value AI companies will ultimately follow the same fundamental economic rules as great companies in earlier eras.

They will not succeed through “magic.”

They must begin by creating customer value.

That means helping customers:

• do something differently or better;
• increase revenue;
• increase efficiency;
• or solve a materially important problem.

Her first detailed example is Thermo Fisher.

She says Thermo Fisher wants to accelerate patient screening so that FDA approvals can happen more quickly.

For a cancer patient who may have only weeks to live, she argues, reducing a process from four weeks to two can literally be a life-or-death difference.

She then says Thermo Fisher has something like 30,000 or 38,000 field employees, immediately warning that she may be misquoting the number.

Those employees, she says, can also benefit from AI productivity.

She notes that Thermo Fisher equipment is ubiquitous in laboratories and argues that AI can make the company’s commercial workforce more efficient.

Friar says the fastest recent increase in Codex adoption inside OpenAI was occurring in the go-to-market organization.

Developers were already using it, but she says the fastest month-over-month growth was coming from GTM teams.

The use case is productivity.

She also says AI is being used in finance, a domain that interests her personally.

This supports her broader argument that AI’s enterprise value extends beyond software engineers.

After customer value, Friar says a durable company needs strong gross margins.

For OpenAI, the main cost-of-revenue input is compute.

Although AI infrastructure is extremely capital intensive, she argues that the effective cost curve is also strongly deflationary.

She gives an example involving model generations, initially saying “ChatGPT 5 to 5.4” and then correcting herself to 4 to 5.4.

She says the “deprecation cost” had fallen by approximately 97% over about two years.

The term may be a transcription error and must be checked.

She characterizes the curve as extraordinary.

Friar then discusses what the transcript identifies as model 5.5, which she says had just been released.

She says OpenAI increased its stated pricing on 5.5 by approximately 2×.

Despite that higher nominal price, she says customers may still experience an effective 20–30% reduction in cost per token because each token is substantially more efficient.

Her point is that sticker price and economic cost per useful unit of intelligence are not necessarily the same.

Both the model designation and figures require verification.

Friar warns against making capital-allocation decisions based only on today’s compute costs.

Because those costs and efficiencies are changing quickly, doing so could substantially misprice future outcomes.

Management therefore has to “lean in” to an expected future cost curve rather than extrapolate the current one indefinitely.

This uncertainty becomes more important because compute is purchased years before it becomes operational.

Friar says her current focus is increasingly on what compute OpenAI can secure for 2028 onward.

She says the Saline, Michigan project likely will not begin producing compute until approximately late 2027 or early 2028.

She says the period where she feels most acutely short of future compute is already 2030, 2031, and 2032.

That means OpenAI must construct business models many years ahead of realized demand.

Friar says confidence increases each year because the actual business has been materially outperforming expectations, and capital markets are becoming more willing to support the company.

Asked how OpenAI forecasts compute needs several years ahead while models and architectures become more efficient, Friar says the company must make multiple assumptions simultaneously.

On a raw per-gigawatt basis, she says compute is actually becoming more expensive because inputs such as:

• electricity;
• memory;
• and other infrastructure components

are increasing in cost.

However, she says advances on the chip side produce more intelligence from that expensive infrastructure, more than offsetting those increases.

Thus, while a gigawatt may cost more, the customer-facing unit of useful output should become substantially cheaper.

Friar distinguishes predictable hardware improvement from less predictable model architecture.

She says OpenAI does not want to overestimate model-side efficiency.

According to the transcript:

• “5.5” was particularly efficient;
• “5.4,” the prior model, was a very large pretrained model;
• 5.4 was expensive and difficult to serve.

Sometimes OpenAI intentionally takes a large pretrained-model step even though it temporarily raises costs.

Subsequent model iterations are then used to drive serving costs back down.

The result is not a perfectly smooth efficiency curve.

For 2026 and 2027, Friar says she can build a relatively conventional bottom-up financial model.

Inputs include:

• known products;
• expected pricing;
• price × quantity;
• consumer adoption;
• number of subscribers;
• weekly active users;
• daily users;
• message volume;
• advertising.

The transcript renders one phrase as “how many wilds do I think I have?”, which is [transcript unclear] and should not be silently corrected.

She says near-term modelling can therefore be reasonably detailed.

Even so, actual demand repeatedly surprises OpenAI to the upside.

For more distant years, Friar says the modelling process partially reverses.

Instead of forecasting every future product and then determining the compute required, OpenAI looks at the compute it has secured and estimates how much revenue that capacity ought to support.

She openly acknowledges that she cannot know precisely where every future revenue stream will come from.

This is a key distinction in her planning framework:

• near term: product and customer model → revenue → compute requirements;
• outer years: committed compute → expected monetizable intelligence → approximate revenue.

Friar recounts building an investor model roughly a year earlier containing agentic revenue.

Her thesis at the time was that developers would eventually interact with agents using natural language and obtain enough productive value to pay perhaps $2,000 per month.

She says investors initially regarded the idea as implausible.

In hindsight she calls the $2,000 figure almost laughable, implying that the value trajectory developed faster or further than the model assumed.

She contrasts that skepticism with the earlier reaction to ChatGPT Pro at $200 per month, when many people had argued nobody would pay such a price.

An interviewer asks why OpenAI needs $122 billion and whether the capital is meant to support requirements through 2031 or 2032.

Another interviewer makes the question more concrete.

He says estimates he has seen put the all-in cost of building 1 gigawatt of AI compute at roughly $50 billion, including:

• land;
• power;
• data-center shell;
• chips;
• other infrastructure.

The transcript initially renders the number as “$50” before the interviewer says “all in around 50 billion,” so the intended unit appears to be billion but should still be verified.

He asks whether OpenAI has to fund the entire amount itself or can use debt and partners, and whether $100 billion therefore purchases roughly two gigawatts or significantly more.

These cost figures are the interviewer’s estimates, not Friar’s.

Friar says OpenAI’s compute strategy changed radically in only two years.

Approximately two years earlier, she describes the company as having:

• one cloud-service provider: Microsoft Azure;
• one principal chip provider: Nvidia;
• one product: ChatGPT;
• one main price point: $20 per month.

She uses a Rubik’s Cube as a metaphor.

At that stage, OpenAI occupied essentially one cube.

Now, the strategy contains many more dimensions and possible combinations.

Friar says OpenAI deliberately diversified across cloud providers.

She names:

• Oracle;
• CoreWeave;
• Microsoft;
• Google Cloud/GCP;
• AWS;
• smaller “neoscalers.”

Her financial reasoning is that cloud-service providers can effectively convert some of OpenAI’s capital expenditure into operating expenditure.

Rather than OpenAI funding every data center upfront, the provider finances and builds capacity, while OpenAI pays as it uses that infrastructure and earns revenue from it.

This lets OpenAI benefit from the providers’ own balance sheets, financing ability, and construction capacity.

Friar says OpenAI also intends to avoid dependence on a single processor architecture.

Her reasoning is that technological leadership can leapfrog between chip suppliers, so a company locked to one supplier may periodically fall off the frontier.

She nevertheless calls Nvidia OpenAI’s “absolute priority partner” and says Nvidia has the frontier chip.

The transcript says OpenAI’s next major “trading run” in the fall would use “Vera Rubins.” This likely contains transcription errors and should be verified rather than corrected.

Friar then mentions a forthcoming “Simon series”, another term requiring verification.

She says OpenAI also has:

• AMD chips in the pipeline;
• Cerebras already online;
• an internal chip project with Broadcom.

She describes Cerebras as particularly good for low-latency workloads such as real-time coding.

Beyond CSP capacity, Friar says OpenAI is beginning to use more built-to-suit infrastructure.

She identifies a data center being developed with SoftBank Energy in Texas.

She says this represents a step beyond relying entirely on CSPs and therefore requires somewhat more direct capital expenditure.

Her larger strategy is not to choose a single financing structure but to maintain many possible configurations.

Friar returns to the Rubik’s Cube metaphor, saying the many configurations represent optionality.

She asks listeners to check her number but says a Rubik’s Cube has something like a quintillion possible arrangements.

The specific mathematical claim should be verified.

She reiterates that her job is to maximize optionality.

Friar also says OpenAI is not yet an investment-grade type of entity able to obtain the cheapest debt financing directly.

Therefore, partnerships with investment-grade infrastructure providers are important because those partners can obtain capital more cheaply and build on OpenAI’s behalf.

An interviewer asks whether the traditional technology stack will effectively merge over the next five years.

Historically:

• Nvidia primarily sold chips;
• Microsoft primarily operated cloud infrastructure;
• consumer applications occupied a separate layer.

The interviewer observes that leading AI companies increasingly operate across several layers:

• OpenAI is developing silicon as well as models and consumer products;
• Nvidia offers silicon and open-source models and is described as increasingly becoming an off-taker;
• Google operates cloud infrastructure, designs chips, and develops models.

He asks whether this convergence simplifies or complicates competition.

The interviewer then offers a strategic hypothesis: every company wants to occupy the layer closest to the customer, because that layer often captures the largest share of ecosystem profits.

Friar agrees.

Friar says this convergence reinforces OpenAI’s desire to occupy the AI intelligence layer.

She notes that roughly a year earlier many people predicted that LLMs would quickly become commodities.

In her view, the opposite has occurred.

The reason is not simply the base model.

As agentic systems emerge, a “harness” adds:

• context;
• memory;
• personalization;
• organization-specific knowledge.

Those layers deepen the value of the underlying model rather than making all models interchangeable.

Friar says her own Codex contains a large memory file that knows:

• who she is;
• that she is OpenAI’s CFO;
• how she likes to write or speak;
• what topics interest her;
• that she is a mother of teenagers.

She says this persistent memory makes the system much more useful to her personally.

Her argument is that the same principle becomes much more powerful inside a company when combined with enterprise information and organizational context.

Friar argues that enterprise value is not just about making structured corporate data searchable.

Organizations contain intuition — knowledge about relationships, incentives, circumstances, and unwritten context that may not appear in conventional databases.

She illustrates the idea through her previous Wall Street experience.

After an earnings call, all quantitative data might indicate that a stock should rise.

But a trader might say the stock will not rise because he knows a particular fund is under pressure and must sell positions, depressing the stock for the next week.

The data alone does not contain that situational knowledge.

Friar describes that as the “intuition of an enterprise.”

She believes AI models are increasingly able to connect company data with memory, context, and this kind of organizational intuition.

That, she says, is what excites CEOs and senior executives: they can begin to see AI as both:

• a top-line/revenue-growth tool;
• an efficiency tool.

Her strategic conclusion is that AI providers want to remain as close as possible to the layer where that customer value is created.

Jason asks the final major question about advertising.

He notes that several historically dominant consumer businesses have relied heavily on advertising and asks whether ads are ultimately how OpenAI can make AI broadly accessible for free.

He also references Anthropic having criticized or mocked OpenAI over advertising during the Super Bowl.

Friar says advertising must follow a core principle:

the user should receive the best result from the model because it is the best result, not because someone sponsored it.

She says that principle “has to hold true.”

She also says OpenAI will always provide an ad-free tier for people who do not want advertising.

An interviewer interjects: “If they pay.”

The transcript therefore does not establish that an ad-free free tier will always exist.

Friar attributes an analogy to a name rendered by the transcript as “Fiji” [transcript unclear].

The analogy is that if Google and Meta had a baby, it would be ChatGPT.

Her reasoning is that ChatGPT potentially combines two valuable advertising signals.

From Google/Search:

• explicit, high-intent queries;
• users directly state what they are trying to find, solve, or buy.

From Meta:

• knowledge about the user and people similar to the user.

Friar argues ChatGPT may have an additional signal beyond those systems: memory and context.

A user might directly tell ChatGPT they want a particular type of shoe, while ChatGPT also knows persistent information about that person.

She says combining:

• intent;
• memory;
• context

could make ChatGPT a highly potent advertising platform.

Within the advertising discussion, Friar says OpenAI knows it has at least 11% of the search market.

She argues the practical share may be significantly higher because conventional measurement treats Google and ChatGPT interactions differently.

Her example:

• a Google page refresh may count as another search;
• a long ChatGPT conversation containing perhaps 50 questions may be counted only once.

Therefore, she argues simple query/session comparisons understate the amount of search-like activity occurring inside ChatGPT.

This is a significant quantitative claim and requires independent verification before publication.

Friar says a powerful advertising business could fund much broader access to AI.

This connects directly to her earlier argument that OpenAI wants intelligence to reach people regardless of their ability to pay.

Advertising is therefore presented not only as a revenue opportunity but as a subsidy mechanism for universal or very broad access.

Friar gives a striking comparison between API and consumer economics.

She says that if she were optimizing only for current revenue per token, she would allocate every available token to the API.

She says API economics are roughly an order of magnitude better than consumer economics on that measure.

However, OpenAI deliberately does not optimize only for the present-day revenue maximum.

Its strategic objective is larger.

Friar says OpenAI believes AI will become an infrastructure or utility layer comparable in some sense to electricity.

In that future, OpenAI wants to serve:

• individual consumers;
• small businesses;
• large enterprises;
• governments;
• ultimately the world at large.

That broad infrastructure vision is how she closes the substantive discussion.

Exact quote: “Like in the end, an IPO, I say this to the team all the time, it's a milestone. It is not a destination.”

Context: Asked whether OpenAI should race Anthropic and SpaceX to the public markets.

Topic: IPO strategy / corporate discipline.

Exact quote: “I feel like my job as a CFO is create optionality for this, not just this company, but just this era that we're living in.”

Context: Explaining the purpose of OpenAI’s fundraising.

Topic: Capital strategy / optionality.

Exact quote: “In the end, the market is a weighing machine, not a popularity machine.”

Context: Rejecting media focus on which AI company goes public first.

Topic: Long-term company value.

Exact quote: “No one remembers who won first, Google or Yahoo, Lyft or Uber.”

Context: Argument that IPO order does not determine the long-term winner.

Topic: Competition / IPO timing.

Exact quote: “The press loves a bit of drama, but in the end we're going to have to build big, sustainable, durable companies.”

Context: Distinguishing public-market narratives from business fundamentals.

Topic: Corporate durability.

Exact quote: “Over 900 million people use ChatGPT weekly, and it's become the noun and the verb.”

Context: Describing ChatGPT’s role as OpenAI’s main consumer interface.

Topic: ChatGPT scale / consumer adoption.

Verification required.

Exact quote: “ChatGPT acts as a front door.”

Context: Explaining how OpenAI’s consumer interface feeds its broader platform strategy.

Topic: Platform strategy.

Exact quote: “The reality is we're very much both. We're not one or the other.”

Context: Rejecting the choice between being a consumer company and an enterprise company.

Topic: Consumer versus enterprise.

Exact quote: “Remember our mission at OpenAI is AGI for the benefit of humanity, not for the benefit of humanity who can pay.”

Context: Explaining why OpenAI continues offering substantial free usage.

Topic: Access / mission.

Exact quote: “Compute is a very scarce resource at the moment.”

Context: Discussion of power, data centers, chips and token supply.

Topic: Compute shortage.

Exact quote: “We're going up that kind of vertical wall of demand right now and there's just not enough tokens available.”

Context: Describing OpenAI’s demand relative to compute supply.

Topic: AI demand / infrastructure.

Exact quote: “There's kind of choke points everywhere.”

Context: Listing energy, land, power, chips, memory, talent, regulation and trust as infrastructure constraints.

Topic: AI supply chain.

Exact quote: “And then trust. I mean, I actually put that as part of the supply chain.”

Context: Transition from physical infrastructure to community acceptance.

Topic: Data centers / public trust.

Exact quote: “Thank you, but no thank you. I will tell you what I need.”

Context: Her description of how local communities respond when companies dictate priorities from above.

Topic: Community engagement.

Exact quote: “I pretty much probably wouldn't hire a finance person today that doesn't know how to use a tool like Codex.”

Context: Comparing AI literacy with knowing Excel.

Topic: Workforce / AI skills.

Exact quote: “If you want to buy more compute good luck to you.”

Context: Describing 2026 infrastructure scarcity.

Topic: Compute availability.

Exact quote: “Talk. Multimodality is here.”

Context: Criticizing smartphone-era thumb-based communication and predicting more natural voice/multimodal interfaces.

Topic: Multimodal AI.

Exact quote: “We're changing into a consumer substrate that I cannot tell you what it is, but by the end of this year, we will unveil it.”

Context: Asked about the OpenAI/Jony Ive hardware project.

Topic: AI hardware / consumer device.

Exact quote: “What Johnny and team are really good at is bringing humanity to devices.”

Context: Describing her experience of the unreleased device.

Topic: Jony Ive / product design.

Exact quote: “It feels very natural, but it feels very lovable.”

Context: Her description of the unreleased consumer device.

Topic: AI hardware / product experience.

Exact quote: “Customer value first.”

Context: Explaining OpenAI’s capital-allocation philosophy.

Topic: Business fundamentals.

Exact quote: “The main input is compute.”

Context: Explaining cost of revenue and gross margins.

Topic: AI economics.

Exact quote: “The shape of the line keeps taking us by surprise to the upside.”

Context: Explaining why even near-term demand forecasts keep being exceeded.

Topic: Revenue forecasting / demand.

Exact quote: “I don't know for certain exactly where it will all come from.”

Context: Discussing outer-year revenue forecasts based on committed compute.

Topic: Forecast uncertainty.

Exact quote: “My job is maximum optionality.”

Context: Rubik’s Cube analogy for cloud, chip and financing diversification.

Topic: CFO strategy.

Exact quote: “A year ago people talked about the commoditization of the LLMs. Um and frankly, it's gone the opposite.”

Context: Explaining why memory, context and agent harnesses make the intelligence layer strategically valuable.

Topic: LLM commoditization / agents.

Exact quote: “That is the intuition of an enterprise.”

Context: Wall Street example of a trader knowing information not visible in numerical data.

Topic: Enterprise AI / institutional knowledge.

Exact quote: “We want to make sure that you know you're always getting the best result based on the model, not by something that was sponsored.”

Context: Stating the principle governing advertising inside ChatGPT.

Topic: Advertising / answer integrity.

Exact quote: “If, you know, Google and Meta had a baby, it would be ChatGPT.”

Context: Describing the potential combination of search intent, user information, context and memory.

Topic: Advertising model.

Exact quote: “Imagine putting memory and context next to intent, you should have a very potent ad platform.”

Context: Explaining ChatGPT’s potential advertising economics.

Topic: Ads / personalization.

Exact quote: “If I was optimizing only for today, I would give every token to the API.”

Context: Comparing API revenue per token with consumer economics.

Topic: Monetization / resource allocation.

Exact quote: “Every token to the API. Order of magnitude more than to the consumer.”

Context: Quantifying the economic difference while explaining why OpenAI nevertheless continues serving consumers.

Topic: API versus consumer economics.