All-In Podcast: OpenAI CFO Reveals New Consumer Device, Compute Scarcity Through 2027, and a Coming Ad Platform That Marries Google's Intent With Meta's Targeting
Sarah Friar joins the All-In Podcast on June 2, 2026, offering the most detailed public disclosure yet of OpenAI's economics, capital strategy, and product roadmap
OpenAI CFO Sarah Friar used a lengthy appearance on the All-In Podcast to lay out the company's strategic and financial architecture in unusually granular terms — touching on a not-yet-announced consumer hardware device, a frank acknowledgment that compute supply remains critically constrained through at least 2027, and a clear-eyed articulation of why OpenAI believes it is building the most potent advertising platform the industry has ever seen. Taken together, the conversation represents one of the most substantive public disclosures the company has made ahead of what will eventually be a landmark IPO.
A New Consumer Device: "Lovable" and Coming Before Year-End
The single most market-moving disclosure in the conversation was Friar's confirmation of a new consumer hardware product designed by Jony Ive's team, set to be unveiled by year-end with a commercial launch expected in early 2026. Friar was careful not to describe the form factor in detail, but her language was deliberate and striking. "What Johnny and team are really good at is bringing humanity to devices," she said. "When you see it, you feel it. It feels natural in some way. It feels very natural, but it feels very lovable." She described it as moving away from the phone paradigm entirely — no more talking with your thumbs — and suggested it is designed around seamless, real-time multimodal interaction. She confirmed she has personally used it, and said it produced something close to a paradigm shift in her own experience. The product sits at the intersection of OpenAI's push into multimodality and its agentic computing vision, and would represent the company's first direct play for the consumer hardware market dominated by Apple.
Compute Is the Binding Constraint — and Will Remain So
Friar was blunt on the supply side: "If you want to buy more compute in 2026, good luck to you — tell me where to find it." She extended that scarcity assessment into 2027, describing the environment as "pretty limited as well, frankly." The Michigan data center in Seline — part of the Oracle complex, with Sam Altman cutting the ribbon within hours of the interview — will not produce usable compute until late 2027 or early 2028. Friar said her current capital planning focus for new compute acquisition is pointed at 2030, 2031, and 2032, reflecting lead times that are now measured in years, not quarters. She noted that OpenAI assumes per-gigawatt compute costs are actually rising, driven by power and memory price inflation, even as the intelligence value extracted per unit continues to improve sharply thanks to chip efficiency gains.
The company's response to this constraint has been a deliberate diversification of its compute stack. Two years ago, OpenAI operated with a single cloud provider (Microsoft Azure), a single chip (Nvidia), and a single product at a single price point. Today it sits across every major CSP — Oracle, CoreWeave, Microsoft, GCP, AWS, and a collection of neoclouds — and is actively multi-chip, running Cerebras for low-latency inference, AMD in the pipeline, and its own proprietary chip being developed with Broadcom. The next major training run, scheduled for the fall, will be conducted on Nvidia's Verrubin architecture. Friar used a Rubik's cube metaphor to explain the strategic intent: maximum optionality across a combinatorially large configuration space. "My job is maximum optionality," she said, "and in a moment where I'm not yet an investment-grade type of entity where I can go get lower-cost debt financing, being able to work with partners to do that is really important."
The Economics: 97% Cost Deflation, Pricing Power, and a Gross Margin Build
On unit economics, Friar disclosed that the cost per token fell approximately 97% from GPT-4 to GPT-4.5 — a deflation curve she described as occurring over roughly two years. With GPT-5, OpenAI raised prices 2x, yet customers are still seeing approximately 20 to 30% cost reduction per token given the efficiency gains in the underlying model. This spread between what OpenAI charges and what it costs to serve is the gross margin engine the company is building toward. Friar was explicit that pricing decisions made on today's cost profile would misprice future value, and that the capital allocation framework requires leaning into forward cost curves rather than current ones.
On revenue mix, she confirmed the business is now roughly 50/50 consumer versus enterprise, a balance she described as intentional rather than accidental. The consumer side is anchored by 900 million weekly ChatGPT users, with engagement scaling sharply up the subscription tiers: free users average seven interactions per day, the first paid tier reaches roughly fifteen, the $20 Plus tier hits approximately three times free, and Pro users generate around eleven times the engagement of free users. Codex, launched from near zero in January, crossed five million users over the weekend of the interview. The fastest adoption of Codex within OpenAI itself is not in engineering but in go-to-market teams, a data point Friar flagged as indicative of where enterprise productivity gains are actually landing.
Advertising: Where Google's Intent Meets Meta's Targeting — With Memory on Top
Perhaps the most underappreciated strategic disclosure in the conversation was Friar's articulation of why OpenAI believes it is building a structurally superior advertising platform. She quoted Figma CEO Dylan Field's framing directly: "If Google and Meta had a baby, it would be ChatGPT." The logic is straightforward but powerful. Google captures high-intent search signals. Meta captures demographic and behavioral targeting. ChatGPT captures both — plus persistent memory. "Imagine putting memory and context next to intent," Friar said. "You should have a very potent ad platform." She confirmed ads are already in testing in the free tier and committed to maintaining an ad-free paid option for users who prefer it, while being clear that ad-supported access is the mechanism that allows OpenAI to pursue its stated mission of making AGI broadly accessible rather than only to paying customers.
She also disclosed that OpenAI believes it holds at least 11% of the global search market — and argued the real share is materially higher because search query counting methodologies undercount ChatGPT's usage relative to Google's page-refresh model. A full ChatGPT conversation with fifty exchanges counts as one interaction in current metrics; a Google session with fifty page refreshes counts as fifty. Advertisers with high-intent audiences will find that asymmetry increasingly difficult to ignore.
Capital Structure and the IPO: Optionality Over Timing
On the IPO question, which dominated the early part of the conversation given news that Anthropic had confidentially filed its S-1, Friar was characteristically disciplined. The $122 billion raise completed in March — which she described as the largest private fundraise in history by orders of magnitude, dwarfing Saudi Aramco's roughly $30 billion IPO — was designed explicitly to maximize flexibility rather than to accelerate or delay a public offering. "An IPO is a milestone, not a destination," she said. "No one remembers who went first, Google or Yahoo, Lyft or Uber." The implicit message to investors is that OpenAI will go public when the capital structure and business trajectory make it optimal, not in response to competitive pressure from Anthropic's filing.
The Agentic Revenue Inflection: From Laughable to Obvious
Friar offered a candid look back at her own forecasting track record as a way of illustrating how quickly the agentic revenue thesis has moved from speculative to mainstream. A year ago, she built investor models showing agentic API pricing of up to $2,000 per month. "Nobody believed," she said. "I don't even know what she's talking about. There's no way that will happen." She noted that the same skeptics had questioned whether anyone would pay $200 per month for ChatGPT Pro. The point is not triumphalism but a genuine signal about the pace of enterprise willingness to pay once agentic workflows demonstrate value — and a warning to analysts still anchoring models to historical SaaS pricing conventions.
The strategic through-line Friar returned to repeatedly is OpenAI's conviction that it is building the intelligence layer of the economy — a utility analogous to electricity, sitting beneath consumer interfaces, developer tools, enterprise deployments, and eventually government applications. The commoditization of LLMs that many predicted a year ago has not materialized; instead, the agentic and memory layer is creating deepening lock-in. "The models are now getting very connected to the memory and context and intuition of your company," she said, "and that's what gets CEOs and C-suite really excited." Whether OpenAI can sustain that position against Google, Anthropic, and a rapidly vertically integrating Nvidia is the question the market will eventually have to price — but for now, Friar has given institutional investors considerably more detail than they had before about how the company intends to win.