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

Summary

• Friar framed an IPO as “a milestone,” not “a destination,” after OpenAI raised $122 billion in March to maximize financing flexibility. She called it the largest private or public raise “by orders of magnitude” versus Saudi Aramco’s roughly $30 billion IPO. Anthropic’s confidential filing does not establish a winner: it still must “run the gauntlet of the SEC.”
• Friar did not directly address the host’s claim that Anthropic had overtaken OpenAI; she instead argued that the companies are pursuing different strategies. OpenAI wants one intelligence foundation spanning ChatGPT’s 900 million weekly users, Codex’s 5 million users, and enterprise products; more users and data should improve personalization, efficiency, gross margin, and ultimately access to compute. Revenue is already about 50/50 consumer and enterprise.
• Compute remains the binding constraint: OpenAI expects insufficient capacity in 2026 and a still-limited market in 2027. Friar called demand a “vertical wall” and treated energy, land, regulation, chips, memory, talent, and community trust as one supply chain. The 1-gigawatt Saline, Michigan project is described as providing 2,500 union jobs, $1 billion in taxes, and $45 million for education and Codex credits without raising local ratepayer bills.
• OpenAI’s investment case assumes intelligence delivered per dollar will improve faster than the all-in cost of a gigawatt rises. Friar put the cost reduction from GPT-4 to GPT-5.4 at “something like 97%” over two years; OpenAI raised GPT-5.5 prices 2×, yet she said customers still receive roughly 20–30% lower cost per token because each token is more efficient.
• The capital strategy is a Rubik’s Cube of infrastructure partners designed to preserve “maximum optionality.” OpenAI moved from one cloud, Azure, and one chip supplier, NVIDIA, to Oracle, CoreWeave, Microsoft, GCP, AWS, small neoclouds, AMD, Cerebras, and an OpenAI chip being developed with Broadcom. CSP financing shifts much of OpenAI’s infrastructure burden from CapEx into usage-linked OpEx.
• OpenAI plans to unveil a new consumer substrate by year-end and sell it early next year, but Friar would not confirm the host’s “puck and earpieces” description. Having tried it, she called the experience “very natural” and “very lovable,” arguing that great design makes the technology “fade away.”
• Advertising could subsidize broad access even though API tokens currently generate an order of magnitude more revenue per token than consumer tokens. Friar said sponsored content must never alter the model’s best answer and that an ad-free tier will remain. Her comparison was that ChatGPT combines high intent with memory and context, alongside Google’s search intent and Meta’s demographic “people like you” signal; she said OpenAI has “at least 11%” of the search market.

Deep dive

1. The IPO is a financing milestone, not the finish line

• Friar’s governing frame: an IPO is “a milestone” and “just another way to fundraise.” The $122 billion March raise supplied “maximum flexibility”; her CFO mandate is to create optionality rather than treat a listing as the company’s destination.
• She agreed the raise was the biggest private or public fundraising round in the comparison, “by orders of magnitude,” contrasting it with Saudi Aramco’s roughly $30 billion IPO. Sequence matters less than durability because “the market is a weighing machine, not a popularity machine.”
• When Jason reported Anthropic’s confidential S-1 filing, Friar rejected the scoreboard: it still must “run the gauntlet of the SEC.” Her historical analogy was blunt: “No one remembers who won first, Google or Yahoo, Lyft or Uber.”

2. One intelligence layer is meant to compound across every interface

• A host’s pushback—that Anthropic had “blown past” OpenAI among developers and corporations—was not conceded or directly rebutted. Friar instead defined OpenAI’s strategy as building “the AI layer, the infrastructure,” with one foundation distributed through many interfaces.
• Her scale evidence: ChatGPT has over 900 million weekly users and has become “the noun and the verb”; Codex reached 5 million users after starting near zero in January; Frontier and other channels address enterprises.
• The compounding mechanism runs from more users and data to better personalization, then model efficiency, lower token costs, higher gross margins, and more money for compute. “ChatGPT acts as a front door.”
• Asked whether gadgets, Sora, and other projects diluted enterprise focus, Friar rejected the consumer-versus-enterprise binary: revenue is about 50/50, and she described extensive current enterprise engagement. Usage rises from seven daily questions for free users to about 15 at the first paid tier, roughly 3× free at Plus, and 11× at Pro. She noted that free users do not receive the latest model and framed free access as a way for people to get a taste of intelligence and move up the commitment curve.

3. Scarce compute constrains 2027 even as new interfaces approach

• A host resurfaced Friar’s earlier rule that 1 gigawatt was roughly equivalent to $10 billion of annual revenue for OpenAI; she did not update that figure. She did say OpenAI faces a “vertical wall of demand,” lacks enough compute in 2026, and sees 2027 as “pretty limited” too.
• Friar counts energy, land, regulation, racks, chips, memory, talent, and trust as supply-chain constraints. Drawing on seven years at Nextdoor, she said communities must be engaged locally rather than told from the top down what they need. At the 1-gigawatt Saline project, she said ratepayers will not fund its power, while Michigan gets 2,500 union jobs, $1 billion in taxes, and $45 million for education and Codex credits.
• Training mostly still happens in the United States, while Friar wants inference to be global and more real-time for agents and multimodal interaction, including video. That leads into OpenAI’s unnamed consumer “substrate”: unveiling by year-end, sales early next year, and an experience Friar described as “very natural” and “very lovable.”

4. Falling unit costs support commitments years ahead of demand

• Friar starts capital allocation with measurable customer value. She said Thermo Fisher wants patient screening completed faster so it can obtain FDA approval faster; for someone with weeks to live, a breakthrough in two weeks instead of four “can literally be life or death.”
• Compute is the main input to cost of revenue, but its deflationary curve is steep. She put the cost reduction from GPT-4 to GPT-5.4 at “something like 97%” over two years. Although OpenAI raised GPT-5.5 prices 2×, she estimated customers still receive a 20–30% lower cost per token through greater efficiency.
• Forecasting for 2026–27 is bottom-up—products, pricing, weekly actives, subscriptions, advertising, and messages—though demand repeatedly surprises to the upside. Outer-year modeling reverses the equation: start with purchased compute, then estimate the revenue it might support.
• OpenAI is buying for 2028 onward; Saline may not produce compute until late 2027 or early 2028, while Friar feels shortest in 2030–32. A year earlier, investors doubted her projection that agentic developers might pay “upwards of maybe $2,000” monthly—just as people had balked at $200 ChatGPT Pro.

5. OpenAI is diversifying the stack while pursuing the profit pool

• Asked whether $122 billion funds OpenAI through 2031–32—and whether a host’s roughly $50 billion all-in estimate means $100 billion buys two gigawatts or five—Friar offered no simple runway answer. Her answer was financing architecture: CSPs shift CapEx into OpEx, with OpenAI paying as it generates revenue and uses the data centers while relying on partners’ ability to build and finance capacity.
• Two years ago, OpenAI had Azure, NVIDIA, ChatGPT, and one $20 price point. Now it uses Oracle, CoreWeave, Microsoft, GCP, AWS, and small neoclouds. NVIDIA remains the priority partner; the next big fall training run is planned for Vera Rubin, while Friar also mentioned AMD, Cerebras’ low-latency chip for developers, an OpenAI chip being developed with Broadcom, and a “Simon series” being plotted.
• On whether the stack will merge, Friar said everyone is trying to stay closest to the customer, where the largest share of ecosystem profits tends to sit. Her differentiation argument is that commoditization has moved in the opposite direction as the agentic “harness” adds memory, context, and enterprise intuition to the model. She illustrated that intuition with a trader who knows a pressured fund must sell a stock even when the available data suggests it should rise.
• Advertising fits that customer-layer thesis. Friar’s principles are that the model’s best result must not be displaced by sponsorship and that an ad-free tier will remain. Her shorthand was, “If Google and Meta had a baby, it would be ChatGPT”: ChatGPT offers explicit intent plus memory and context, while Google supplies search intent and Meta supplies demographic “people like you” signals. She said OpenAI has “at least 11%” of the search market.
• Yet the allocation tension is explicit: optimizing only for today would send “every token to the API,” where revenue per token is an order of magnitude higher than in the consumer product. The broader strategy is to serve consumers, small businesses, enterprises, and governments, including through free access, as an AI infrastructure-layer utility.