OpenAI Funds Post-Nvidia Shift via $122 Billion Cash Pile
All-In Podcast
CFO Sarah Friar treats AI infrastructure as a global utility, leveraging a multi-cloud strategy and hardware partnerships to drive down token costs while preparing for specialized enterprise adoption.
## Background
Published on June 2, 2026, by the All-In Podcast on YouTube, this episode features OpenAI CFO Sarah Friar. The page description highlights a wide-ranging discussion covering OpenAI's IPO timeline, the AI arms race with Anthropic and Google, navigating compute bottlenecks, a preview of a new hardware device, company economics, and their future advertising strategy.
## TL;DR
OpenAI CFO Sarah Friar discusses the company's historic $122 billion fundraising round and its strategic approach to a future IPO, emphasizing flexibility over rushing to market. She addresses the intense rivalry with Anthropic, detailing OpenAI's compounding platform strategy, massive user growth, and balanced 50/50 consumer-to-enterprise revenue split. Friar also previews a massive 1-gigawatt data center in Michigan, a highly anticipated hardware collaboration with Jony Ive, and the company's multi-chip, multi-cloud 'Rubik's Cube' infrastructure strategy designed to manage massive compute demands through 2032.
## Atmosphere & style
The conversation is fast-paced, highly strategic, and intellectually charged, blending high-finance pragmatism with visionary technological ambition. Friar maintains a calm, confident, and occasionally playful demeanor, easily laughing off competitive jabs from the hosts while delivering sharp, data-driven insights. The dynamic feels like an exclusive, high-stakes boardroom briefing where massive economic shifts and physical infrastructure realities are laid bare.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=0s) OpenAI's historic March fundraising round reached $122 billion, designed to provide maximum operational flexibility rather than treating an IPO as the final destination.
- [02:53](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=173s) Friar dismisses competitive pressure from Anthropic's confidential S1 filing, pointing to OpenAI's compounding single-model platform strategy that drives down token costs.
- [02:53](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=173s) ChatGPT has surpassed 900 million weekly active users, with Codex experiencing explosive growth to 5 million users over a single weekend.
- [05:49](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=349s) OpenAI's revenue is evenly split 50/50 between consumer and enterprise sectors, rejecting the binary classification of being just one or the other.
- [07:42](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=462s) Sam Altman is breaking ground on a massive 1-gigawatt data center in Saline, Michigan, in partnership with Oracle to address severe compute bottlenecks.
- [12:03](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=723s) OpenAI is collaborating with legendary designer Jony Ive on a secret, highly natural consumer hardware substrate expected by late 2026 or early 2027.
- [16:07](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=967s) Compute costs are experiencing massive deflation, with GPT-4 to GPT-4o seeing a 97% cost reduction over two years, making historical cost models quickly obsolete.
- [19:40](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1180s) OpenAI has transitioned to a highly diversified 'Rubik's Cube' infrastructure strategy, partnering with multiple cloud providers and chipmakers to shift capex to opex.
- [25:48](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1548s) The integration of deep contextual memory and organizational intuition is what prevents LLMs from becoming commoditized and drives massive enterprise value.
- [29:04](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1744s) OpenAI commands at least 11% of the search market and is exploring a highly potent, intent-driven advertising model that will coexist with an ad-free tier.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=0s) The Historic $120 Billion Fundraising Round and the IPO Race
The conversation opens with a discussion of OpenAI's massive, historic fundraising efforts. The interviewer notes that OpenAI's CFO, Sarah Friar, has just completed what is arguably "the most successful fundraising round in history." Friar confirms that they are actually raising north of $120 billion, driven by the belief that AI represents the most significant technological era to date, with massive implications for global productivity, affluence, and quality of life. When asked about the timing of a potential IPO—especially in light of rumors that SpaceX is preparing to go public and a *Wall Street Journal* article suggesting an advantage to IPOing early—Friar emphasizes that an IPO is merely a milestone and "not a destination." She explains that the $122 billion raised in March was designed to give the company maximum flexibility and optionality in this rapidly evolving era.
The scale of this private fundraising is unprecedented, dwarfing previous public and private markets. The speakers note that prior to this, the largest IPO in history was Saudi Aramco at approximately $30 billion. Friar points out that the market has grown significantly, with substantial capital sitting on the sidelines due to stock buybacks and M&A returning cash to shareholders. Addressing the competitive pressure of who goes public first, Friar quotes the famous market adage that "the market is a weighing machine, not a popularity machine," noting that history does not care whether Google or Yahoo, or Lyft or Uber, went first. The ultimate goal is to build a highly sustainable, durable company, with fundraising serving as a critical tool to achieve that.
- **Fundraising Total:** North of $120 billion (specifically referencing $122 billion raised in March).
- **Historical Comparison:** Saudi Aramco's IPO was previously the largest at ~$30 billion.
- **CFO Philosophy:** Friar views her primary job as creating "optionality" for the company and navigating the broader AI era rather than rushing to a public market destination.
[02:53](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=173s) The Anthropic Rivalry and OpenAI's Multi-Pronged Strategy
The interview takes a dramatic turn when Jason Calacanis chimes in with "breaking news," announcing that Anthropic has just confidentially filed its S1. He asks Friar if this puts OpenAI in "third place" in the race to file. Friar laughs off the pressure, noting that filing is just the beginning of running "the gauntlet of the SEC." Calacanis pushes further, suggesting that Anthropic has "blown past OpenAI" in terms of developers, corporate adoption, and revenue, and asks how OpenAI allowed this to happen despite their early, massive lead.
Friar firmly rejects this premise, outlining OpenAI's distinct, compounding platform strategy. Rather than focusing on a single niche, OpenAI is building a foundational AI infrastructure layer served up on a single model, which feeds multiple interfaces. This includes ChatGPT—which has over 900 million weekly active users and has become both "the noun and the verb" for how people experience AI—as well as Codex, their coding tool that surged to 5 million users over the weekend from nearly zero in January. She explains that this single-model approach creates a compounding advantage: more users generate more data, which allows for better personalization, greater model efficiency, lower token costs, and higher gross margins to fund massive compute needs.
- **ChatGPT Weekly Active Users:** Over 900 million.
- **Codex Growth:** Reached 5 million users over a single weekend, up from nearly zero in January.
- **Global Expansion:** OpenAI's economic research team identified Africa as the fastest-growing continent for usage, with Azerbaijani and Kazakh being the fastest-growing languages.
- **User Engagement Metrics:** Free users average 7 turns (questions) per day; the first paid tier doubles that to 15; Plus users ($20/month) do 3x the free tier; Pro users do 11x the free tier.
[05:49](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=349s) Balancing Enterprise and Consumer Missions
Addressing criticisms that OpenAI may have spread itself too thin with projects like the Sora video generator at the expense of enterprise focus, Friar dismisses the "binaryisms" of being labeled strictly a consumer or enterprise company. She reveals that OpenAI's revenue is currently balanced at roughly a 50/50 split between the two. Under the leadership of Denise Dresser, who took over as Head of Revenue in December, OpenAI's enterprise arm is "firing on all cylinders," with Friar personally engaging with clients ranging from Thermo Fisher in Boston to major Wall Street banks and Travelers Insurance.
At the same time, Friar emphasizes that OpenAI's core mission is "AGI for the benefit of humanity," which means they cannot leave the consumer behind. This mission is why they offer a robust free tier, allowing users to "get a taste of intelligence" before moving up a steep commitment curve. She compares this transition to the evolution of the mobile phone, where users initially only expected to make calls on a flip phone but now rely on smartphones for almost every aspect of daily life.
- **Revenue Split:** Approximately 50/50 between consumer and enterprise.
- **Key Leadership:** Denise Dresser, Head of Revenue since December.
- **Corporate Clients Mentioned:** Thermo Fisher, Travelers Insurance, and various New York banks.
[07:42](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=462s) The Compute Bottleneck and the 1-Gigawatt Michigan Data Center
The conversation shifts to the physical infrastructure powering the AI revolution. The interviewer recalls a metric Friar shared 18 months prior: the economic trade-off of "gigawatts to cash," where 1 gigawatt of power roughly equates to $10 billion in annual revenue. Friar confirms that compute remains an incredibly scarce resource, describing the demand curve as a "vertical wall." She praises co-founders Greg Brockman and Sam Altman for their foresight in aggressively buying up compute the previous year, noting that even by 2026, the industry will face a severe deficit.
Friar details the shifting choke points in the AI supply chain, which range from energy, land, and regulatory approvals to chips, memory spikes, and talent. She also highlights "trust" as a critical, often overlooked component of the supply chain. To illustrate this, she shares a "sneak preview" of Sam Altman's activities that very day: he is in Saline, Michigan, breaking ground on a massive 1-gigawatt data center as part of their Oracle complex. Drawing on her seven years of experience at Nextdoor, Friar explains that local communities must be engaged from the bottom up rather than top-down.
- **The Gigawatt Equation:** 1 gigawatt of power is roughly equivalent to $10 billion in annual revenue.
- **New Infrastructure Project:** A 1-gigawatt data center in Saline, Michigan, built in partnership with Oracle.
- **Community Commitments in Michigan:**
- OpenAI is paying for its own power and infrastructure so local ratepayers' bills do not increase.
- The project will bring 2,500 union jobs (electricians, HVAC, etc.).
- It will generate $1 billion in local taxes.
- OpenAI is investing $45 million in local education, specifically providing Codex credits.
[12:03](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=723s) The Future of Compute, Global Inference, and the Johnny Ive Hardware Collaboration
Looking ahead to 2026 and 2027, Friar warns that compute availability will remain extremely tight across the industry. She explains the geographic split in infrastructure: AI training will continue to happen primarily on US soil for national security reasons, while inference must be distributed globally to enable real-time, low-latency agentic experiences. This compute scarcity forced OpenAI to make tough choices, such as limiting the rollout of their resource-heavy Sora video model. However, Friar notes that the future lies in true multimodality, moving away from the "disease" of people staring down and communicating with their thumbs to actually speaking naturally with their technology.
This transition to voice and multimodality leads to a discussion about OpenAI's highly anticipated, secret hardware collaboration with legendary designer Jony Ive. While Friar jokingly refuses to confirm if the device is an earpiece—quipping that "Jony will come and steal my teenage son" if she reveals too much—she promises that OpenAI will unveil a new "consumer substrate" by the end of the year or early next year. Having personally tested the device, Friar describes the experience as incredibly natural, seamless, and "lovable," praising Ive's team for their unique ability to "bring humanity to devices" and make the underlying technology fade away.
### [16:07](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=967s) Customer Value, Compute Deflation, and Capital Allocation
The conversation shifts to how modern, outsized returners deploy capital at a higher Return on Capital (ROC) than anyone else, and how OpenAI plans to build a durable, high-value company. Sarah Friar explains that the great companies of this era will not be "magical" anomalies; they will look like the great companies of prior eras by starting with customer value—helping customers drive revenue or gain efficiency. She points to Thermo Fisher Scientific as an example. They have roughly 38,000 field sales representatives and want to get patient screenings done faster to secure FDA approvals quicker. In life-or-death scenarios like cancer, "the difference between a breakthrough in four weeks and two weeks can literally be life or death." Interestingly, Friar notes that the fastest-growing month-over-month takeoff of Codex within OpenAI is actually in their own go-to-market (GTM) team, as they seek more productivity, alongside exciting developments in finance.
To turn customer value into great gross margins, OpenAI must manage its main cost of revenue: compute. Friar highlights a massive deflationary curve on compute costs, noting that from GPT-4 to GPT-4o, the cost deprecated by roughly 97% in just two years. With their newest model, GPT-5.5, OpenAI raised prices twofold, yet customers still experience a 20% to 30% cost reduction per token because the model is vastly more efficient. Friar warns that making capital allocation decisions based on today's cost profile risks mispricing the outcome. Looking ahead, her current compute focus is on securing capacity for 2028 and beyond. For instance, OpenAI's Seline, Michigan data center is not expected to yield compute until late 2027 or early 2028.
- **Compute Cost Deflation:** GPT-4 to GPT-4o saw a 97% cost reduction over a two-year span.
- **GPT-5.5 Pricing:** Despite a 2x price increase, customers see a 20% to 30% cost reduction per token due to efficiency gains.
- **Future Planning:** Current compute procurement is focused on the years 2028 through 2032, with the Michigan data center expected online around late 2027/early 2028.
### [19:40](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1180s) Forecasting Multi-Year Compute Needs and the "Rubik's Cube" Strategy
When asked how OpenAI forecasts compute needs multiple years out amidst rapid architectural advancements, Friar explains that they assume raw compute on a per-gigawatt basis is getting more expensive due to rising power and memory costs. However, the intelligence gained on the other side due to chip-level improvements more than compensates for this. For near-term years like 2026 and 2027, Friar builds a bottoms-up model based on products, pricing, and user metrics (P * Q, weekly actives, messages, etc.), though she admits "the shape of the line keeps taking us by surprise to the upside." For outer years, they reverse-engineer the math, mapping purchased compute to expected revenue. She laughs recalling how a year ago, she modeled "agentic revenue" where developers would pay upwards of $2,000 a month—a notion people found laughable when users were already complaining about ChatGPT Pro costing $200.
Addressing the sheer scale of capital required—with estimates suggesting 1 gigawatt of AI compute costs about $50 billion all-in—Friar details OpenAI's transition from a simple setup to a highly diversified "Rubik's Cube" strategy. Just two years ago, OpenAI had one Cloud Service Provider (CSP) in Microsoft Azure, sat on one chip (Nvidia), and had one product (ChatGPT) at a single $20 price point. Today, they partner with multiple CSPs (including Oracle, CoreWeave, Microsoft, GCP, AWS, and smaller neoclouds), which effectively shifts capital expenditure (capex) into operational expenditure (opex), allowing them to pay as they generate revenue.
- **Multi-Chip Strategy:** While Nvidia remains the absolute priority partner (with the next big training run in the fall on Rubin, followed by the Feynman series), OpenAI has diversified with AMD, Cerebras (a low-latency chip favored by developers), and their own custom silicon being co-developed with Broadcom.
- **Built-to-Suit Infrastructure:** OpenAI is beginning to move beyond standard CSP models, recently announcing a built-to-suit data center in Texas with SoftBank Energy, which will require more direct capex.
- **Rubik's Cube Metaphor:** Friar notes a Rubik's Cube has "something like a quintillion" different configurations, representing the maximum optionality she seeks to build for OpenAI before they reach an investment-grade status to secure lower-cost debt.
### [25:48](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1548s) The Merging Tech Stack and the Power of Contextual Memory
The hosts discuss how the traditional tech stack is merging, with silicon providers building models, cloud companies designing chips, and application companies spinning their own silicon. Friar explains that everyone is trying to reside closest to the customer to capture the largest portion of the ecosystem's profits. While critics predicted the commoditization of Large Language Models (LLMs) a year ago, the opposite has occurred due to the "harness" of context and memory. Friar shares a personal example: her own Codex has a massive memory file. "It knows that I'm me, it knows I'm the CFO of OpenAI, it knows how I like to write things... it actually also knows that I'm a mom of teenagers."
This level of memory and context becomes incredibly powerful when brought into an enterprise environment, capturing the "intuition" of a company. Friar shares an anecdote from her days working on Wall Street: all the data in the world might suggest a stock should go up after an earnings call, but a quick call to a trader would reveal, "Yeah, stock's not going up, Sarah... I know this fund is under pressure and they need to sell down their book." That unquantifiable intuition exists in every enterprise, and connecting models to this organizational memory is what has CEOs and C-suite executives excited about driving both top-line revenue and efficiency.
### [29:04](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1744s) The Future of Advertising on ChatGPT
In a rapid-fire closing question, the hosts point out that two of the three greatest consumer businesses of our lifetime—Meta and Google—are ad-based, and ask about OpenAI's commitment to an ad-supported model. Friar asserts that OpenAI will always stick to its core principles: search results must be determined by the model's quality, not by sponsorships, and they will always provide an ad-free tier for users who want it. However, she quotes Instacart CEO Fidji Simo, saying, "if Google and Meta had a baby, it would be ChatGPT."
Friar explains that OpenAI already commands at least 11% of the search market—and likely much more, because while a Google page refresh counts as a search, a single multi-turn ChatGPT conversation of 50 questions only counts as one. This conversational format captures incredibly high user intent. Furthermore, while Meta targets users based on demographics ("people like you"), OpenAI pairs high intent with deep contextual memory. This combination creates an incredibly potent potential ad platform to fund free access for the world. Friar concludes by noting that if she were optimizing only for short-term revenue, she would route "every token to the API." Instead, OpenAI is playing its own game, treating AI infrastructure like a utility—similar to electricity—to serve consumers, small businesses, large enterprises, and governments alike.
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=0s) An IPO is a milestone, not a destination. — Sarah Friar
> [00:00](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=0s) The market is a weighing machine, not a popularity machine. — Sarah Friar
> [02:53](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=173s) ChatGPT has become both the noun and the verb for how people experience AI. — Sarah Friar
> [05:49](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=349s) We cannot leave the consumer behind in our mission of AGI for the benefit of humanity. — Sarah Friar
> [07:42](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=462s) The demand curve for compute is a vertical wall. — Sarah Friar
> [12:03](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=723s) Jony will come and steal my teenage son if I say too much. — Sarah Friar
> [16:07](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=967s) The difference between a breakthrough in four weeks and two weeks can literally be life or death. — Sarah Friar
> [16:07](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=967s) Making capital allocation decisions based on today's cost profile risks mispricing the outcome. — Sarah Friar
> [29:04](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1744s) If Google and Meta had a baby, it would be ChatGPT. — Sarah Friar
> [29:04](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1744s) If I were optimizing only for short-term revenue, I would route every token to the API. — Sarah Friar
## Mentioned
OpenAI, Sarah Friar, Sam Altman, Greg Brockman, Denise Dresser, Jony Ive, Anthropic, Google, Meta, Oracle, Microsoft Azure, Nvidia, AMD, Broadcom, Cerebras, SoftBank Energy, Thermo Fisher Scientific, Travelers Insurance, ChatGPT, Codex, Sora
OpenAI Chief Financial Officer Sarah Friar argues that sustaining leadership in artificial intelligence requires building strategic and capital optionality through compute diversification, rapid model cost deflation, and persistent contextual memory, rather than prioritizing an early initial public offering.
Capital Optionality and Strategic IPO Timing (0:00)
Friar frames OpenAI's fundraising—surpassing $120 billion, including $122 billion recorded in March—as a means to secure long-term independence during an unprecedented technological transition, exceeding past records such as Saudi Aramco's approximately $30 billion public offering. Addressing competitive moves like Anthropic's confidential S-1 registration, she contends that an IPO remains a financing milestone rather than an ultimate destination, arguing that market longevity matters more than being first to file.
To manage immense capital demands before reaching an investment-grade rating for low-cost debt, OpenAI employs what Friar terms a "Rubik's Cube" capital structure. The company has shifted from relying solely on Microsoft Azure to contracting across multiple cloud providers, including Oracle, CoreWeave, Google Cloud Platform, and Amazon Web Services. This structure converts fixed capital expenditures into variable operating expenses, while OpenAI selectively pursues direct-capex projects, such as a built-to-suit data center in Texas with SoftBank Energy.
Compute Scarcity, Energy Bottlenecks, and Infrastructure Geography (7:42)
Friar describes industry-wide compute demand as a "vertical wall," projecting supply shortages through at least 2026 and 2027. Under Friar's metric, one gigawatt of power roughly translates to $10 billion in annual revenue, against an estimated all-in deployment cost of $50 billion per gigawatt. Supply constraints span power generation, land acquisition, advanced silicon, memory chips, and municipal trust. In Saline, Michigan, OpenAI and Oracle broke ground on a one-gigawatt facility, committing to cover all dedicated power infrastructure to protect local utility rates, while pledging 2,500 union jobs, $1 billion in local taxes, and $45 million in educational coding credits.
Geographically, Friar expects primary model training to stay inside the United States due to national security concerns, whereas inference infrastructure must be deployed globally to support low-latency agentic applications. To insulate itself against chip shortages, OpenAI maintains Nvidia as its core partner for future Rubin and Feynman architectures, while incorporating AMD and Cerebras chips and co-developing custom silicon with Broadcom.
Deflationary Unit Economics and Contextual Moats (16:07)
Despite escalating infrastructure scale, Friar highlights rapid deflation in unit compute costs: token costs dropped roughly 97% between GPT-4 and GPT-4o over two years. For GPT-5.5, OpenAI doubled nominal pricing, yet customers saw effective per-token costs decrease by 20% to 30% due to architectural efficiencies. Because immediate unit economics fluctuate quickly, Friar bases multi-year infrastructure planning on longer horizons, currently procuring power and capacity for 2028 through 2032.
Friar rejects predictions that foundation models will commoditize, asserting that retention relies on software context and accumulated memory. OpenAI's revenue is currently divided roughly 50/50 between enterprise clients and consumer products under revenue head Denise Dresser. Enterprise deployments, such as Thermo Fisher Scientific utilizing models across 38,000 sales representatives to accelerate clinical research and FDA processes, rely on institutional context to capture operational intuition that standard algorithms miss.
Consumer Expansion, Multimodal Hardware, and Advertising Potential (12:03)
ChatGPT has reached over 900 million weekly active users, with usage scaling from an average of 7 daily queries on the free tier to 11 times that level among Pro subscribers. Its coding tool, Codex, reached 5 million users shortly after release, with the highest usage growth occurring in Africa and in languages like Azerbaijani and Kazakh. To expand input formats beyond mobile screens, OpenAI is partnering with designer Jony Ive on a dedicated multimodal device slated for late this year or early next year, focused primarily on natural voice interaction.
To sustainably fund broad consumer access, Friar outlines potential monetization through high-intent advertising. Estimating that OpenAI already captures at least 11% of search activity—a figure she considers undercounted due to multi-turn conversational sessions—Friar argues that pairing direct user intent with contextual memory yields an ad model more effective than demographic targeting. However, she maintains that model responses will remain unsponsored and that OpenAI will preserve an ad-free subscription option.
While Friar outlines aggressive scaling across consumer, enterprise, and custom silicon initiatives, OpenAI's long-term trajectory depends on navigating persistent multi-year power shortages, successfully bringing delayed multi-gigawatt facilities online by 2028, and managing steep operational expenses across competing cloud ecosystems.
## Background
本期内容由 All-In Podcast 频道于 2026 年 6 月 27 日发布在 YouTube 平台。页面简介显示，节目邀请了 OpenAI 首席财务官 Sarah Friar 做客，深入探讨了 OpenAI 的 IPO 时间线考量、与 Anthropic 和 Google 的 AI 军备竞赛、算力瓶颈与全新硬件设备预览、商业经济学与芯片布局，以及潜在的广告业务战略。
## TL;DR
在访谈中，OpenAI 首席财务官 Sarah Friar 详述了公司创纪录的融资进展与资本配置哲学，重申 IPO 仅是募资里程碑而非企业终点。针对算力短缺挑战，她分享了 1 吉瓦电力对应百亿美元年营收的测算逻辑、密歇根超大数据中心的社区落地承诺，以及向多云与多芯片架构转移的策略。同时，Friar 首度确认了与 Jony Ive 团队合作开发的人性化消费级 AI 硬件原型，并揭示了 ChatGPT 结合高意图搜索与深度上下文的巨大广告潜力，强调 OpenAI 宁可牺牲部分短期 API 变现利润，也要构建普惠全社会的 AI 基础设施底座。
## Atmosphere & style
对谈整体节奏紧凑、交锋激烈而又不失幽默风趣，四位主持人的轮番追问充满了华尔街投资人特有的敏锐与直率。Sarah Friar 展现了极强的控场能力与战略定力，不仅用播客经典开场白轻松化解上市竞争的火药味，在探讨百亿级重资产投资与严酷算力荒时，语气更是从容笃定。从回忆北爱尔兰童年接触大英百科全书的温情震撼，到面对前沿硬件设计与超级基础设施宏图时的澎湃自豪，整个现场始终洋溢着见证历史性技术革命的兴奋感与务实雄心。
## Bulletin
- [00:00](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=0s) OpenAI 首席财务官 Sarah Friar 证实公司近期完成超 1200 亿美元估值规模（含 3 月敲定的 122 亿美元融资）的巨额私募，并强调 IPO 仅是募资里程碑而非战略终点，公司不会被外界与 Anthropic 的上市竞赛节奏所裹挟。
- [03:01](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=181s) ChatGPT 周活跃用户已突破 9 亿，Codex 开发者从年初近乎为零激增至 500 万以上；公司营收结构已达到 C 端与 B 端各占约 50% 的均衡分布，单位推理成本在两年内断崖式下降 97%。
- [09:53](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=593s) 算力供给面临严重赤字，OpenAI 正与甲骨文合作在密歇根萨林破土动工规模达 1 吉瓦的数据中心，并承诺全额承担能源基建成本、不推高当地电费，同时创造 2500 个工会岗位与 10 亿美元税收。
- [14:43](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=883s) Sarah Friar 正式确认 OpenAI 正在与前苹果设计总监 Jony Ive 联合打造全新消费级 AI 硬件原型机，预计将于今年年底或明年初正式向全球揭晓。
- [22:30](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1350s) OpenAI 算力基建摆脱对单一云服务商和单款芯片的依赖，构建多云与多芯片体系，未来算力真正的极度匮乏周期将出现在 2030 至 2032 年。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=0s) IPO 只是企业发展与募资的里程碑而非终点站，媒体热衷制造针锋相对的竞争叙事，但企业长青的核心在于构建可持续的底层价值与长期财务韧性。
- [03:01](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=181s) OpenAI 践行“单一基座模型、多元触达界面”飞轮战略，以 ChatGPT、Codex 与 Enterprise 矩阵沉淀数据并摊薄 Token 成本，释放充沛现金流反哺算力。
- [05:49](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=349s) 破除 B 端与 C 端的非黑即白二元对立，目前商业营收维持 50/50 的均衡结构，既通过企业服务创造深度价值，又坚持 C 端免费普及以践行 AGI 使命。
- [05:49](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=349s) 用户从免费向付费层级跃迁时交互深度呈指数级爆发，Pro 专业用户的日交互频次达到免费用户的 11 倍，展现出类似功能机向智能手机演进的不可逆依赖。
- [07:42](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=462s) 电力到现金的转化法则确立，1 吉瓦电力规模约对应每年 100 亿美元营收产出，算力供应链正在能源、审批、芯片和顶尖人才等全链路遭遇严重堵塞。
- [09:53](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=593s) 大型数据中心落地需深度扎根本地社区，通过承担电力增量成本、提供工会就业、贡献税收及设立教育基金等务实承诺化解居民敌意。
- [12:03](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=723s) 2026 至 2027 年全球算力仍处于极度紧缺状态，OpenAI 采取训练算力锚定美国本土、推理算力全球低延迟分布的双重布局。
- [14:43](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=883s) 自然多模态交互将终结“大拇指打字”的畸形沟通习惯，与 Jony Ive 合作的全新硬件设备已进入原型测试阶段，强调化繁为简与注入人性。
- [16:01](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=961s) 算力单位经济学呈现急剧通缩曲线，GPT 系列单位推理成本降幅达 97%，模型智能密度跃升让客户在单价提升的情况下实际任务成本净降 20% 至 30%。
- [16:01](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=961s) 资本开支决策具备高度前瞻性，目前重心已锁定在 2028 年乃至 2030-2032 年的算力管线储备，以此应对漫长的数据中心建设交付周期。
- [19:26](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1166s) 算力需求预测采用双轨制：近期依托产品转化自下而上建模，远期则基于采购的算力底座反向推导营收空间，智能体未来单月单客价值具极高上限。
- [22:30](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1350s) 资本架构从单薄走向多元“魔方”，全面引入微软 Azure、甲骨文、AWS、GCP 等多云服务商，并将高强度 CapEx 转化为按需付费的 OpEx 以优化现金流。
- [22:30](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1350s) 芯片供应链由英伟达核心支撑逐步扩展至 AMD、Cerebras 以及与博通联合定制的自研芯片，确保算力弹性和极低延迟表现。
- [29:09](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1749s) ChatGPT 兼具 Google 的高意图搜索需求与 Meta 的精准画像特征，并叠加长期记忆能力，潜藏着极高转化率的广告商业化爆发力，但坚持输出纯粹客观答案与提供无广告选项的底线。
- [31:07](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1867s) 尽管 API 业务在单 Token 变现效率上具有数量级层面的优势，OpenAI 仍坚持重金投入消费端，旨在将自身打造为如水电般普及的现代社会 AI 通用基础设施。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=0s) 史诗级融资与对 IPO 节奏的理性思考
OpenAI 首席财务官 Sarah Friar 登台 All-In 播客，主持人开场即点出 OpenAI 刚刚完成了历史上最具轰动性的私募融资。Sarah Friar 证实 OpenAI 筹集了超过 1200 亿美元（具体为 3 月份敲定的 122 亿美元巨额融资，此轮募资赋予了团队极大的财务灵活性）。她感慨 AI 标志着人类迄今经历的最庞大时代，社会才刚开始触及生产力爆发所带来的红利，而作为 CFO，她的核心使命是在这个大时代中为公司创造最大化的“选择权（Optionality）”。面对台下四位主持人的轮番提问，Friar 风趣地用播客经典开场白“长期听众，首次连线”（Longtime listener, first time caller）化解了同台交锋的紧张感。
David Sacks 随后抛出重磅话题：华尔街日报近期评论 AI 公司越早上市越占优，随着 SpaceX 推进 IPO 进程，市场极为关注 OpenAI 与 Anthropic 的上市竞赛。Sarah Friar 对此予以坚决纠正，她向团队始终传达一个理念：IPO 仅仅是一个“里程碑（Milestone）”，绝非企业的“终点站（Destination）”，绝不能把上市当作某种战略归宿来运营公司，它本质上只是一种募资途径。她借用传奇投资格言指出，“市场短期是一台投票机，长期是一台称重机”。历史上没人会去纠结 Google 和 Yahoo 谁先上市、Lyft 和 Uber 谁先敲钟，先发上市与否并不决定胜败。媒体热衷于制造针锋相对的戏剧冲突，但企业唯一的目标是构建体量庞大、可持续且具备长期韧性的业务。
- **创纪录的募资体量**：在此之前历史上规模最大的 IPO 是沙特阿美（Saudi Aramco）约 300 亿美元的规模，而当下私市单轮融资与未来 AI 龙头的潜在上市规模均远超 2000-2001 年互联网泡沫时期的体量。
- **充沛的市场流动性**：美股近些年频繁发生巨额股票回购与兼并收购，大量资金沉淀在场外等待优质标的，资本市场对头部 AI 企业承接力十足。
---
### [03:01](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=181s) OpenAI 与 Anthropic 的竞争分歧及飞轮战略
Jason Calacanis 在对谈中突然抛出“突发新闻”，称 Anthropic 刚刚秘密提交了 S-1 上市申请表格，以此追问 OpenAI 是否在上市节奏上落入下风，甚至在开发者生态、企业级落地和营收上被 Anthropic 迎头赶上。Sarah Friar 坦然回应，秘密提交申请仅是第一步，后续仍需直面 SEC 漫长而严苛的审查流程。更重要的是，她明确阐述了 OpenAI 与竞品截然不同的商业与架构哲学。
```mermaid
flowchart TD
accTitle: OpenAI 单一基座与多元产品飞轮机制
accDescr: 展示从统一模型层到多元交互界面，再通过数据回流与规模效应降低成本并反哺算力的闭环结构
subgraph Base ["统一基座模型层"]
Foundation["单一模型基座 (Single Foundation)"]
end
subgraph Interfaces ["多元触达界面 (Multi-Interfaces)"]
Consumer["ChatGPT (9亿+周活消费级用户)"]
Dev["Codex (500万+开发者)"]
Enterprise["Frontier (企业级定制套件)"]
end
subgraph Flywheel ["正向飞轮效应"]
DataScale["用户规模扩增与极致个性化"]
CostAdvantage["大模型推理效率提升降低Token成本"]
CapitalGeneration["高毛利释放巨额现金流反哺算力"]
end
Foundation --> Consumer
Foundation --> Dev
Foundation --> Enterprise
Consumer --> DataScale
Dev --> DataScale
Enterprise --> DataScale
DataScale --> CostAdvantage
CostAdvantage --> CapitalGeneration
CapitalGeneration --> Foundation
```
Friar 详细拆解了这一“单底座、多接口”飞轮的底层逻辑：OpenAI 坚持构建统一的 AI 基础设施底座，并向外延伸出多元界面。ChatGPT 是面向大众的超级入口，目前每周活跃用户已突破 9 亿，成为全球大众认知 AI 的“代名词”与日常动词；不仅覆盖欧美，其经济研究团队最新数据显示，非洲成为增速最快的大陆，阿塞拜疆语与哈萨克语则是使用增速最快的语种。面向开发者，Codex 用户量刚刚突破 500 万大关（年初几乎从零起步）；面向商用客户，则有 Enterprise 矩阵。统一基座意味着飞轮的正向复合累积：更多用户带来更多场景数据，进而驱动更深度的个性化；随着模型参数与架构优化，单位 Token 交付成本被大幅摊薄，毛利率同步提升，最终形成反哺天量算力采购的充沛现金流。
- **业务体量指标**：ChatGPT 周活超 9 亿人次；Codex 开发者从 1 月初的几乎为零飙升至 500 万以上。
- **长尾地域爆发**：全球化扩散迅速，非洲成为增长最快的区域，阿塞拜疆语（Azerbaijani）与哈萨克语（Kazakh）位居增速榜首。
---
### [05:49](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=349s) 破除“C端还是B端”的二元论与梯级商业化曲线
针对外界质疑 OpenAI 战略分散——既做智能硬件、又做 Sora 视频生成，导致在企业级（B 端）精力不足的论调，Sarah Friar 作出有力反驳。她直言外界总喜欢陷入非黑即白的二元对立（“你们到底是 2C 还是 2B？”），而现实情况是 OpenAI 兼具两端优势，目前的营收结构已经极为健康地达到了接近 50/50 的平衡分布。OpenAI 在企业端绝未松懈，新任首席营收官 Denise Dresser（2025 年 12 月入职）执行力极强；Friar 分享了自己的行程：仅过去一周内，她就奔赴波士顿拜访了赛默飞世尔（Thermo Fisher Scientific），在纽约会晤了多家华尔街银行，并与旅行者保险（Travelers）及多家顶级科技公司高层深入探讨了全面接入方案。
与此同时，Friar 强调 OpenAI 绝不可能抛弃 C 端，因为公司的根本使命是“造福全人类的 AGI”，绝非仅仅服务于有付费能力的人群或大企业雇员。OpenAI 坚定推行大规模免费策略，正是为了让全人类尽早体验到智能的价值。而一旦用户尝到了智能的甜头，其使用深度和向上付费的跃升曲线是极为惊人的：
| 用户层级 / 订阅方案 | 每日交互轮次（Turns） | 相对免费层级的使用强度 |
|:--- |:--- |:--- |
| **免费用户 (Free)** | 约 7 次提问 / 天 | 基准 (1x) |
| **入门付费层级 (Tier 1)** | 约 15 次提问 / 天 | 约 2x |
| **Plus 订阅用户 ($20/月)** | 约 21 次提问 / 天 | 约 3x |
| **Pro 专业级用户** | 约 77 次提问 / 天 | 约 11x |
Friar 用功能机（Flip Phone）向智能手机的演变作比喻：最初大家觉得翻盖手机无非就是打打电话，如今智能终端已接管生活方方面面。人类对智能的依赖与使用习惯也正沿着这条路径不可逆地加速演进。
---
### [07:42](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=462s) 算力荒暴击：电力变现模型与供应链全面瓶颈
对话切入深层基础设施议题。主持人提到 Sarah Friar 在 18 个月前提出过一个震动业界的经济学公式：“电力到现金的转化”，即 **1 吉瓦（GW）电力规模大致对应每年 100 亿美元的营收产出**。彼时 OpenAI 激进储备带电土地和超大数据中心容量的举动曾被外界视作疯狂，甚至遭受冷嘲热讽，但如今看来却极具远见，行业已普遍面临严重的算力供给赤字。
Friar 确认算力供需严重失衡，OpenAI 业务当前正在攀爬“垂直的绝壁需求墙”，市场上可用的 Token 产出根本无法满足暴涨的诉求。幸亏去年与联合创始人 Greg Brockman、Sam Altman 顶住重重阻力大举采购算力，“即便到了 2026 年，算力也依旧会不够用”。在拆解算力供应链的梗阻时，Friar 勾勒出了一整套动态转移的卡点图景：
- **能源与土建审批**：首先且最核心的是电力（Power）与带电土地（Powered Land），以及能否获得允许高速开工建设的监管与环评许可。
- **机柜、芯片与元器件**：硬件供应链条环环相扣，高带宽内存（HBM）等存储元器件暴涨及芯片配额争夺仍在持续。
- **顶尖人才储备**：工程与基础科研人才供给告急；身为斯坦福大学校董（Trustee），Friar 极度忧虑当前教育体系能否为国家和产业源源不断输送顶尖数理与工程专才。
- **社会信任与落地阻力**：公众接纳度是供应链隐蔽却关键的一环。采访进行时，Sam Altman 正亲赴密歇根州萨林（Saline, Michigan），准备在两小时后为全新的大型数据中心基础设施剪彩，表明 OpenAI 正在把技术基建纵深扎入美国本土重工业与能源腹地。
### [09:53](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=593s) 密歇根1吉瓦数据中心破土与社区共赢策略
OpenAI 首席财务官 Sarah Friar 在现场透露了一项重磅消息：公司正准备为一座规模达 1 吉瓦（GW）的超大型数据中心破土动工，该项目是其与甲骨文（Oracle）合作综合体的一部分。谈及如此庞大的基建落地，Sarah 强调获取社区信任至关重要，绝不能将项目所在地的民众甩在身后。她分享了自己在社区社交平台 Nextdoor 担任首席执行官长达七年的切身体会——扎根本地社区是极其艰难细致的工作，任何自上而下的傲慢命令都会遭到居民“谢谢，但请离开”的抵制，因为社区居民非常清楚自己真正需要什么。
为了彻底消除当地社区对超大型数据中心的敌意与担忧，OpenAI 采取了极为透明且务实的沟通方案，向密歇根萨林（Saline）当地作出了四项极具分量的承诺：第一，明确承诺绝不会推高当地居民的电费账单，OpenAI 会全额承担自身基建和电力扩张的成本，绝不让普通纳税人和用电户买单；第二，为当地带来 2,500 个高质量的工会就业岗位，涵盖电工、暖通空调（HVAC）工程师等专业工种；第三，仅这一座数据中心就将为密歇根州带来高达 10 亿美元的税收贡献；第四，额外出资 4,500 万美元用于当地教育事业，发放 Codex 学习额度，让下一代在步入职场前就熟练掌握前沿生产力工具。Sarah 形象地打比方道，过去招财务绝不会录用一个不懂 Excel 的人，而今天她也绝不会雇佣一个不会使用 Codex 等 AI 工具的财务人员。
- **电力与成本隔离**：企业独立承担所有能源基础设施投资，确保当地电价不受冲击。
- **高质量就业**：创造 2,500 个涵盖电气、暖通工程等领域的工会岗位。
- **税收贡献**：单座密歇根数据中心预计上缴税款达 10 亿美元。
- **教育扶持**：投资 4,500 万美元教育基金，通过 Codex 额度资助学生掌握 AI 时代的工作技能。
---
### [12:03](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=723s) 算力荒现状：2026-2027年全球算力极度短缺与代际体验
面对 Chamath 关于全行业电力可用度与算力供需格局的追问，现场展开了一场关于“算力瓶颈”的现实碰撞。Chamath 吐槽自己使用竞品 Anthropic 时经常遭遇限流打击——早上 10 点半系统弹窗显示“下午 2 点半再见”，这种体验在真实工作流中简直无法容忍，而 ChatGPT 却极少出现这种情况。Sarah 对此解释称，OpenAI 团队在 Token 配额上一直保持相当慷慨的策略，哪怕在免费层级也尽最大努力开放访问，目的是让每一个人都能切身感知到 AI 的能力。
Sarah 动情地回忆起自己在北爱尔兰动乱时期（The Troubles）一个偏僻小社区长大的童年往事：当全套大英百科全书（Encyclopaedia Britannica）第一次送到家门口时，对年幼的她来说宛如“拨云见日，晴空万里”。OpenAI 希望如今世界各地用 AI 写作业的孩子们，也能在屏幕前体会到那种认知被彻底点亮的震撼。然而温情背后是严酷的算力现实。Sarah 直言，眼下在 2026 年如果有谁想去市场上买到更多现成算力，“那我只能祝你好运，因为连我都不知道上哪儿去弄”；到了 2027 年，全行业的算力供给同样高度吃紧（现场笑称马斯克成了极少数手里算力过剩并成功变现的人）。
在算力架构的顶层设计上，OpenAI 清晰地划分了“训练”与“推理”的双重版图：模型训练主要锚定在美国本土进行，这关乎美国国家资产与国家安全层面的考量；而推理算力则必须实现真正的全球化低延迟分布。尤其随着 AI 进入智能体（Agentic）与多模态时代，系统对低延迟交互的要求呈几何级上升。此前在推出 Sora 视频生成模型时，OpenAI 正是因为遭遇极端严重的算力挤兑，才不得不做出了大幅收紧权限的艰难决断。
---
### [14:43](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=883s) 与 Jony Ive 合作的神秘 AI 硬件与多模态交互革命
Sarah 指出，多模态交互的崛起正在从根本上终结过去二十年科技强加给人类的畸形习惯。她感叹“用大拇指打字沟通已经成了一种疾病”：街上的人全都低头看手机，甚至她家里的青少年晚上并排坐在沙发上，彼此说话竟然还在用大拇指发短信。随着多模态技术的成熟，人机交互正迅速回归到自然的语音与视线交流，但这背后需要庞大的实时计算能力作为支撑。
谈及与前苹果传奇设计主管 Jony Ive 合作打造的新硬件时，主持人追问该设备究竟是便携圆盘（puck）还是智能耳机。Sarah 幽默地打趣道：“我要是承认那是耳机，Jony 可能会冲过来把我十几岁的儿子拐走——虽然我搞不好会同意把儿子送给他。”玩笑过后，她首度正面确认：OpenAI 正在联合打造一种全新的消费级硬件载体，目前原型机已经成型，她本人已经亲自试用过，该设备预计将于今年年底或明年初正式向全球揭晓。
在描述试用感受时，现场提问这是否带来了如同当年初代 iPhone 问世时的震撼颠覆。Sarah 动容地表示，Jony Ive 及其团队最非凡的能力在于“为机械设备注入人性（humanity）”。那种体验无法用冰冷的工业参数去形容，但一旦上手就能直观感受到它极其自然、让人油然而生喜爱之情（lovable）。在顶级设计面前，一切冰冷繁琐的技术细节都会彻底退入幕后，真正诠释了“将至简做到极致才是最难的”（simple is hard）。
---
### [16:01](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=961s) 资本配置与单位经济学：从算力通缩曲线到2030年代前瞻布局
面对 CFO 最核心的资本配置模型问题，Chamath 询问 OpenAI 如何像历史上那些伟大的超级资本配置者一样，找到持续带来超额资本回报率（ROC）的核心引擎。Sarah 剖析了 OpenAI 的商业闭环：持久高价值的商业巨头从不是靠变魔术建立的，其基石永远是为客户创造真实的业务增量。她以生命科学巨头赛默飞世尔（Thermo Fisher）为例，AI 帮助其加速患者筛选流程，从而让新药更快通过 FDA 审批——对于仅剩数周生命的晚期癌症患者而言，研发周期提前两周就意味着生死之别；同时，赛默飞世尔遍布全球实验室的 3.8 万名一线销售人员，也在通过 AI 全面提升拓客效率。有趣的是，在 OpenAI 内部，代码与生产力工具 Codex 普及渗透增速最迅猛的部门并非研发，而是业务拓展与销售（GTM）团队，紧随其后的便是财务部门。
```mermaid
flowchart TD
accTitle: OpenAI 资本配置与单位经济学闭环
accDescr: 展示从客户价值创造、算力成本通缩到远期基础设施前瞻投资的循环飞轮
subgraph ValueCreation ["价值创造与商业化"]
A["创造真实客户价值<br/>(医疗提效/GTM提效)"] --> B["摆脱成本加成定价<br/>转向按价值创造收费"]
end
subgraph Economics ["单位经济学与通缩"]
B --> C["高毛利率模型建立"]
D["算法与硬件效率提升<br/>(模型代际成本降幅达97%)"] --> C
C --> E["Token 综合成本降低<br/>反哺客户与应用普及"]
end
subgraph CapitalSpend ["资本配置与前瞻性基建"]
E --> F["超前资本开支决策"]
F --> G["锁定远期电力与算力<br/>(布局 2028-2032 周期)"]
G --> A
end
```
在毛利率与营业成本层面，Sarah 揭示了算力端令人惊叹的极速通缩曲线。从 GPT-4 到后续迭代版本（如 GPT-4o 系列），单位推理成本在短短两年内出现了高达 97% 的断崖式下降。在最新的前沿模型定价上，尽管 OpenAI 将模型调用单价提高了两倍，但由于模型单 Token 的执行效率与智能密度大幅跃升，终端客户实际承担的任务成本反而享受了 20% 至 30% 的净降幅。
正因如此，CFO 在进行数十亿甚至上百亿美元的资本开支决策时，绝不能静态地参照当下的硬件成本，否则必然会严重误判投资回报。Sarah 直言，她当前大部分精力是在为 2028 年及更长远的算力做采购准备。像密歇根萨林的数据中心，最早也要到 2027 年底或 2028 年初才能真正输出算力；而放眼行业，真正让她感到算力供给极度匮乏、甚至令人夜不能寐的周期，实际上是 2030、2031 以及 2032 年。面对这种极其漫长的基建交付滞后，OpenAI 必须在当下以超常规的魄力提前下注。
### [19:26](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1166s) 算力需求预测：从自下而上建模到外生算力推导
在被问及如何跨越数年预测瞬息万变的算力需求，并权衡单位能耗带来的价值跃迁与技术架构不确定性时，Sarah Friar 详细拆解了财务与算力规划的底层假设。一方面，若按每吉瓦（GW）的基础成本核算，算力本身实际上正在变得更贵，因为电力价格、内存价格等基础设施硬性成本都在上升；但另一方面，由于底层芯片的推陈出新与硬件效能摊销，每单位算力产出的“智能”呈现出爆发式增长，完全抵消并超越了硬件成本的上涨。因此，最终分摊给客户的单位智能服务价格实际上大幅走低。在测算中，财务团队非常克制，不会预先对模型算法本身的能效优化做出过于激进的假设。比如，后续迭代的 GPT-5.5 在推理效率上表现惊人，而更早的超大规模预训练模型（如 GPT-5.4 阶段的探索）不仅极其昂贵，服务部署难度也很大。OpenAI 的策略是先通过大规模预训练确立能力边界，再经过数轮模型架构迭代与精炼，把推理成本彻底打下来。
对于近期的财务测算（如 2026 至 2027 年），Sarah Friar 采用的是非常扎实的自下而上（Bottoms-up）模型：基于清晰的产品矩阵，通过消费端的“单价 × 销量”（$P \times Q$）逻辑，结合周活跃用户数（WAU）、日活跃用户数（DAU）、用户发信量、订阅转化率曲线以及广告收入的潜力量化未来的现金流。这一阶段的增长曲线往往超出预期向上突破。然而，当把视野延伸到更长远的未来几年时，传统自下而上的预测就会失效，规划逻辑将转变为基于所采购算力规模的反向算法——即测算固定规模的算力底座最终应对应多少营业收入。虽然当下无法确切知道每一分钱来自哪个具体产品形态，但商业化的爆发力屡屡击碎市场预期。例如，Sarah Friar 曾在一年前向投资人展示了一套“智能体（Agentic）收入模型”，设想在智能体时代将自然语言编程工具交付给开发者，预估开发者愿意为此支付高达每月 2000 美元的费用。当时这套模型被投资人视作无稽之谈，甚至遭到嘲笑，“人们当时甚至觉得 ChatGPT Pro 收 200 美元一个月就已经疯了，断言绝不可能有人为此掏钱。”
- 算力成本构成的二元性：电力与内存等物理成本上升，但芯片代际进化带来的智能吞吐量大幅跑赢成本涨幅。
- 预测周期的双轨机制：近期（2026-2027）依靠产品线、订阅率与消息量的自下而上精细建模；远期依靠“算力采购规模 → 对应营收潜力”的反向推导。
- 定价认知的突破：从最初市场质疑 200 美元/月的 ChatGPT Pro，到如今测算每月高达 2000 美元的开发者智能体套件。
---
### [22:30](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1350s) 算力战略与资本架构：魔方模型与多云/多芯片布局
当讨论深入到具体资本开支时，主持人提出业界普遍预估建设 1 吉瓦（GW）AI 算力（涵盖土地、电力外壳、芯片及全套设施）全包成本约为 500 亿美元，那么动辄千亿美元级别的资金到底能支撑多少吉瓦，OpenAI 又是否需要亲自承担全部资本支出？Sarah Friar 回顾道，OpenAI 的算力战略在过去两年发生了翻天覆地的转变。两年前的架构极其单薄脆弱：只有微软 Azure 一个云服务商（CSP）、只依赖英伟达一家芯片、核心产品只有 ChatGPT、单一定价 20 美元/月。Sarah Friar 常用“魔方（Rubik's Cube）”做比喻，当时的 OpenAI 就像是魔方最底层孤零零的一个小色块。
为了支撑如此庞大的算力扩张，OpenAI 采取的核心策略是引入多元云服务商与多芯片体系，将高强度的资本开支（CapEx）转嫁并转化为按需付费的运营开支（OpEx）。通过借助具备投资级信用评级的云巨头进行重资产建设与债务融资，OpenAI 能够随着算力中心的实际使用与收入增长逐步支付费用，极大释放了现金流灵活性。如今，OpenAI 已经同时运行在微软 Azure、甲骨文（Oracle）、CoreWeave、谷歌云（GCP）、亚马逊 AWS 以及多家小型新型云平台之上；在芯片层面上，英伟达仍然是最核心的先锋合作伙伴，秋季的大规模训练将运行在 Vera Rubin 架构上，后续还将跟进 Feynman 系列，但与此同时 AMD 芯片已被列入交付管线，Cerebras 芯片已上线运行并为实时代码编写提供极低延迟支持，此外还在与博通（Broadcom）联合研发自研定制芯片。此外，OpenAI 也开始
### [29:09](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1749s) 广告商业化潜力：Google 与 Meta 的基因结合
主持人在节目的快速问答环节抛出了最后一个核心商业议题：回顾当代商业史上最伟大的三大消费级商业模式——iPhone、Meta 广告网络和 Google 广告网络，其中两家完全建立在广告业务之上，即便苹果也在逐步渗透广告生态。然而，OpenAI 极少公开讨论广告战略。针对外界关于免费版 ChatGPT 进行广告测试的传闻，以及竞争对手 Anthropic 在“超级碗”期间对 OpenAI 植入广告的嘲弄，主持人尖锐地提问：广告是否正是 OpenAI 实现向全球用户免费普及 AI 的终极解法？
OpenAI 首席财务官 Sarah Friar 首先阐明了公司在广告业务上的底线原则：必须始终确保用户获得的是模型计算出的最佳答案，而非受赞助商利益操纵的内容；同时，OpenAI 承诺永远会保留无广告的付费订阅版本。随后，她引用了董事会成员 Fidji Simo 的生动比喻指出：“如果 Google 和 Meta 生了一个孩子，那它就是 ChatGPT。”在 Google 方面，用户搜索展现出极其明确的高意向性（High Intent）——例如直接表达“我想买双帅气的鞋坐在舞台上”——这种意图对广告主而言极具商业价值。按统计口径，OpenAI 目前已占据至少 11% 的搜索市场份额，而实际使用量远超数字本身，因为传统 Google 搜索将每一次页面刷新都计为一次查询，但在 ChatGPT 中，一段包含多达 50 个来回问答的深度会话在统计上也仅被算作一次。
而在 Meta 的维度上，其广告优势在于利用人口统计学标签和“相似人群”画像进行推测，ChatGPT 则在此基础上进一步叠加了强大的“长期记忆与上下文能力”。当高意图搜索、个性化记忆与深厚上下文结合在一起时，将诞生一个前所未有的超高转化率广告平台。这种变现能力可以有效覆盖高昂的算力成本，从而真正实现让全球广大公众大规模免费使用顶尖 AI。
```mermaid
flowchart TD
accTitle: ChatGPT 广告潜力架构
accDescr: 展示 ChatGPT 融合 Google 搜索意图与 Meta 用户画像并叠加模型记忆能力的商业逻辑
subgraph Inputs["核心优势融合"]
A["Google 基因：高意图搜索需求"]
B["Meta 基因：精准用户画像与群体特征"]
C["OpenAI 独有：长期记忆与深度对话上下文"]
end
subgraph Platform["ChatGPT 商业化引擎"]
D["超高精度广告投放与意图匹配"]
end
subgraph Impact["商业与社会价值"]
E["支撑高额算力成本支出"]
F["向全球公众大规模开放免费 AI 服务"]
end
A --> D
B --> D
C --> D
D --> E
E --> F
```
- **搜索市场份额真实占比**：虽然名义统计已占有至少 11% 的搜索市场，但传统搜索引擎将单次刷新计为一次搜索，而 ChatGPT 将数十轮的交互只算作一次，实际用户意图交互频次远高于表观数据。
- **两大不可妥协的广告准则**：模型的输出结果必须保持客观纯粹，绝对不按赞助金额排序输出；始终为用户提供纯净无广告的付费层级选项。
---
### [31:07](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1867s) Token 算力分配抉择与 AI 公共基础设施战略
谈到商业变现的最优解时，Sarah Friar 回应了主持团队中关于“单 Token 营收产出比”的质疑。她坦言，如果 OpenAI 的目标仅仅是短期利润最大化，财务上最优的策略就是“把每一个 Token 都分配给 API 接口”。相较于消费级终端产品，面向开发者的 API 业务在单 Token 变现效率上具有数量级层面的优势。
然而，Sarah Friar 强调 OpenAI 正在“下属于自己的下一盘大棋”（playing our own game）。公司的长期愿景是成为整个社会的 AI 基础设施层，就像自来水和电力等公共事业一样无处不在。为了实现这一未来蓝图，OpenAI 必须构建多层次的生态体系，不仅要服务于大型企业和开发者，更要覆盖全球普通消费者、微小型企业乃至各级政府机构。正是基于这种超越当下财务报表的战略定力，OpenAI 选择在消费端持续重金投入。随着全场的热烈掌声，访谈在对 OpenAI 宏大基础设施构想的赞叹中圆满落幕。
- **Token 变现效率权衡**：API 业务的单 Token 创收能力比消费端高出一个数量级，但过度追求短期变现会偏离普及 AI 的长期使命。
- **终局定位**：将 AI 视为类似电力的现代社会通用基础设施（AI infrastructure layer），全面渗透并赋能全社会各个阶层。
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=0s) IPO 仅仅是一个‘里程碑（Milestone）’，绝非企业的‘终点站（Destination）’，绝不能把上市当作某种战略归宿来运营公司。 — Sarah Friar
> [00:00](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=0s) 市场短期是一台投票机，长期是一台称重机。 — Sarah Friar
> [07:42](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=462s) 即便到了 2026 年，算力也依旧会不够用。 — Sarah Friar
> [12:03](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=723s) 在 2026 年如果有谁想去市场上买到更多现成算力，那我只能祝你好运，因为连我都不知道上哪儿去弄。 — Sarah Friar
> [14:43](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=883s) 用大拇指打字沟通已经成了一种疾病。 — Sarah Friar
> [14:43](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=883s) 我要是承认那是耳机，Jony 可能会冲过来把我十几岁的儿子拐走——虽然我搞不好会同意把儿子送给他。 — Sarah Friar
> [14:43](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=883s) 将至简做到极致才是最难的（simple is hard）。 — Sarah Friar
> [19:26](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1166s) 人们当时甚至觉得 ChatGPT Pro 收 200 美元一个月就已经疯了，断言绝不可能有人为此掏钱。 — Sarah Friar
> [29:09](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1749s) 如果 Google 和 Meta 生了一个孩子，那它就是 ChatGPT。 — Fidji Simo
> [31:07](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1867s) 如果 OpenAI 的目标仅仅是短期利润最大化，财务上最优的策略就是把每一个 Token 都分配给 API 接口。 — Sarah Friar
> [31:07](https://www.youtube.com/watch?v=TjrShuj_Zsg&t=1867s) OpenAI 正在下属于自己的下一盘大棋（playing our own game）。 — Sarah Friar
## Mentioned
Sarah Friar, David Sacks, Jason Calacanis, Chamath Palihapitiya, Sam Altman, Greg Brockman, Denise Dresser, Jony Ive, Fidji Simo, 埃隆·马斯克 (Elon Musk), OpenAI, Anthropic, Google, Meta, SpaceX, 沙特阿美 (Saudi Aramco), SEC, 甲骨文 (Oracle), Nextdoor, 微软 Azure (Microsoft Azure), CoreWeave, 亚马逊 AWS (Amazon AWS), 谷歌云 (GCP), 英伟达 (NVIDIA), AMD, Cerebras, 博通 (Broadcom), 赛默飞世尔 (Thermo Fisher Scientific), 旅行者保险 (Travelers), ChatGPT, Codex, Frontier, Sora, iPhone, GPT-4, GPT-4o, GPT-5.4, GPT-5.5, Vera Rubin, Feynman
## Video description
(0:00) OpenAI CFO Sarah Friar joins the show!
(0:31) How OpenAI thinks about its IPO timeline
(3:31) OpenAI, Anthropic, Google: The AI arms race
(7:43) Navigating the compute crunch and AI bottlenecks, device preview!
(15:53) OpenAI's economics
(26:08) Push into chips, the cloud
(29:32) OpenAI's ad business and strategy
Thanks to our partners for making this possible!
EY - Agentic AI is introducing a new investment discipline. As AI shifts to consumption-based models, EY connects spend to enterprise value.
https://www.ey.com/en_us/insights/ai/agentic-ai-token-costs?WT.mc_id=3501318&AA.tsrc=sponsorship
NYSE - Thank you to our partner, the New York Stock Exchange - a modern marketplace and exchange for building the future. It all happens at the NYSE.
https://www.nyse.com
Plaud - Never miss a moment. Plaud, our official wearable AI note-taking partner at All-In Liquidity Summit, captured every insight.
https://www.plaud.ai
Follow Sarah Friar:
https://x.com/thefriley
Apply for Summit 2026:
https://allin.com/events
Follow the besties:
https://x.com/chamath
https://x.com/Jason
https://x.com/DavidSacks
https://x.com/friedberg
Follow on X:
https://x.com/theallinpod
Follow on Instagram:
https://www.instagram.com/theallinpod
Follow on TikTok:
https://www.tiktok.com/@theallinpod
Follow on LinkedIn:
https://www.linkedin.com/company/allinpod
Intro Music Credit:
https://rb.gy/tppkzl
https://x.com/yung_spielburg
[Source](https://www.youtube.com/watch?v=TjrShuj_Zsg)
Source quotes
[0:00] Sarah Friar: IPO 仅仅是一个‘里程碑（Milestone）’，绝非企业的‘终点站（Destination）’，绝不能把上市当作某种战略归宿来运营公司。
[0:00] Sarah Friar: 市场短期是一台投票机，长期是一台称重机。
[7:42] Sarah Friar: 即便到了 2026 年，算力也依旧会不够用。
[12:03] Sarah Friar: 在 2026 年如果有谁想去市场上买到更多现成算力，那我只能祝你好运，因为连我都不知道上哪儿去弄。
[14:43] Sarah Friar: 用大拇指打字沟通已经成了一种疾病。
[14:43] Sarah Friar: 我要是承认那是耳机，Jony 可能会冲过来把我十几岁的儿子拐走——虽然我搞不好会同意把儿子送给他。
[14:43] Sarah Friar: 将至简做到极致才是最难的（simple is hard）。
[19:26] Sarah Friar: 人们当时甚至觉得 ChatGPT Pro 收 200 美元一个月就已经疯了，断言绝不可能有人为此掏钱。
[29:09] Fidji Simo: 如果 Google 和 Meta 生了一个孩子，那它就是 ChatGPT。
[31:07] Sarah Friar: 如果 OpenAI 的目标仅仅是短期利润最大化，财务上最优的策略就是把每一个 Token 都分配给 API 接口。
[31:07] Sarah Friar: OpenAI 正在下属于自己的下一盘大棋（playing our own game）。
Source quotes
[0:00] Sarah Friar: An IPO is a milestone, not a destination.
[0:00] Sarah Friar: The market is a weighing machine, not a popularity machine.
[2:53] Sarah Friar: ChatGPT has become both the noun and the verb for how people experience AI.
[5:49] Sarah Friar: We cannot leave the consumer behind in our mission of AGI for the benefit of humanity.
[7:42] Sarah Friar: The demand curve for compute is a vertical wall.
[12:03] Sarah Friar: Jony will come and steal my teenage son if I say too much.
[16:07] Sarah Friar: The difference between a breakthrough in four weeks and two weeks can literally be life or death.
[16:07] Sarah Friar: Making capital allocation decisions based on today's cost profile risks mispricing the outcome.
[29:04] Sarah Friar: If Google and Meta had a baby, it would be ChatGPT.
[29:04] Sarah Friar: If I were optimizing only for short-term revenue, I would route every token to the API.
Original transcript
[0:00] Open AAI's CFO, Sarah Frier.
[0:02] >> We got to get right to it. You have just
[0:04] completed what I [music] regard as the
[0:05] most successful fundraising round in
[0:08] history.
[0:10] >> We're going to raise [music] actually
[0:11] north of $120 billion. We think AI is
[0:15] the biggest era that we've [music] seen
[0:16] to date. We're just starting to
[0:18] understand what it's going to mean for
[0:19] global productivity and with that,
[0:21] [music] you know, hopefully more
[0:22] affluence, better lives for everyone.
[0:25] Luck is whatever the preparation [music]
[0:27] meets opportunity, but you got to grab
[0:29] it.
[0:31] >> Longtime listener, first time caller.
[0:34] Quite exciting to get to hang out with
[0:37] all the bros here. Hello. [laughter]
[0:39] >> We weren't sure how to start this off,
[0:41] but I thought the best thing was to
[0:42] allow our URSTW cryptosar to maybe
[0:46] >> save comments. I saw an article today, I
[0:48] think it might have been in the Wall
[0:49] Street Journal, that the perception is
[0:52] that there's an advantage to IPOing
[0:55] earlier if you're an AI company. So now
[0:58] we know SpaceX is going and then the
[1:00] question is when's when are open AI and
[1:03] Anthropic going to go? And I'm curious,
[1:06] how do you think about that? Do you
[1:08] think there is a little bit of a race on
[1:10] or you know, you haven't made a decision
[1:11] about that yet? Like in the end, an IPO,
[1:14] I say this to the team all the time,
[1:16] it's a milestone. It is not a
[1:18] destination. Do not run your company as
[1:21] if that's some sort of destination. It's
[1:23] just another way to fund raise. We just
[1:25] did, you heard me on on the the s the
[1:27] sizzle reel, raise 122 billion dollars
[1:30] in March, and that was to give ourselves
[1:33] maximum flexibility. I feel like my job
[1:36] as a CFO is create optionality for this
[1:41] not just this company but just this era
[1:44] that we're living in.
[1:44] >> So Sarah was that that point in
[1:47] fundraising is that the biggest private
[1:49] republic up until the SpaceX IPO?
[1:51] >> It is.
[1:52] >> It is by orders of magnitude. I think
[1:54] the largest IPO to date was the Sidio
[1:56] Ramco which was about $30 billion. So it
[1:59] is actually incredible that you're going
[2:00] to have potentially three IPOs at a
[2:03] scale that will be bigger even than 2001
[2:07] 2000 that that time frame there was a
[2:09] lot that went on in the market too but
[2:11] the market has grown and by the way the
[2:13] other thing going on in the market is
[2:14] like if you look at buybacks M&A and so
[2:17] on it's actually a lot of capital keeps
[2:19] being returned back to shareholders and
[2:21] cash so there is a lot of money sitting
[2:23] on the sidelines but in the spirit of
[2:26] like the question David, I think in the
[2:30] end you want to you'll be measured,
[2:32] right? It's the in the end the market is
[2:34] a weighing machine, not a popularity
[2:35] machine. No one remembers who went
[2:37] first, Google or Yahoo, Lyft or Uber.
[2:40] And I say that not because whether I
[2:41] want to be first or second, but I just
[2:43] think it, you know, the the press loves
[2:45] a bit of drama, but in the end, we're
[2:47] going to have to build big, sustainable,
[2:49] durable companies, and fundraising will
[2:51] be a key component of doing exactly
[2:53] that.
[2:53] >> Sarah, breaking news.
[2:55] >> Oh my god, so many people coming at me.
[2:57] Hi, [laughter] Jason. No, it is it is it
[2:59] is hard balancing four interviewers at
[3:01] the same time.
[3:01] >> It's okay. This is my world, by the way.
[3:04] So, I'm good with this. Jason,
[3:05] >> Anthropic just uh confidentially filed
[3:09] their S1. So, does that mean you're
[3:11] third place in terms of the filing?
[3:13] >> It does not mean anything yet because
[3:15] [laughter] you have to run now the
[3:17] gauntlet of the SEC and who knows how
[3:19] long that takes for anyone.
[3:21] >> Yeah. Is it is there though a benefit to
[3:24] them going forward? And I think
[3:26] unpacking the rivalry with Anthropic is
[3:29] on everybody's minds. So just I guess
[3:32] you can't talk too much about IPOs. So
[3:34] I'll just pivot to Anthropic was far
[3:38] behind and now they've really um I think
[3:41] everybody would agree in the industry
[3:43] now blown past OpenAI in terms of
[3:45] developers and corporations and it seems
[3:48] revenue. So did how did that happen at
[3:52] OpenAI when you had such a tremendous
[3:53] lead? How did Anthropic blow past you
[3:57] guys?
[3:57] >> So let's talk a little bit about a
[3:59] strategy. Our strategy is different,
[4:01] right? So we are building the AI layer,
[4:04] the infrastructure and it's really
[4:06] important that there's a single
[4:07] foundation but then with many interfaces
[4:10] out into the world. So chat GPT is one
[4:13] to the consumer. Over 900 million people
[4:16] use Chat GPT weekly and it's become the
[4:19] noun and the verb. It's how most people
[4:21] experience um AI for the first time.
[4:24] Kind of fun fact, our economic research
[4:26] team just showed me um the fastest
[4:28] growing continents now are Africa.
[4:30] Probably not totally surprising since it
[4:32] started a small base. fastest growing
[4:34] languages are um Azerbjani and um what
[4:39] Kazakhstani what is it's Kazak um which
[4:42] it's kind of incredible to talk about
[4:44] where it's going so multiple interfaces
[4:46] chatgpt of course there's um codeex um
[4:49] just hit 5 million over the weekend
[4:51] we're really proud of that coming from
[4:53] almost zero in January 5 million users
[4:55] go codeex um help me prepare for this
[4:58] little special up here too um there's of
[5:01] course Frontier our enterprise offering
[5:03] ing and everything every other way that
[5:05] we can get out there to reach businesses
[5:07] of all sizes. That is a very different
[5:09] strategy. We think that because it's
[5:11] served up on one model, there's a
[5:13] compounding element of advantage that
[5:16] comes from that. More users, more data,
[5:18] more ability to personalize chats as a
[5:22] front door. As we as models get bigger,
[5:25] there's more efficiency. That should
[5:26] lower the overall cost to give you a
[5:28] token in the world. That should compound
[5:30] to higher gross margins, ultimately more
[5:33] ways to pay for compute, and then access
[5:35] to compute is one of the really big
[5:37] competitive advantages at the moment.
[5:39] So, you know, we have to all run our own
[5:41] races, but we all have to recognize
[5:43] we're part of an ecosystem that also
[5:46] needs to bring people along
[5:48] collectively.
[5:49] >> Did you spread a little bit too then too
[5:50] many projects? People were talking about
[5:52] this new gadget, Sora, and and then
[5:55] maybe not enough focus on enterprise. Is
[5:57] that a fair assessment of if there was a
[5:59] mistake in the last year that was it?
[6:01] >> No, I I think that the world loves to go
[6:03] to binaryisms like are you a consumer
[6:06] company Sarah? Are you an enterprise
[6:07] company? The reality is we're very much
[6:10] both. We're not one or the other. Right
[6:12] now our revenue is getting pretty
[6:14] balanced about 50/50. We are incredibly
[6:17] focused on the enterprise. Like I spend
[6:19] so much of my time with I mean just even
[6:22] in the last week I could tell you I've
[6:23] been to see Thermopisher in Boston. I
[6:25] was with a bunch of banks in New York. I
[6:28] was on the phone with travelers on
[6:30] Friday. I spent this morning on the
[6:32] phone with a tech company. It doesn't
[6:34] matter the vertical. People are really
[6:36] moving on AI right now. Our new head of
[6:38] revenue, Denise Dresser, in seats since
[6:40] December. She is a force of nature. And
[6:43] so I think the enterprise broadly
[6:45] speaking is really firing on all
[6:47] cylinders. But we don't want to leave
[6:49] the consumer behind. Remember our
[6:50] mission at OpenAI is AGI for the benefit
[6:54] of humanity, not for the benefit of
[6:56] humanity who can pay or for the benefit
[6:58] of humanity who live in an enterprise,
[7:00] but very broadbased. Um, it's why we
[7:04] offer so much free because we want
[7:06] people to get a taste. Once they get a
[7:09] taste of intelligence, the ability to
[7:11] come up a commitment curve is
[7:12] incredible. Our free users do about
[7:14] seven turns, seven questions a day. Our
[7:17] first paid tier do double that about 15.
[7:20] Our our real paid tier the plus 20
[7:23] bucks. Hopefully you're all on it or
[7:24] higher about 3x and pro about um 11x
[7:30] over a free user. So remember when you
[7:32] got your flip phone and you're like yeah
[7:34] I don't know what it does make some
[7:35] calls. Now that same phone think of all
[7:38] the things it does for you. That's the
[7:40] path we're on with intelligence right
[7:41] now. Sorry.
[7:42] >> You said something very influential. I
[7:44] think it was about 18 months ago for a
[7:46] lot of us in the industry where you
[7:48] framed a very simple economic trade-off
[7:50] which was gigawatts to cash and I think
[7:53] you said one gigawatt is roughly
[7:55] equivalent to about $10 billion a year
[7:58] of revenue to open AAI. So common number
[8:01] one was this one gigawatt equals10
[8:03] billion a year of revenue for you. But
[8:05] it's not just you because you can
[8:07] probably extrapolate that to anthropic
[8:09] and other folks Gemini
[8:11] >> but then you were really at the
[8:12] forefront of getting access to power and
[8:14] data centers and powered land. It seemed
[8:17] a little crazy but now it looks like
[8:19] hold on there's a huge deficit of
[8:21] supply. Can you just unpack all of that
[8:23] and explain
[8:25] >> both the spectrum of where we are and
[8:27] then those specific economics and if
[8:28] that's changed. So first of all, yes,
[8:31] compute is a very scarce resource at the
[8:34] moment. I mean, what we see in our
[8:35] business, we're going up that kind of
[8:37] vertical wall of demand right now and
[8:39] there's just not enough tokens
[8:40] available. So we I'm very grateful that
[8:44] I got to work alongside Greg and Sam. I
[8:47] think we're very precient on this and
[8:49] last year we were definitely taking some
[8:51] you know arrows in the back about why
[8:53] are they out there buying all this
[8:54] compute and I think thank god we did
[8:56] because in 26 we still won't have enough
[8:59] compute. Um where are we on the compute
[9:02] continuum there's kind of choke points
[9:04] everywhere and and I think they will
[9:06] continue to move back and forth. I mean
[9:07] you all talk about this and know this as
[9:10] well as anyone um here whether it's
[9:13] energy first and foremost um land power
[9:16] we get regulatory um environment such
[9:19] that we can build quickly um when you
[9:22] get into the racks and chips themselves
[9:24] clearly do we have enough um in that
[9:26] supply chain memory spike is is on at
[9:29] the moment access to great talent um do
[9:32] we have enough people coming through our
[9:33] education system I really worry about
[9:35] this right now I'm a trustee at Stanford
[9:38] and you know I see just that you know we
[9:40] need to keep the focus on education and
[9:43] science um and then trust I mean I
[9:45] actually put that as part of the supply
[9:47] chain um Sam right now is in Selen
[9:50] Michigan he's going to be cutting the
[9:51] ribbon in about two hours so you are
[9:53] getting a sneak preview but they told me
[9:55] it was okay to say it in the room um
[9:57] that will be you know sticking shovels
[9:59] in the ground on a 1 gigawatt data
[10:01] center which is part of our or Oracle
[10:03] complex really important there on the
[10:05] trust side that we don't leave
[10:07] communities behind. I spent seven years
[10:10] of my life working at Next Door doing
[10:12] the hard work of what it means to be
[10:14] local and you cannot tell people from
[10:17] top down what they need because they
[10:19] will tell you thank you but no thank
[10:20] you. I will tell you what I need. And so
[10:23] in a data center like that actually
[10:25] spending a lot of time in the community
[10:26] saying number one, we're not going to
[10:28] raise your electricity bills. We're
[10:30] going to pay for our infrastructure and
[10:31] our power. It will not be the rateayer
[10:33] that has to pay. Number two, we're going
[10:35] to bring jobs, 2500 union jobs, good
[10:39] jobs like electricians, HVAC, and so on.
[10:43] We are going to pay our taxes, a billion
[10:46] dollars in taxes just for that data
[10:48] center into Michigan. And on top of
[10:50] that, we're going to invest $45 million
[10:53] going into education for codeex credits
[10:56] to do what you all talked about this
[10:57] weekend is like anyone who's not like
[11:00] coming in filled to their new job. I
[11:02] have teenagers using Codeex. It would be
[11:04] like I would never hire a finance person
[11:06] didn't know how to use Excel. And I
[11:08] pretty much probably wouldn't hire a
[11:09] finance person today that doesn't know
[11:11] how to use a tool like Codeex. Um so
[11:14] that you know so when I think about
[11:16] investment we're having to invest ahead
[11:18] of demand. That means we need to both be
[11:20] able to find all of the compute and all
[11:23] the pieces and then pay for it. So that
[11:25] goes back to your capital question on
[11:26] IPO. And then on the other side on the
[11:29] economics, look, the economics do
[11:30] continue to get better. They're getting
[11:32] better on multiple fronts. I think we
[11:34] are doing a better job of actually
[11:36] showing true value to our customers. And
[11:38] I think you get beyond kind of a cost
[11:40] plus type pricing into something that
[11:42] feels more akin to the value being
[11:45] created. Now, scarcity of tokens helps
[11:47] because it's causing a bit of a
[11:49] compression. Can you talk about that in
[11:50] just like without specific names where
[11:54] you know the landscape exists today in
[11:57] terms of all the power that's available
[11:59] and all the demand that exists across
[12:01] everybody. Yep.
[12:03] >> What's going to happen over the next
[12:04] year just at the current course and
[12:06] speed of what is available of the data
[12:08] centers that's available of the tokens
[12:10] that's available of the infrastructure
[12:11] that's available for everybody because
[12:13] you know I told this story last week but
[12:16] you know I'll use anthropic and one of
[12:17] the frustrating things is at some point
[12:18] it just says you know 10:30 it's like
[12:21] all right Jimoth see you at 2:30.
[12:22] >> Yeah.
[12:22] >> And that's not a viable experience
[12:25] >> right. Um
[12:26] >> and in fairness to chat GPT actually
[12:28] I've never had that with
[12:29] >> Yeah. We we're quite generous with our
[12:32] tokens and again on purpose we're trying
[12:34] to drive access so people understand
[12:37] because if you're on that free tier not
[12:39] actually getting the latest model but
[12:40] we're trying to put it in your hands so
[12:42] you get a sense for it by the way
[12:44] because you know if you're a kid um
[12:47] doing homework like I think about when I
[12:49] grew up and the encyclopedia bratannicas
[12:52] showed up at the front door in Northern
[12:54] Ireland in a tiny little community in
[12:56] the middle of the troubles. It was like
[12:57] the clouds parted and so we want to make
[12:59] sure that people get that feeling by the
[13:02] way but the landscape right now in 26 if
[13:05] you want to buy more compute good luck
[13:07] to you like tell me cuz I don't know
[13:09] where else to find it. I mean as you
[13:10] know
[13:11] >> well I was going to say Elon ironically
[13:13] ended up being the one person that had
[13:16] too much compute in a way but good job
[13:19] on like figuring out how to sell that
[13:21] off. Um in 27 it's pretty limited as
[13:25] well frankly. Now there's a couple of
[13:27] things shifting around when we talk
[13:28] about compute. There's training that
[13:31] mostly still all happens here in the
[13:34] United States for USG reasons for making
[13:37] sure that a national asset and effect is
[13:39] happening on US soil. For inference, we
[13:42] want that to be global. And I think
[13:44] particularly in an agentic world, you
[13:46] want much more kind of real time. Even
[13:48] for things like Sora and video, which by
[13:50] the way, yeah, we have, you know, we had
[13:52] to make a really tough choice because we
[13:54] didn't have enough compute and we said a
[13:56] lot
[13:57] >> right now. Yeah, video does, but video
[13:59] is not over. Like in particular, when
[14:02] you start to think about where AI is
[14:04] taking us into more multimodality. So
[14:07] remember, we've all been taught by the
[14:10] last generation of technology to talk
[14:12] with our thumbs. It's a disease. You
[14:15] walk around, everyone's looking down.
[14:16] They don't look up anymore. Teenagers
[14:18] sit on my sofa at night and talk to each
[14:21] other with their thumbs. I'm like, who
[14:22] are you talking to? And my son will be
[14:24] like him. I'm like, okay, talk.
[14:27] Multimodality is here. Um hopefully I
[14:30] think you all talked about it this
[14:32] weekend. You're talking to your tool. I
[14:33] talk to codeex every day. And so that is
[14:36] changing rapidly, but that is going to
[14:38] need much more kind of real time compute
[14:40] because it's an odd experience if I was
[14:42] talking to
[14:43] >> Johnny I of this puck these earpieces.
[14:46] So maybe tell us a little bit about that
[14:47] problem if you've admitted it now.
[14:49] >> If I if I tell you it's an earpiece,
[14:50] Johnny will come and steal my teenage
[14:52] son. I might give it to him. give them
[14:54] to
[14:54] >> but you do believe that there should be
[14:56] some
[14:57] >> we're changing into a consumer substrate
[15:00] that I cannot tell you what it is but by
[15:02] the end of this year we will unveil it
[15:04] early next year I have seen it I've
[15:07] tried it I am a hand talker right now
[15:09] I'm sitting on my paradigm
[15:12] >> shift when when yeah when you used it
[15:13] was it like having an iPhone for the
[15:15] first time
[15:16] >> it's very what Johnny and team are
[15:19] really good at is bringing humanity to
[15:23] devices. And I don't really know how to
[15:25] explain that well, but when you see it,
[15:27] you feel it.
[15:28] >> It feels natural in some way.
[15:29] >> It feels very natural, but it feels very
[15:32] lovable
[15:33] >> really.
[15:34] >> And I can't really explain what that
[15:36] emotion is cuz
[15:37] >> intimate in some way in terms of
[15:39] >> technology,
[15:40] >> not taking your phone out and it's it's
[15:42] seamless is what I've heard from people
[15:43] that played with it.
[15:44] >> Technology is very um can be very
[15:46] mechanistic, but we all know great
[15:49] design just makes everything fade away,
[15:52] right? what um at the time you know the
[15:54] simple is hard.
[15:56] >> Yeah.
[15:56] >> But I I think this is a very this story
[15:59] just going back to the earlier question.
[16:00] So putting on
[16:01] >> the CFO hat, help us understand the
[16:05] capital allocation model that you use
[16:07] cuz a lot of businesses over the last
[16:09] decade, two decades that have kind of
[16:11] been these outsized returners have found
[16:14] some unique way to deploy capital at a
[16:16] higher ROC than anyone else. and then
[16:18] you end up plowing all your capital into
[16:20] that higher ROC bucket.
[16:23] >> What is that for you guys? And how do
[16:25] you think about that le that portfolio
[16:27] approach to having more of these kind of
[16:28] big returner shots and is there an
[16:30] engine where that gets better over time?
[16:32] >> There has to be because in the end the
[16:35] durable highvalue companies created in
[16:38] this era, I don't think they're not
[16:40] going to be magical. They're going to
[16:42] look like the great companies of prior
[16:43] eras. They're going to create customer
[16:45] value. starts with a customer um and
[16:49] really helps the customer do something
[16:51] different, better, more revenue, more
[16:54] efficiency, right? Thermmaisher wants to
[16:57] be able to get um patient screening done
[17:00] faster so they get FDA approval faster.
[17:03] That's really important. Like if you
[17:04] have a form of cancer where you have
[17:07] weeks to live, the difference between a
[17:09] breakthrough in four weeks and two weeks
[17:11] can literally be life or death. They
[17:13] also have, I'm going to misquote this,
[17:15] but something like 30,000 38,000 people
[17:19] in the field selling those amazing like
[17:21] if you walk into any lab in the country,
[17:23] you'll just see thermopisher plastered
[17:25] all over every device. Those people want
[17:27] to be more efficient going to work. Like
[17:29] the the fastest takeoff of codecs within
[17:31] OpenAI right now is actually in our go
[17:34] to market team. Our devs are there, but
[17:36] like if you look at the pace of growth
[17:38] kind of month over month, it's all in
[17:39] GTM. So they want more productivity out
[17:42] of their GTM team and of course um
[17:44] they're doing things in areas like
[17:46] finance which I get really excited
[17:47] about. So customer value first from that
[17:50] now you need to get to great gross
[17:52] margin. So how do you get to great gross
[17:54] margin? You're looking at like the cost
[17:55] of revenue. The main input is compute.
[17:58] The good news on compute is that there
[18:00] is a massive deflationary curve on cost
[18:02] right from chat GBT uh five to 54 I
[18:08] think the deprecation of cost was
[18:09] something like 97%. It's like kind of an
[18:12] amazing curve. Actually, I'm slightly
[18:14] from four to 54, it was 97%. But that
[18:18] happened in like two years.
[18:20] >> That's kind of wowing, right?
[18:22] >> Um, even our newest model, if you look
[18:24] at 55 that we just released, we're
[18:26] trying to now translate that back to the
[18:28] customer. So, we actually raise prices
[18:30] on 552x.
[18:32] But if you look at what the cost of the
[18:33] customer is, they're probably still
[18:35] getting a break of about 20 to 30% cost
[18:38] reduction per token because just much
[18:40] more efficient per token. So there's a
[18:42] lot to do in that envelope and and part
[18:45] of making an alloc a capital allocation
[18:47] decision is having to
[18:49] >> if you make it on today's cost profile,
[18:51] >> you actually might mispric the outcome.
[18:54] So you have to lean in a little on the
[18:55] cost profile. And then as we think about
[18:58] like the builds, yeah, you are having to
[19:00] make like really my focus today on
[19:03] compute is what's the compute I can buy
[19:05] for 28 onwards. Like that Michigan data
[19:09] center in Seline, I don't think we will
[19:11] be getting compute out of it until
[19:13] probably end of 27, early 28. So that's
[19:16] where you're starting to make your bets.
[19:18] And in fact, where I feel most short of
[19:20] compute right now is starting to look at
[19:22] 30, 31, 32. So you're having to create a
[19:26] business model. Um now the good news is
[19:28] each year goes by we get more confidence
[19:31] in the build. We're seeing it massively
[19:33] outperform and so that's giving us more
[19:36] and more confidence and the market is
[19:38] coming towards us much more.
[19:40] >> All right. So how are you making the
[19:42] compute need forecast multiple years out
[19:46] accounting for all of the architectural
[19:49] and model advancements that are
[19:50] happening where call it value or utility
[19:53] per unit of power is going up and help
[19:57] us understand how you kind of estimate
[20:00] that given that there's a lot of
[20:02] technology development going on that has
[20:03] a high kind of variance to it.
[20:06] >> Yeah. Yeah. So we we do have to make
[20:08] multiple um assumptions both on the
[20:11] compute itself. So we assume right now
[20:13] that compute actually on a per gigawatt
[20:15] is getting more expensive because power
[20:17] is getting more expensive, memory is
[20:18] getting more expensive and so on.
[20:20] However, the the intelligence that we
[20:23] get on the other side out because of the
[20:26] deprecation on the chip side is is more
[20:28] than making up for that. So in in terms
[20:30] of a per unit sold to a customer, it
[20:32] should actually get a lot less expensive
[20:34] for
[20:34] >> no model improvement in that.
[20:36] That's just
[20:37] >> exactly that's just the chip itself. We
[20:39] don't try to overestimate on the model
[20:41] side because sometimes like 55 is an
[20:44] incredibly good model on the efficiency
[20:46] side, but if you look at something like
[20:48] 54, the prior model, it was a really
[20:50] large pre-trained model. It was very
[20:52] expensive. It was actually hard to
[20:54] serve. And sometimes we want to do that
[20:56] really big pre-train moment and then we
[20:59] take multiple model turns to be able to
[21:01] kind of drive down on the cost side. I
[21:03] mean in the in the near term like in 26
[21:06] and 27 I clearly build a model that's
[21:09] bottoms up. So I know what my products
[21:12] are. I have a sense of what the pricing
[21:14] will be. Um you know P time, you know,
[21:16] consumer P* Q. How many Wows do I think
[21:19] I have? I can see what the shape of the
[21:20] line is. How many of them will subscribe
[21:23] advertising coming in is also still
[21:25] related to how many weekly activives,
[21:28] how many dailies, how many messages and
[21:30] so on. So you can you can do actually a
[21:32] pretty good model job in 26 and 27. That
[21:36] said, the shape of the line keeps taking
[21:38] us by surprise to the upside. When you
[21:40] get into the outer years, you're
[21:42] actually looking more at the compute
[21:43] you've bought and almost just doing an
[21:46] algorithm the other way that's saying
[21:47] this amount of compute should equate
[21:49] somewhat to this amount of revenue. I
[21:51] don't know for certain exactly where it
[21:53] will all come from. Like a year ago, I
[21:55] built a model for investors that showed
[21:57] agentic revenue. And the story was,
[22:00] we're going to have this thing. We're
[22:02] going to be in the agent era. We're
[22:03] going to hand it to a developer with
[22:05] natural language. They're going to be
[22:07] able to build and we think they will pay
[22:10] upwards of maybe $2,000 a month for it,
[22:14] which is kind of laughable in hindsight,
[22:16] but nobody believed. Um, they were like,
[22:18] I don't even know what she's talking
[22:19] about. There's no way that will happen.
[22:21] And $2,000 a month, remember when people
[22:23] were losing their minds over Chat GBT
[22:25] Pro being at $200? Oh my god, no one
[22:29] will ever pay for that.
[22:30] >> Yeah.
[22:30] >> So why 122 billion? Does it take you to
[22:33] 2031, 2032? Like how do you get the
[22:35] calculus on the capital needs as you do
[22:37] that modeling?
[22:38] >> You maybe even more specific. So the
[22:41] estimates I've seen is that to stand up
[22:43] 1 gawatt of AI compute costs about $50
[22:47] billion in cap land, PowerShell chips,
[22:51] everything all in around 50 billion.
[22:54] Do you have to front all of that money
[22:56] when you create a new data center or how
[22:59] much of it do you do? How much of it can
[23:01] you get debt for? Does a 100red billion
[23:04] raise only get you two gigawatts or does
[23:06] it get you five? Like what does it get
[23:08] you?
[23:08] >> It's it's a great question. So if you
[23:10] look at our compute strategy um and it's
[23:13] crazy how fast the world has changed. So
[23:15] just two years ago we were literally one
[23:18] we had one CSP we worked with Microsoft
[23:20] Azure um we we sat on one chip Nvidia we
[23:24] had one product chat GBT one price point
[23:27] $20 a month so I often use a Rubik's
[23:29] cube as kind of my metaphor so we were
[23:31] like one cube in the bottom today if you
[23:34] look at our strategy it's been to go
[23:36] first of all multi multiple CSPs because
[23:39] what CSPs do for us in effect is they
[23:41] shift capex into opex so you pay as you
[23:44] get the revenue so as you're actually
[23:46] utilizing the data centers. So in effect
[23:48] we are writing somewhat on their ability
[23:50] to build and have capex and um
[23:53] financing. So today we sit on top of
[23:55] every CSP Oracle um coreweave um
[23:59] Microsoft GCP AWS and a bunch of small
[24:02] neocalars on the chip side. We've also
[24:05] um gone for a program of being
[24:07] multi-chip um because we want to make
[24:09] sure you're always on the frontier. I
[24:11] think if you're only on one chip,
[24:12] there's just inherently a moment where
[24:14] you can't be on the frontier because
[24:16] there's some leaprogging that happens.
[24:18] So today, Nvidia remains our absolute
[24:20] priority partner. They have the Frontier
[24:22] chip. Our next big trading run in the
[24:24] fall will be done on Ver Rubin. We're
[24:26] really excited about that. And now we're
[24:28] plotting kind of the Fineman series
[24:30] that's coming. But we also now have
[24:32] chips in the pipeline from AMD. Um
[24:35] Cerebrris is already online. It's been
[24:37] an incredible low latency chip. Great
[24:39] for devs, for example, that want
[24:41] real-time coding. And there's our own
[24:43] chip that we're working on with
[24:44] Broadcom. And then beyond that, there's
[24:46] other ways we've diversified. So now,
[24:47] think about that Rubik cube. It's become
[24:49] much more multi-dimensional. And it
[24:52] allows us to effectively utilize
[24:55] investment grade CSPs in order to be
[24:58] able to go fast and push it back to be
[25:00] more opex, not capex. Now, we are
[25:03] starting to shift gears into more of a
[25:05] built to suit type environment. We
[25:08] announced a data center we're building
[25:10] with SoftBank Energy um down in Texas.
[25:13] That's the beginning of something that's
[25:14] beyond a CSP. There's a little bit more
[25:17] capex required there. And then finally,
[25:19] I think as the world progresses,
[25:21] remember we've done all that just in two
[25:22] years. The reason I like a Rubik's cube
[25:24] is again, please chat GBT this, but I
[25:27] think a Rubik's cube has something like
[25:28] a quintilion different um forms it can
[25:31] come up with. And so it just gives us a
[25:33] lot of optionality. So remember what I
[25:35] said, my job is maximum optionality and
[25:39] in a moment where I'm not yet an
[25:40] investment grade type of entity where I
[25:43] can go get lowerc cost debt financing,
[25:45] being able to work with partners to do
[25:47] that is really important.
[25:48] >> So do you think that in 5 years from now
[25:50] the stack is just merged together? What
[25:53] do I mean? In traditional or historical
[25:56] markets, you'd have Nvidia sell the
[25:57] chips, but that's all they do. And then
[25:59] you'd have, you know, Microsoft just run
[26:01] a cloud. That's all they would do. And
[26:03] then you would have a consumer app.
[26:05] That's all you would do. But now we see
[26:07] everybody doing everything. You know,
[26:08] you guys have silicon that you're
[26:10] spinning. You have models that you make.
[26:13] You may or may not eventually decide
[26:14] that you need to be some form of a
[26:16] NeoCloud yourself. If you look at
[26:18] Nvidia, they have incredible silicon,
[26:20] but they also have their own open source
[26:21] models. They're increasingly becoming an
[26:23] offtaker. Google is a cloud company
[26:26] first, but they also have a chip. Now
[26:27] they have models. So it's all merging.
[26:32] is if that continues to happen, does
[26:34] that make the competitive landscape
[26:36] simpler or easier?
[26:38] >> I mean, I think where everyone is trying
[26:40] to make sure they reside is the layer
[26:42] that is closest to the customer where
[26:45] usually you take the the largest portion
[26:48] of the profits of the ecosystem, right?
[26:50] No one wants to find themselves
[26:54] away. Absolutely. And so that's why
[26:57] today when I think about our position
[26:59] and comes back to where I started why we
[27:01] want to be that AI intelligence layer is
[27:05] because a year ago people talked about
[27:07] the commoditization of the LLMs. Um and
[27:10] frankly it's gone the opposite because
[27:13] as you start building an agentic layer
[27:15] and we've all started to use this word
[27:16] harness but the harness is what brings
[27:19] the context the memory right it I I have
[27:23] in my codeex I have a whole ginormous
[27:26] memory file where it knows that I'm me
[27:29] it knows I'm the CFO of openai it knows
[27:31] how I like to write things how I like to
[27:33] say things it knows what I'm interested
[27:35] in it actually also knows that I'm a mom
[27:38] teenagers I mean it just carries all
[27:40] this memory and that makes the model
[27:43] more powerful for me. Now think about
[27:45] what happens when that memory in that
[27:46] context is brought into an actual
[27:49] enterprise environment. So now it's not
[27:51] just even about the data that resides
[27:54] there, but I always think about the the
[27:56] intuition of like back when I worked on
[27:57] Wall Street, right? There was all the
[27:59] data in the world that told you what a
[28:01] stock should do post an earnings call.
[28:04] But give me one second. Then you called
[28:07] your trader and the trader be like,
[28:09] "Yeah, stock's not going up, Sarah." And
[28:11] I'm like, "What are you talking about?
[28:12] Like all the numbers say it did this, it
[28:14] did this, it did this." He's like,
[28:14] "Yeah, I know, but I know this fund is
[28:17] under pressure and they need to sell
[28:19] down their book and that is going to
[28:21] kill the stock for the next week."
[28:23] Right? That is the intuition
[28:25] of an enterprise. Like it's the best
[28:27] example I always think of because I came
[28:29] out of a financing world, but there's
[28:30] this intuition in every walk of life.
[28:34] And that's where I think the models are
[28:35] now getting very connected to the memory
[28:38] and context and intuition of your
[28:40] company. And that's what gets CEOs and
[28:43] seuite really excited because they're
[28:45] like okay now I really see how this is
[28:47] going to add value to drive my revenue
[28:49] line my top line but also you know I can
[28:52] think about as an efficiency play as
[28:54] well. And so back to what you're asking,
[28:56] I think what people want to make sure is
[28:57] they stay as close to that value as
[28:59] possible
[29:00] >> and be flexible enough to pivot as you
[29:03] have to rap.
[29:04] >> Yeah.
[29:04] >> But
[29:04] >> sorry Jason,
[29:05] >> it's quite all right. Um it's been
[29:07] wonderful and and you've been so great
[29:09] with the details. One final detail
[29:11] question. Rapid fire. Three great
[29:13] greatest consumer businesses of our
[29:15] lifetime. iPhone, Meta Advertising
[29:18] Network, and Google's advertising
[29:20] network. Two of those three are ad based
[29:21] and and even Apple has a sprinkling of
[29:23] ads.
[29:24] >> Haven't heard you talk about ads much.
[29:26] People tell me they're seeing some ads
[29:28] in the experiment in the free version.
[29:30] What is your commitment to the ad
[29:32] version? You guys got a little uh
[29:34] trolled by Anthropic during the Super
[29:37] Bowl? Oh, you're going to have ads. But
[29:38] is ads the solution to making this free
[29:41] for the world?
[29:42] >> Yeah. So, first of all, on the ad front,
[29:45] you know, we we want to stick by our
[29:46] principles. We want to make sure that
[29:48] you know you're always getting the best
[29:50] result based on the model, not by
[29:52] something that was sponsored. So that
[29:53] has to hold true. And I think the second
[29:56] thing is that we'll always provide a
[29:58] free a tier, sorry, an ad free tier for
[30:00] people that just don't want ads. But
[30:02] with that said, if you took if you took
[30:06] Fiji says this really well, if you know
[30:09] Google and Meta had a baby, it would be
[30:11] chat GPT because what you have in Google
[30:13] search and by the way we know we have at
[30:16] least 11% of the search market. It's a
[30:18] lot more because actually when you do a
[30:20] Google search and the page refreshes
[30:22] that counts as one in chat GBT when you
[30:25] do a whole conversation where you might
[30:26] ask 50 questions that also only counts
[30:28] as one. So in reality, we have a much
[30:30] higher portion, very high intent. That
[30:33] is great for advertisers because I'm
[30:35] effectively telling you what I'm doing,
[30:36] right? I want really cool shoes to sit
[30:39] on the stage. I'm telling you what I
[30:41] want to go by. In Meta's case, right,
[30:43] they use this like people like you sort
[30:46] of intent. So they have the demographic.
[30:48] We have more than that because we have
[30:50] memory, right? I just told you it knows
[30:52] who I am. So imagine putting memory and
[30:54] context next to intent. You should have
[30:56] a very potent ad platform which gives
[30:59] you an ability to offer up massive
[31:02] access to the world writ large because
[31:05] now you can pay for it. And I think back
[31:07] to a question you asked Freeberg like if
[31:09] you look at um the revenue per token
[31:12] right now. If I was optimizing only for
[31:15] today, I would give every token to the
[31:17] API,
[31:17] >> right?
[31:18] >> Every token to the API order of
[31:20] magnitude more than to the consumer.
[31:22] However, I told you we're playing our
[31:24] own game. We have a strategy where we
[31:26] believe there's an AI infrastructure
[31:28] layer, a utility like electricity, and
[31:30] in a future state, you'll want to be
[31:32] able to serve the world at large,
[31:34] consumers, small businesses, large
[31:37] enterprises, governments. That's our
[31:39] strategy.
[31:39] >> Ladies and gentlemen, the CFO of Open
[31:41] AI, Sarah Frier. Well done. Fabulous.
[31:43] [music]
[31:52] [music]
[32:00] >> [music]
