OpenAI CFO Deep Dive: IPO Stance, Compute Shortage, and Jony Ive's New Hardware Revealed

In a recent episode of the All-In Podcast, OpenAI Chief Financial Officer Sarah Fryer sat down for an in-depth interview, candidly discussing the company's historic fundraising round, IPO stance, competition with rival Anthropic, the compute shortage crisis, and the mysterious new hardware being developed with legendary designer Jony Ive. This conversation not only revealed OpenAI's strategic positioning but also reflected the core challenges and opportunities facing the entire AI industry.

The Logic Behind the Largest Private Fundraise in History

The interview opened with an industry-shaking figure: OpenAI had just completed a fundraising round exceeding $122 billion, which Fryer called "the most successful fundraise in history." For context, the previous largest IPO globally was Saudi Aramco's approximately $30 billion—OpenAI's private round dwarfs it by several multiples.

The structure of this fundraise itself is worth examining. The core difference between a private placement and an IPO lies in disclosure obligations and liquidity: private investors are typically institutional or accredited investors who accept lower liquidity in exchange for earlier access to high-growth companies. OpenAI's valuation in this round was approximately $300 billion, meaning investors are betting that its discounted future cash flows far exceed current revenue—this kind of "expectation-based pricing" is especially common in AI, where markets believe in the nonlinear growth of technology curves.

Notably, the valuation behind this round reflects venture capital's classic bet on the "technology S-curve." In AI, investors employ "options pricing thinking" rather than traditional DCF (Discounted Cash Flow) models—when a technology could restructure entire economic systems, traditional P/E (Price-to-Earnings) or P/S (Price-to-Sales) frameworks break down, replaced by pricing the "window of technological monopoly." Historically, Microsoft in the PC era, Google in the search era, and Amazon in the cloud computing era all experienced similar phases of "forward-looking valuation," ultimately validated by actual market share. Unlike traditional industries, AI companies' moats often don't appear on current income statements but rather manifest in data flywheels, model iteration speed, and ecosystem lock-in effects—which is why investors are willing to price them on forward-looking logic rather than current financial metrics.

Facing speculation about AI companies "racing to IPO," Fryer's stance was notably measured. She repeatedly emphasized one point: "An IPO is a milestone, not a destination. Don't run a company with going public as the goal—it's just another form of fundraising." She cited the classic market metaphor—"the market is ultimately a weighing machine, not a popularity contest"—and noted that "no one remembers whether Google or Yahoo went public first, or whether Lyft or Uber went public first."

During the interview, the hosts also broke "breaking news": competitor Anthropic had confidentially filed an S1. An S1 filing is the registration statement required by the U.S. Securities and Exchange Commission (SEC) before an Initial Public Offering (IPO), containing core information about the company's financial condition, business model, and risk factors. Filing an S1 is merely the starting point of the IPO process; the SEC then conducts multiple rounds of review and issues comment letters that the company must address with revisions, a process that typically takes months or even over a year. Notably, companies can choose "confidential filing," keeping information opaque before the official public disclosure—this has become standard practice among tech unicorns. Fryer responded that filing itself doesn't mean much, since there's still the SEC's lengthy review process ahead.

OpenAI vs Anthropic: Fundamentally Different Strategic Paths

The hosts bluntly pointed out that the industry widely believes Anthropic has "overtaken" OpenAI in terms of developers, enterprise customers, and even revenue. Facing this pointed question, Fryer didn't deflect but responded from the angle of strategic differentiation.

She emphasized that OpenAI's core positioning is building the "AI infrastructure layer"—a single model foundation that reaches the world through multiple interfaces. Specifically:

• ChatGPT: Consumer-facing, with over 900 million weekly active users, it has become synonymous with AI ("the noun and the verb"). The fastest-growing region is Africa, and the fastest-growing languages are Azerbaijani and Kazakh.
• Codex: A programming tool that just surpassed 5 million users over the weekend, starting from nearly zero at the beginning of the year.
• Frontier: Enterprise product line.

Fryer argued that this "single model, multiple interfaces" strategy has compounding advantages: more users bring more data and stronger personalization capabilities, while efficiency improvements from scaling models ultimately drive down per-token costs and improve gross margins. She also addressed external skepticism about OpenAI "spreading itself too thin and neglecting the enterprise market"—the company's revenue structure is now quite balanced, with consumer and enterprise businesses each accounting for roughly half.

1 Gigawatt Equals $10 Billion: The Harsh Reality of Compute Shortage

The most industry-relevant portion of the interview was the discussion of compute economics. The hosts referenced OpenAI's famous conversion formula from roughly 18 months ago: 1 gigawatt (GW) of compute roughly corresponds to $10 billion in annual revenue.

1 gigawatt (GW) = 1 billion watts, a basic unit for measuring large data center power consumption. Construction costs for a 1GW data center typically run in the $10 billion range, with annual electricity costs also reaching billions of dollars. OpenAI's "1GW ≈ $10 billion annual revenue" formula essentially links compute density directly to commercial output as an industry rule of thumb. By comparison, traditional internet data centers typically operate at the tens of megawatts (MW) scale—AI training and inference demands have brought the data center industry into an entirely new dimension. This also explains why energy supply, land approvals, and grid expansion have become core bottlenecks for AI expansion.

The fundamental reason U.S. grid infrastructure takes 5 to 10 years to build lies in the complex approval system spanning federal, state, and local regulatory layers, along with land acquisition challenges for transmission line construction. Much of the existing U.S. grid was built in the mid-20th century, with design capacity that never anticipated the demand explosion of the AI era. Meanwhile, data center siting faces the dual constraints of "power availability" and "cooling water sources"—which is why OpenAI chose Saline, Michigan, an area with both Great Lakes water resources and relatively ample grid capacity, though still far short of keeping pace with AI demand growth.

It's worth adding that AI training clusters (such as NVIDIA H100/H200 racks) can reach power densities exceeding 100kW per rack, while traditional server racks typically run at just 5-10kW—a 10-20x difference. This means that for the same floor space, AI workloads consume several times the electricity of traditional IT loads, placing entirely new demands on cooling systems, power distribution architecture, and grid interconnection points. Liquid cooling technologies (including Direct Liquid Cooling/DLC and immersion cooling) are rapidly replacing traditional air cooling to address the thermal challenges of ultra-high power density.

Fryer acknowledged that compute is currently an extremely scarce resource: "Our business is climbing a vertical wall of demand, and there aren't enough available tokens." She expressed gratitude that the company had aggressively procured compute despite skepticism—"Thank God we did, because even by 2026, our compute still won't be enough."

She further outlined the "chokepoints" in the compute supply chain: energy, land, power, regulatory approvals, racks and chips, memory (currently experiencing price surges), and even talent and trust. When asked about the compute landscape over the next two years, her answer was blunt: "If you want to buy more compute in 2026, good luck; 2027 is also quite limited."

Interestingly, Sam Altman was at that moment in Saline, Michigan, breaking ground on a 1GW data center. Fryer specifically emphasized commitments to the local community: no increase to residential electricity rates, 2,500 union jobs, approximately $1 billion in tax payments, and $45 million invested in educational Codex credits. She views "trust" as part of the supply chain, candidly stating that "you can't tell a community top-down what they need."

Capital Allocation and Cost Curves: The CFO's Core Thinking

As CFO, Fryer offered a detailed breakdown of OpenAI's capital allocation model. She noted that the truly enduring, high-value companies of this era "won't be magic—they'll be like great companies of past eras"—starting from customer value and then achieving excellent gross margins.

On the cost side, she revealed a stunning compute deflation curve: from ChatGPT-4 to version 5.4, service costs dropped approximately 97%, all within roughly two years. Even with the latest 5.5 model, though externally priced higher, the dramatic improvement in per-token efficiency means actual customer costs still decreased by approximately 20% to 30%.

Regarding compute forecasting, Fryer explained OpenAI's dual assumptions: on one hand, per-gigawatt compute costs are rising due to electricity and memory price increases; on the other hand, intelligence improvements from chip depreciation more than compensate, so end-customer unit costs continue falling. She admitted that the further out the forecast, the more it resembles "reverse engineering"—projecting potential revenue backward from already-purchased compute.

She also shared a compelling anecdote: a year ago, when she was modeling agentic revenue for investors, she predicted developers might be willing to pay $2,000 per month—no one believed it at the time; just as when ChatGPT Pro was priced at $200, people said "no one will pay." Agentic AI refers to AI systems capable of autonomous planning, tool invocation, and multi-step task execution, distinct from traditional single-turn Q&A interactions with large language models. The core technical breakthrough of agentic AI systems lies in the maturation of "Function Calling/Tool Use" capabilities, transforming LLMs from passive respondents into active executors. Typical agentic frameworks like LangChain, AutoGen, and CrewAI decompose complex tasks into sub-task chains and grant models the ability to call external APIs, databases, and code execution environments, achieving the qualitative shift from "answering questions" to "completing work." The commercial significance of this transformation is that users are willing to pay far more for "completed results" than for "receiving information," directly supporting higher ARPU (Average Revenue Per User). In agentic scenarios, completing a single task may consume thousands or even tens of thousands of tokens (the basic unit of text processed by models), far exceeding ordinary conversation. This directly drives explosive growth in inference compute demand and explains why Fryer predicts "compute still won't be enough even by 2026." The decline in token costs (97% over two years for OpenAI) is a key driver of AI adoption—the lower the cost, the more commercially viable agentic applications become, and the higher the monthly fee ceiling developers are willing to pay. This positive flywheel is the core economic logic behind OpenAI's bet on the "AI infrastructure layer."

From a cost structure perspective, the typical architecture of agentic AI systems includes a Planner, Tool Use layer, Memory module, and Executor. A complete agentic task may involve dozens of LLM calls, each consuming hundreds to thousands of tokens, potentially totaling tens of thousands of tokens. At GPT-4-level historical pricing (approximately $30 per million output tokens), a single complex agentic task could cost several dollars—it's precisely the deflation curve of token costs dropping over 95% from 2022 to 2024 that truly opened the commercial space for agentic applications, making it feasible for developers to pay higher monthly fees.

Diversified Compute Strategy: From Single Cloud to the "Rubik's Cube" Approach

Regarding how long $122 billion can last and how many gigawatts it can build, Fryer used the "Rubik's Cube" as a metaphor to describe the multi-dimensional evolution of OpenAI's compute strategy. Just two years ago, the company relied on a single cloud provider (Microsoft Azure), a single chip (NVIDIA), a single product (ChatGPT), and a single price point ($20/month).

Today, OpenAI has built a highly diversified compute foundation:

• Multiple Cloud Service Providers (CSPs): Oracle, CoreWeave, Microsoft, GCP, AWS, and a batch of smaller neo-scalers. The key role of CSPs is converting capital expenditure (CapEx) into operating expenditure (OpEx). CapEx (Capital Expenditure) refers to one-time large investments in long-term assets (like servers and data centers) that are depreciated over time on the balance sheet; OpEx (Operating Expenditure) refers to pay-as-you-go daily operational costs that directly hit the current period's income statement. For OpenAI, which is not yet profitable and faces cash flow pressure, converting compute procurement from CapEx to OpEx through CSPs means "pay for what you use," greatly reducing financial risk and capital commitment.
• Multi-chip strategy: NVIDIA remains the primary partner (large-scale training this fall will run on Vera Rubin, with Feynman series already in planning); simultaneously deploying AMD, Cerebras (already online, offering low latency ideal for real-time programming), and custom chips developed in partnership with Broadcom. NVIDIA's chip naming convention uses scientists' names—after Hopper (H100) comes Blackwell (B200), followed by Vera Rubin (expected mass production 2025-2026). The Vera Rubin architecture will be the first to introduce HBM4 memory, with theoretical bandwidth roughly 50% higher than Blackwell, while supporting FP4 precision training to further compress per-token compute costs; the Feynman series represents a longer-term roadmap milestone. OpenAI's early lock on Vera Rubin capacity means it has first-mover advantage in next-generation training infrastructure—this is the deeper context behind "compute still won't be enough in 2026": even when new chips reach mass production, capacity will have been pre-booked by top customers. Cerebras employs a unique Wafer-Scale Engine architecture, fabricating an entire silicon wafer into a single chip, processing massively parallel computation at extremely low latency, particularly suited for multimodal real-time inference scenarios.
• Build-to-suit: Such as the data center being built in Texas in partnership with SoftBank, marking a shift from pure CSP models toward greater CapEx investment.

Fryer's core philosophy is "maximizing optionality," especially critical during a phase when OpenAI has not yet achieved investment-grade status and cannot access low-cost debt financing—making partnerships with collaborators essential.

Regarding the industry trend of "everyone doing everything" (NVIDIA building models, Google making chips), she believes everyone is competing for "the layer closest to the customer"—because that layer typically captures the largest share of ecosystem profits. This is precisely why OpenAI insists on being the "AI intelligence layer." She specifically noted that a year ago the industry worried LLMs would be commoditized, but with the development of the agentic layer and the "harness" (the framework carrying context and memory), the opposite has occurred.

Jony Ive Teams Up with OpenAI: Mysterious New Hardware First Disclosed

The most anticipated topic in the interview was OpenAI's new hardware being developed with legendary designer Jony Ive. Jony Ive is the former Chief Design Officer of Apple, who led the design of the iMac G3, iPod, iPhone, iPad, and Apple Watch—products that defined their eras. He is regarded as one of the most influential figures in industrial design history. He left Apple in 2019 to found the independent design firm LoveFrom. The core of his design philosophy is "human-centered technological embodiment"—presenting complex technology in the most natural, emotionally resonant form possible. Given iPhone's history of redefining human-computer interaction, if OpenAI and Ive's collaboration can package AI capabilities into a "seamless" physical medium, its industry impact could rival that of smartphones disrupting the PC era.

To understand the significance of this collaboration, one must recognize the core dilemma of current AI hardware: existing AI interactions are primarily parasitic on smartphones and PCs—hardware forms designed for the pre-AI era. From an industrial design history perspective, every shift in computing paradigm has been accompanied by a revolution in interaction medium: from command line to graphical interface (mouse + display), from desktop to mobile (touchscreen + sensors)—each transition birthed new hardware forms and interaction languages. The core design challenge for AI-native devices is the "boundary of ambient awareness"—the device needs to be "smart" enough to understand context while being "restrained" enough not to violate privacy. Jony Ive's design principles established at Apple—"good design makes technology disappear"—are precisely the key to resolving this contradiction. The smartphone interaction paradigm (touchscreen + app icons) was born in 2007, fundamentally an interface optimized for human fingers and visual attention. The core assumption of AI-native devices is: when AI can continuously perceive the environment, understand context, and proactively respond to needs, the interaction flow of "open app → input command → wait for result" itself becomes the bottleneck. The failures of early attempts like Humane AI Pin and Rabbit R1 revealed the core challenge of AI hardware: it's not that technology isn't good enough, but how to achieve "ambient awareness + instant response" seamlessly without disturbing the user. Jony Ive's design philosophy—letting technology recede so humanity emerges—directly addresses this pain point.

Fryer didn't reveal the specific form factor but confirmed "it will be unveiled by the end of this year," calling it a kind of "consumer substrate." She has personally experienced the prototype. When asked whether it's like "the paradigm shift of using the iPhone for the first time," she described: "What Johnny and the team are truly great at is infusing humanity into devices... when you see it, when you use it, you can feel it." She used words like "natural," "lovable," "intimate," and "seamless" to describe the product.

This hardware direction is consistent with Fryer's repeated emphasis on the "multimodal" trend. Multimodal AI refers to models capable of simultaneously processing text, voice, images, video, and other forms of information. Compared to pure text interaction, multimodal (especially real-time voice and vision) places extremely demanding requirements on inference latency—users expect millisecond-level responses, not waiting several seconds. Real-time voice and visual interaction typically requires inference latency under 200 milliseconds (the psychological threshold for humans perceiving "instant response"). This constraint challenges traditional "large model centralized inference" architectures, driving rapid development of Edge Inference and Model Distillation technologies. NVIDIA's Jetson series, Qualcomm's AI chips, and Apple's Neural Engine are all hardware solutions optimized for edge inference scenarios. This places fundamentally different demands on compute architecture: the training side pursues throughput, while the inference side pursues low latency. As multimodal interaction becomes widespread, real-time inference compute will become the next scarce resource—this is one of the deeper reasons OpenAI is proactively deploying a multi-chip strategy. OpenAI's deployment of Cerebras's low-latency architecture is precisely a strategic move to pre-position for multimodal real-time scenarios. She criticized the contemporary habit of "speaking with thumbs"—"it's a disease, people walk around with their heads down"—and argued that the multimodal era has arrived, people will speak directly to their tools, and this will require even more real-time inference compute.

Advertising Monetization: Enormous Potential Within Principled Boundaries

Near the end of the interview, the hosts posed a key monetization question: of the three greatest consumer businesses in history (iPhone, Meta ads, Google ads), two are advertising-based—why does OpenAI rarely discuss advertising?

Fryer's response reflected clear principled boundaries: always deliver the best results based on the model, not sponsored content; and always preserve an ad-free paid tier for users who don't want to see ads. But she simultaneously painted a highly compelling advertising vision, quoting colleague Fiji: "If Google and Meta had a baby, it would be ChatGPT."

The logic is this: Google Search has high intent, and ChatGPT goes even further—a single conversation containing 50 questions counts as just one query, meaning actual intent density far exceeds surface-level metrics; Meta's advantage lies in precise audience profiling. ChatGPT possesses both intent and memory—"it knows who I am."

From the fundamental logic of advertising technology, Google Search ads' core asset is "keyword intent signals"—when a user types "buy running shoes," advertisers are willing to pay premium prices for that clear purchase intent (some keyword CPCs exceed $50). But search queries are typically fragmented 3-5 word phrases lacking context. ChatGPT's conversational data contains complete decision chains: a user might progress within the same session from "learning about running" to "comparing brands" to "asking about purchase channels"—this intent evolution trajectory is far more valuable to advertisers than a single keyword. More critically, if ChatGPT accumulates users' long-term memory (preferences, historical decisions, life contexts), its user profiling precision will surpass Meta's behavioral data—the latter relies on inferring from users' historical behavior, while the former comes directly from users' active expression. Combining memory, context, and intent could theoretically build an extraordinarily powerful advertising platform—this is the commercial meaning behind "it knows who I am."

She candidly stated that if she were optimizing only for current revenue, she would allocate every token to the API (revenue an order of magnitude higher). But OpenAI chooses to "play its own game"—believing that the AI infrastructure layer will become a public utility akin to electricity, serving consumers, small businesses, large enterprises, and governments worldwide.

Conclusion

This interview comprehensively presented OpenAI's deep logic across capital, compute, product, and strategic dimensions. From the largest fundraise in history to candidly acknowledging compute shortages, from strategic divergence with Anthropic to the mysterious hardware collaboration with Jony Ive, the core message Fryer conveyed was clear and consistent: OpenAI is betting on the long-term future of "AI as public infrastructure," using "maximizing optionality" as its financial principle, and proactively positioning ahead of demand under extreme compute scarcity constraints. For practitioners and investors following the evolution of the AI industry landscape, this first-hand information is invaluable.