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

2026-06-02 - 32 min - source - Read full transcript

Sarah FriarChamath Palihapitiya (host)Jason Calacanis (host)David Sacks (host)David Friedberg (host)

Key insights

OpenAI raised $122 billion in March 2026, the largest private fundraising round in history and more than four times Saudi Aramco's ~$30 billion IPO record.

Friar frames the raise as being about maximizing 'optionality' for OpenAI and for the broader AI era, not preparation for an imminent IPO. She contrasts it directly against the prior record holder, Saudi Aramco, to underline the scale shift the market has undergone.

ipo-timing

Friar insists an IPO is 'a milestone, not a destination' and warns against running the company as if going public is the goal.

She tells her team this explicitly, arguing that treating an IPO as a finish line distorts decision-making; instead it's just one more fundraising tool. She adds that markets are 'weighing machines, not popularity machines' and nobody remembers who IPO'd first between Google/Yahoo or Lyft/Uber, only who built the more durable company.

ipo-timing

OpenAI's core strategy is one foundation model exposed through many interfaces, not separate product bets, which Friar argues compounds into lower per-token costs and higher margins as it scales.

She names ChatGPT (900M+ weekly users), Codex (5 million users, up from near zero in January), and Frontier (the enterprise offering) as different doors into the same underlying model. More users and more data feed a single system, which she says should lower the cost per token and raise gross margins over time.

openai-strategy

OpenAI's revenue is now roughly 50/50 between consumer and enterprise, pushing back on the narrative that it neglected enterprise while chasing consumer projects like Sora.

Friar rejects the binary framing of 'consumer company vs. enterprise company,' pointing to a packed schedule of enterprise visits (Thermo Fisher, banks, Travelers) and crediting new head of revenue Denise Dresser, in the role since December, with driving enterprise momentum.

openai-strategy

Compute remains OpenAI's binding constraint: Friar says the company is 'going up a vertical wall of demand' and still won't have enough compute through 2026.

She lists choke points spanning energy, land, power, regulatory approval speed, chip/rack supply, a current memory spike, talent (she flags concern as a Stanford trustee about whether the education system produces enough people), and public trust in communities hosting data centers.

compute-economics

Friar's 'gigawatts to cash' framework holds that 1 gigawatt of compute converts to roughly $10 billion a year of OpenAI revenue, while standing up that gigawatt costs about $50 billion all-in.

She first stated the framework publicly about 18 months before this conversation and says it still roughly holds, though it's evolving as she now models compute purchases years ahead - she describes being most compute-short for 2030-2032, not just the near term.

compute-economics

OpenAI has diversified from a single-CSP, single-chip, single-product company two years ago into a multi-CSP, multi-chip portfolio specifically to convert CapEx into OpEx before it reaches investment-grade credit status.

Two years ago OpenAI ran on one cloud (Azure), one chip (Nvidia), one product (ChatGPT), one price ($20/month). Today it sits on top of Oracle, CoreWeave, Microsoft, GCP, AWS and neoscalers, and is adding AMD and Cerebras chips plus a custom Broadcom chip alongside Nvidia (its priority partner, whose Vera Rubin chips power the next major training run). It's now also moving into build-to-suit deals like a SoftBank Energy data center in Texas, which require more direct CapEx.

compute-economics

Model cost-per-token has fallen roughly 97% across recent GPT generations, and even after a 2x price increase on the newest model, customers still net a 20-30% per-token cost reduction from efficiency gains.

Friar cites this deflation curve as the mechanism that lets OpenAI raise headline prices while still passing savings to customers, and as a key input to how the company models future gross margin even as raw compute (power, memory) gets more expensive on a per-gigawatt basis.

compute-economics

OpenAI and Jony Ive's team will unveil a new consumer hardware device by the end of 2026, with purchases opening in early 2027.

Friar, who has used a prototype, describes its differentiator as emotional and design-driven rather than a spec upgrade - it 'feels very natural' and 'very lovable,' bringing what she calls humanity to devices, and reportedly removes the need to take a phone out at all.

consumer-ai-devices

Friar argues durable AI value is shifting from the model itself, which is trending toward commodity, to the 'harness' - the memory, context, and agentic scaffolding wrapped around it.

She uses her own experience with Codex's memory of her role and preferences as an example, then extends the idea to enterprises: once a model is fused with a company's institutional context and intuition (she compares it to a trader who knows a fund is under pressure beyond what the raw data shows), it becomes far stickier and more valuable than the underlying LLM alone.

openai-strategy

OpenAI will always keep an ad-free tier for paying users but is actively testing ads in its free tier, betting ChatGPT can out-monetize Google and Meta by combining their two ad models.

Friar says the company undercounts its own search-competitive footprint (a 50-question ChatGPT conversation counts as a single interaction the way a Google search does), giving it high purchase intent similar to Google, while its persistent memory gives it personal context similar to Meta - a combination she argues makes for 'a very potent ad platform.'

ai-monetization

If OpenAI optimized purely for near-term revenue per token, it would route all compute to enterprise API customers over consumer ChatGPT users - but it deliberately doesn't.

Friar says API revenue per token is an order of magnitude higher than consumer revenue per token today, yet OpenAI's strategy treats AI as a broad utility layer, akin to electricity, meant to serve consumers, small businesses, large enterprises, and governments together rather than chasing the highest-margin segment alone.

ai-monetization

Media referenced

Wall Street Journal article on AI IPO timing - article - Referenced by a host as the source of the idea that there's an advantage to IPOing earlier if you're an AI company, kicking off the SpaceX/OpenAI/Anthropic IPO-race discussion.

Companies

OpenAI - Friar is OpenAI's CFO; the interview covers its fundraising, compute strategy, and monetization.
Anthropic - OpenAI's chief rival, discussed as having confidentially filed its S-1 and having overtaken OpenAI on some developer/enterprise metrics.
SpaceX - Cited as the company whose IPO announcement reignited the question of when OpenAI and Anthropic will go public.
Microsoft - OpenAI's original and still-primary cloud partner (Azure); one of several CSPs in OpenAI's current multi-cloud compute strategy.
Nvidia - OpenAI's priority chip partner; next major training run will use Nvidia's Vera Rubin chips.
Oracle - One of OpenAI's multiple cloud service providers; the Saline, Michigan 1-gigawatt data center is part of OpenAI's Oracle complex.
CoreWeave - Named as one of OpenAI's multiple cloud service providers.
Google - Named among OpenAI's cloud providers (GCP) and discussed as a search/ads comparison point for ChatGPT.
AWS - Named as one of OpenAI's cloud service providers.
AMD - One of the chip vendors OpenAI has added to diversify beyond Nvidia.
Cerebras - Already online as an OpenAI chip partner; described as a low-latency chip good for real-time coding use cases.
Broadcom - Partner on OpenAI's custom in-house chip.
SoftBank - Partner (SoftBank Energy) on a built-to-suit data center OpenAI is building in Texas, beyond its CSP relationships.
Thermo Fisher - Cited as an enterprise customer example: using OpenAI tools to speed up patient screening and FDA approval timelines.
Nextdoor - Friar's prior company (she spent 7 years there); referenced as the source of her community-engagement approach to data center rollout.
Stanford - Friar is a trustee; cited in her comments about worrying whether the education system produces enough AI talent.
Meta - Compared to Google as one of the two great ad-based consumer businesses; referenced in the 'if Google and Meta had a baby' ChatGPT-ads analogy.
Saudi Aramco - Its ~$30 billion IPO cited as the prior record for largest IPO, which OpenAI, Anthropic, and SpaceX are all expected to exceed.

Techniques and frameworks

Gigawatts to cash - Friar's framework, first articulated about 18 months earlier, that 1 gigawatt of AI compute equates to roughly $10 billion a year of OpenAI revenue.
Rubik's Cube compute-diversification metaphor - Friar's mental model for OpenAI's shift from one CSP/one chip/one product/one price two years ago to a multi-CSP, multi-chip, multi-product, multi-price portfolio today.
The harness - Term Friar uses for the memory, context, and agentic scaffolding wrapped around a model, which she argues is becoming the real differentiator as raw LLMs commoditize.

Summary

Sarah Friar, OpenAI's CFO, joins the full All-In quartet for a wide-ranging interview taped the same day OpenAI's Sam Altman was cutting the ribbon on a new 1-gigawatt data center in Saline, Michigan. The conversation opens on the IPO question raised by SpaceX's move toward going public: is there a race between OpenAI and Anthropic (which had just confidentially filed its S-1) to IPO first? Friar reframes the premise entirely, insisting an IPO is "a milestone, not a destination" and that OpenAI's record-setting $122 billion raise in March 2026 (more than four times the prior record, Saudi Aramco's ~$30 billion IPO) was about maximizing optionality, not signaling imminent public-market timing.

Pressed on whether Anthropic has overtaken OpenAI in developer and enterprise traction, Friar lays out OpenAI's strategy: one foundation model exposed through many interfaces - ChatGPT (900M+ weekly users), Codex (5 million users, up from near zero in January), and the enterprise-focused Frontier offering - rather than separate, siloed products. She pushes back on the idea that OpenAI over-indexed on consumer projects like Sora at enterprise's expense, noting revenue is now roughly 50/50 between the two, with new head of revenue Denise Dresser driving enterprise momentum since December.

A large portion of the interview digs into compute economics, Friar's specialty. She reiterates her "gigawatts to cash" framework (1 gigawatt ≈ $10 billion/year of revenue, ~$50 billion all-in to build) and describes OpenAI's transformation from a single-cloud, single-chip company two years ago into a diversified, multi-CSP (Oracle, CoreWeave, Azure, GCP, AWS) and multi-chip (Nvidia, AMD, Cerebras, a custom Broadcom chip) operation - a shift she frames as converting CapEx into OpEx while OpenAI isn't yet investment-grade enough to finance builds cheaply on its own. She also walks through the Michigan data center's local commitments (2,500 union jobs, $1 billion in local taxes, no rate hikes for residents, $45 million into Codex education credits), drawing on her prior experience at Nextdoor doing community-level trust-building.

The interview surfaces two notable new-product signals: a Jony Ive-designed consumer hardware device, to be unveiled by the end of 2026 and purchasable in early 2027, which Friar - having used it - describes as emotionally distinctive rather than spec-driven ("very natural," "very lovable"); and an emerging ads strategy for ChatGPT's free tier, paired with a permanent commitment to an ad-free paid tier. Friar argues ChatGPT could out-monetize both Google (high purchase intent) and Meta (personal context via memory) combined, while stressing that OpenAI deliberately doesn't optimize purely for revenue-per-token (which would favor enterprise API traffic) because its strategy treats AI as a broad utility layer meant to serve consumers, businesses, and governments alike.

Underlying the whole conversation is Friar's argument that the AI stack is converging - chipmakers building models, clouds building chips, model companies building hardware - and that the real prize is owning the "harness": the memory, context, and agentic scaffolding around a model, which she sees replacing the model itself as the primary source of differentiation and lock-in.

Notable Quotes

"An IPO, I say this to the team all the time, it's a milestone. It is not a destination. Do not run your company as if that's some sort of destination." - Sarah Friar

"The market is a weighing machine, not a popularity machine. No one remembers who won first, Google or Yahoo, Lyft or Uber." - Sarah Friar

"We're going up that kind of vertical wall of demand right now, and there's just not enough tokens available." - Sarah Friar

"If you took what I - Fiji says this really well - if, you know, Google and Meta had a baby, it would be ChatGPT." - Sarah Friar

"A year ago people talked about the commoditization of the LLMs. And frankly, it's gone the opposite because as you start building an agentic layer... the harness is what brings the context, the memory." - Sarah Friar