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

Why does it matter? Because the most expensive deliberate under-monetization in tech history might also be the most rational.

OpenAI's CFO spent an hour on stage handing out real numbers — and several of them fundamentally reframe the AI investment thesis. Token costs collapsed 97% in two years. The company is already anxious about compute for 2030-2032. And every day, OpenAI walks away from an order-of-magnitude revenue premium by serving consumers instead of routing all tokens to the API.

• Token costs fell 97% from GPT-4 to GPT-4.5 in two years — OpenAI raised prices on its newest model anyway, capturing the spread rather than passing it through • The Michigan gigawatt data center breaking ground today won't produce tokens until late 2027 or early 2028; the real compute anxiety is 2030-2032 • API earns an order of magnitude more per token than consumer — OpenAI ignores this gap deliberately • ChatGPT holds at least 11% of the search market with persistent memory layered on top, positioning it as the highest-signal ad platform ever assembled

Token costs fell 97% in two years — and OpenAI raised prices anyway

When OpenAI launched its newest model, it raised prices 2x — even as token costs had already fallen 97% from GPT-4 to GPT-4.5 over just two years. Customers still got a break: roughly 20-30% cheaper per token due to efficiency gains. But OpenAI captured the majority of that deflationary spread rather than competing it away.

Friar's framing of the capital allocation logic is precise: "Part of making a capital allocation decision is having to — if you make it on today's cost profile, you actually might misprice the outcome." The curve moves fast enough that anchoring to today's economics systematically undervalues the future business.

The structural consequence: any company controlling the customer-facing layer can expand gross margins even as input costs collapse, because willingness-to-pay doesn't deflate at the same velocity as compute. The entire debate over whether hyperscale AI infrastructure spending will earn returns largely misses this dynamic. Whoever owns the customer relationship captures the spread between a rapidly falling cost denominator and sticky pricing power — and that gap is compounding in incumbents' favor.

OpenAI is already worried about 2030 compute — not 2026

OpenAI is already anxious about compute availability for 2030-2032. Not 2026. "In 26 we still won't have enough compute." The follow-up is what recalibrates timelines: "Where I feel most short of compute right now is starting to look at 30, 31, 32."

The Michigan gigawatt data center — Sam Altman was cutting the ribbon on it as Friar spoke — won't produce tokens until late 2027 or early 2028. Shovels in the ground today; revenue in three years. "If you want to buy more compute, good luck to you — tell me, 'cause I don't know where else to find it."

The choke points shift continuously rather than forming a single queue: energy and land, permitting speed, rack and chip supply chains, memory (currently spiking), talent pipelines, and community trust. The Michigan pitch required $1 billion in state taxes, 2,500 union jobs, and $45 million in Codex education credits just to get the community on board. Each layer of the supply stack unclogs on its own timeline, and the full chain only aligns when all of them clear together.

Any model of AI supply that assumes relief in two to three years is working off the wrong time axis.

API tokens earn 10x more than consumer — and OpenAI routes compute to consumers anyway

API tokens earn an order of magnitude more than consumer tokens — and OpenAI routes compute to consumers anyway. Friar is direct: "If I was optimizing only for today, I would give every token to the API. Every token to the API order of magnitude more than to the consumer."

She doesn't do that. The free consumer tier burns compute at lower revenue per token by design. The strategic logic: "We have a strategy where we believe there's an AI infrastructure layer, a utility like electricity, and in a future state, you'll want to be able to serve the world at large."

Benchmarking OpenAI's financial health against pure API monetization misses the entire bet. The consumer base isn't a revenue drag — it's the hedge against model-layer disintermediation. If foundation models commoditize, whoever has the deepest context and memory layer wins the transition. The "inefficient" consumer allocation today is insurance against losing the game entirely at the layer that eventually matters most.

ChatGPT is what you'd get if Google and Meta had a baby — and memory makes it something neither can replicate

"If you know Google and Meta had a baby, it would be ChatGPT." The framing earns its precision. Google has high purchase intent — search reveals what you're about to buy. Meta has demographic targeting — people like you, psychographic, social graph. ChatGPT has both, plus something neither company can match: persistent memory. It knows who you are across every conversation, not just what you typed five minutes ago.

OpenAI already holds at least 11% of the search market — and that's almost certainly an undercount. A full Google session with multiple page refreshes counts as one query; a 50-turn ChatGPT conversation also counts as one. The engagement depth is structurally incomparable.

"Imagine putting memory and context next to intent. You should have a very potent ad platform." This isn't a defensive ad play — it's an attempt to build the highest-signal advertising surface ever assembled and use that revenue stream to fund genuinely universal access. Results stay model-driven, never sponsored. A paid ad-free tier always exists. But beneath that, a combination of purchase intent, demographic context, and persistent personal memory that has never existed simultaneously in one product is being assembled.

The LLM commoditization thesis didn't just slow down — it inverted

A year ago, conventional wisdom had foundation models commoditizing. Margins would compress, no one would have durable advantage at the base layer, and the real money was in applications built on top.

"Frankly, it's gone the opposite."

The inversion came from the agentic context layer. Friar's Codex instance carries a memory file that knows she's OpenAI's CFO, knows her writing style, knows she has teenagers. That specificity makes the model more powerful for her than for any other user — and switching to a competitor means losing all of it.

At enterprise scale, what accumulates is harder to replace than data. "It's not just even about the data that resides there, but the intuition of an enterprise" — the institutional knowledge that explains why a stock won't move even when the earnings numbers are right, the kind of judgment a trader carries that no database captures. That intuition gets encoded into the context layer over time.

Switching costs are compounding in real time. Model benchmarks are increasingly the wrong unit of competition. The real contest is over who accumulates the richest context layer first — and that asset doesn't transfer between providers.

The $122B raise is a credit arbitrage play — and it explains everything about the multi-CSP strategy

OpenAI's multi-CSP strategy is primarily a balance-sheet maneuver, not a technical one. "What CSPs do for us in effect is they shift capex into opex — so you pay as you get the revenue." OpenAI rides Oracle, CoreWeave, Microsoft, GCP, and AWS's balance sheets to build compute scale it can't yet finance cheaply on its own.

Two years ago: one CSP, one chip (Nvidia), one product (ChatGPT), one price point ($20/month). Today the chip pipeline includes AMD, Cerebras, and OpenAI's own Broadcom silicon in development — with Nvidia remaining the priority partner for the next major training run on Vera Rubin.

"In a moment where I'm not yet an investment grade type of entity where I can go get lower-cost debt financing, being able to work with partners to do that is really important."

The $122B raise is designed to preserve optionality while OpenAI's credit profile catches up to its ambitions. The endgame is reaching investment-grade status, accessing cheap debt directly, and reducing dependency on partner balance sheets. The frontier AI capital race isn't just about who raises the most — it's about who structures the balance sheet to deploy it most efficiently through leverage while they wait.

The gap between a free user and a Pro user is 11x daily engagement — not just a pricing tier

A free ChatGPT user asks about 7 questions a day. First paid tier doubles it to about 15. ChatGPT Plus at $20 a month: 3x free. Pro at $200 a month: 11x free — roughly 77 turns a day.

"Once they get a taste of intelligence, the ability to come up a commitment curve is incredible."

Standard subscription metrics miss what's actually happening. CAC and monthly churn don't capture what it means when usage grows 11x across pricing tiers. Each step doesn't just unlock features — it produces a structurally heavier user who has reorganized meaningful parts of their daily thinking around the tool.

"Remember when people were losing their minds over ChatGPT Pro being at $200? Oh my god, no one will ever pay for that." They paid. And agentic subscriptions at $2,000 a month — which Friar pitched to investors a year ago to visible disbelief — are landing now. The ceiling on the commitment curve keeps moving up. The right question for valuing this consumer base isn't conversion rate. It's where on that curve users eventually stabilize.

The race isn't over models — it's over who gets to be electricity

Friar's utility analogy is the tell. Utilities don't win through superiority; they win through entrenchment. Once the infrastructure layer is established, everything else builds on top of it on the utility's terms. OpenAI's consumer subsidies, its memory accumulation strategy, its advertising ambitions — all of them point toward the same destination: becoming the entity that sets those terms before anyone else can.

The goal isn't to win the AI race. It's to become its infrastructure.

---

Topics: OpenAI, AI infrastructure, compute scarcity, LLM economics, advertising, capital allocation, agentic AI, IPO strategy, CFO, Sarah Friar, data centers, enterprise AI, token pricing