Yahoo Finance

Sarah Friar on Training AI at Altitude

Sarah Friar· CFO at OpenAI
·~11 min·English·Yahoo Finance
AI CompanyAgentsBusiness StrategyInference
TL;DR

OpenAI's CFO reads a $50B-ARR quarter as the payoff of a deliberate sequence: land enterprises on cyber, train new products on their hardest problems first, then ride the frontier down the cost curve toward 1.2 billion consumers.

01The Quarter

A 50/50 Company Now

<strong>Enterprise revenue doubled quarter over quarter, leaving OpenAI split evenly between enterprise and consumer</strong> on a 70% jump in run-rate revenue to roughly $50 billion in ARR.

our enterprise business doubled quarter over quarter, and that leaves us now about 50/50, a very unified platform and very balanced. So, 50% of revenue coming from enterprises.

— Sarah Friar, Yahoo Finance
Key Insight
The 50/50 split is the headline, but the quieter signal is the SMB line (OpenAI's 'business SKU') growing more in one quarter than in the prior six combined. Friar frames the goal as showing up for every segment at once, consumer, developer, small business, enterprise, and government, rather than trading consumer scale for enterprise revenue.
02Land and Expand

Cyber Is the Front Door

<strong>OpenAI lands big enterprises on cyber, then expands as they wire in their own context, memory, and hooks</strong>: its first Visa deal was cyber, and Amgen started on IP security.

what helps them keep coming back for more is you lift out of just the Frontier model layer and into products where they're using their context from their organization, they're using memory, and they're building hooks.

— Sarah Friar, Yahoo Finance
Key Insight
Cyber is a shrewd wedge because it is the one budget line that is simultaneously urgent, board-visible, and recurring. Once the customer's own data and workflows are wired in, switching cost rises even as rival models ship weekly. The lock-in is not the model; it is the integration around it.
03Sequencing

Training at Altitude

<strong>OpenAI ships new products to enterprises first because their problems are harder and more connected</strong>, solving them at altitude before bringing the product down to 1.2 billion consumers.

I usually call this training at altitude because enterprise customers typically have more difficult problems to solve

— Sarah Friar, Yahoo Finance
Key Insight
The altitude metaphor reframes a constraint as a strategy. OpenAI is compute-constrained and cannot yet serve a personal agent to 1.2 billion people cheaply, so 'enterprise first' is partly a cost-rationing decision dressed as a capability-building one. Both can be true at once.
04The Moat

Years of Memory Is the Moat

<strong>OpenAI isn't worried that rivals roll out consumer agents faster, because ChatGPT already holds years of memory and context</strong>; 300 million people ask it health questions every week.

I'm not worried that we're going to get left behind on that front because we've been working with consumers now for multiple years, and there's a lot of memory and context already stored within ChatGPT.

— Sarah Friar, Yahoo Finance
Key Insight
The moat she describes is switching cost built from intimacy: health and finance questions are exactly the data people will not re-enter elsewhere. A competitor can match the model in a week, but it cannot retroactively own three years of a user's private context.
05Myth-Bust

OpenAI Makes the Cheaper Model

<strong>Friar's claim is that the frontier model itself makes the cheap models, so OpenAI's small models undercut Chinese open-source ones</strong>, and she puts Luna at a fifth to a sixth of the cost.

a model like Luna is actually about 1/5 1/6 the cost of say a Chinese open source model. That is one of the myths of 2026 that I think we have to myth bust.

— Sarah Friar, Yahoo Finance
Key Insight
The argument is that cheapness is downstream of the frontier, not a separate race: the same frontier work that pushes capability is what trains the cheaper small models beneath it. If true, it reframes open-weight competition as price leadership following capability leadership rather than undercutting it. The claim is self-serving and unverified here, but internally coherent.
06Economics

Right Token, Right Task, Right Price

<strong>When OpenAI cut Luna's price 80%, usage rose nearly tenfold</strong>, and model families now let the system pick which sub-model answers, so cost follows the task.

we dropped pricing 80% about two weeks after we launched it, and we saw use go up almost 10x.

— Sarah Friar, Yahoo Finance
Key Insight
A 10x jump on an 80% cut points to demand that is highly price-sensitive. Friar gives no baseline or revenue figures, so it is a usage signal, not proof the extra volume paid for the lower price. Model families are the bet that routing cheap tasks to cheap models is what makes that trade work at scale.
07Sovereignty

You Just Count Weight

<strong>On AI sovereignty, Friar argues capability is still concentrated: the US had 52 frontier models in 2025, China 37, Europe 1</strong>, so a country building its own model isn't yet the sensible move.

The US had 52 frontier models in 2025. China had 37 and Europe had one. And so in a world where you need to be ahead of potentially your attackers in a space like cyber, you just count weight.

— Sarah Friar, Yahoo Finance
Key Insight
'You just count weight' is a prizefight metaphor: you back the heavier hitter, not the local one. The uncomfortable implication for sovereign-AI efforts is that Friar, whose company benefits from concentration, is also naming a real asymmetry. In adversarial domains like cyber, second-best capability is not a smaller edge, it is a losing one.
08Go-to-Market

Show Up as a Trusted Partner

<strong>Solving the problem isn't enough; enterprises buy from partners they already trust</strong>, and selling through Amazon's AWS Bedrock opened government and financial-services doors in just the last couple of months.

going to market with Amazon and AWS Bedrock, that's allowed us to get into government and into financial services because that's already a trusted platform that now we are able to go to market with.

— Sarah Friar, Yahoo Finance
Key Insight
It is a telling admission from a frontier lab: in regulated markets the binding constraint is not model quality but procurement trust, and OpenAI is renting Amazon's. The partner that already passed the security review is worth more than a better benchmark; distribution through trust, not capability, is what opened the door.