OpenAI's business model is a negative snowball: the more it burns, the more it loses
A model company has to spend ten times last year's training cost on the next generation of models, while revenue only comes back at twice last year's training cost. Add 2, subtract 10, and cash flow goes more negative every year. Only on the day it stops burning does the income statement suddenly look good.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Model companies are a negative snowball
Freda reduces OpenAI's business model to a minimal equation: assume last year's training cost is 1, next year's revenue brings back 2, but the next generation model's training cost is 10 — so plus 2 minus 10 equals -8. That means large models are essentially a negative snowball: every year burns more cash than the year before. There are only two ways to turn it positive: either next year's revenue is more than twice last year's training cost, or you stop burning on a model 10 times bigger. Anthropic CEO Dario subdivides the latter further: either you hit a physical limit and can't train further, or the Scaling Law slows down and it isn't worth burning another 10x.
— FredaThe income statement looks best the day you stop burning
Freda uses Netflix as an analogy: Netflix's cash flow was negative for years, also a negative snowball, worst in 2019 at negative 3 billion; in 2020 the pandemic halted content production, content costs disappeared, and cash flow suddenly flipped positive to 2 billion. Model companies are the same — the thing you burn money on only truly generates cash flow on the day you stop burning it, and at that moment the income statement looks very good, which Freda calls "very much violence aesthetics." But he also concedes Netflix will never stop investing in content entirely, only the growth rate comes down; model companies are the same — training costs no longer multiplying several times a year is enough for margins.
— FredaThe market underestimates OpenAI's enterprise side
OpenAI has four revenue lines: ChatGPT at over 70% (including enterprise GPT), API, Agent (including the SoftBank partnership), and new products. The company itself forecasts the latter two lines growing faster and taking a larger share in future. Freda points to a non-consensus view: everyone treats OpenAI as a pure 2C company, but today its consumer and enterprise shares are roughly equal, and enterprise users already number in the millions. Enterprise as a super entry point will develop very fast, because Google Workspace actually makes it easy for OpenAI to plug in directly — the software charging model is simple, and as long as users keep paying, Google doesn't care whether it sits at the top of the acquisition funnel.
— FredaOpenAI's real rival is Google
OpenAI internally believes its biggest competitor was never Anthropic or xAI, but Google. Google's position is too good: horizontally it has Search, YouTube and enterprise-facing Workspace; vertically it extends from cloud to in-house chips. Gemini has 650 million MAU, GPT over 1 billion. Google can fight a price war from two angles: absolute price, where TPU vertical integration lets the same model produce higher margins; and bundling, where a 20-dollar YouTube user adds Gemini for 10 more. Freda thinks the reasonable medium-to-long-term landscape is Google taking mid-to-low-end subscriptions while GPT keeps the high-end market at several hundred or even several thousand dollars a month.
— FredaThe Google threat is equally real
Freda thinks the market's negative sentiment on OpenAI and positive sentiment on Google are a bit too one-sided. Google's search ads haven't fallen today because everyone still uses GPT as a big Wikipedia, asking why the sky is blue — a kind of Q&A Google never monetised anyway. But insurance and law firms are Google's biggest advertisers — buying car insurance today might mean googling and clicking the top ad slot, and that question is far too suited to asking GPT, which may already know the user drives badly and can tailor the best one. Once GPT starts doing ads and agentic commerce, the landscape will change. Total US ad revenue is only so big; if OpenAI wants to sell ads, it has to take them from Google and Meta.
— FredaThe two model companies assume opposite directions
Over 80% of Anthropic's revenue is 2B, and within a year it went from losing 2 for every 1 earned to gross margins roughly comparable to OpenAI's. From 100 million to 1 billion in revenue, OpenAI took less than two quarters, Anthropic four or five; but from 1 billion upward it reverses — Anthropic's annualised revenue went from 1 billion to 7 billion in under ten months. Freda points to an interesting divergence in the two companies' financial forecasts: Anthropic assumes total compute stops growing after 2028, lifting margins by training costs no longer growing rapidly; OpenAI assumes model ROI rises year by year, so one dollar returns five the next year. So it isn't that Anthropic spends less — it assumes a higher rate of return.
— FredaRobinhood is a case of playing a bad hand well
Robinhood is the best-performing stock in the S&P 500 this year, but Freda says its business model is fundamentally poor — securities trading is strongly cyclical, heavy trading in bull markets and dead in bear markets. It can do four things: diversify (already over 11 business lines and over 100 million in revenue), grab market share to smooth the cycle, pricing power (crypto commissions went from 10 points to 60 points in three years), and cost-side discipline (operating costs flat since 2022, a horizontal line). Freda stresses being honest with yourself about whether you're earning beta or alpha: Coinbase has had no alpha relative to Bitcoin since 2022, while Robinhood has strong alpha relative to both Bitcoin and the Nasdaq.
— FredaAI revenue has to tear a hole in the labour market
Freda proposes the concept of "electronic revenue": US online advertising is about 260 billion, e-commerce commissions about 100 billion, subscriptions about 50 billion, totalling about 400 billion. If OpenAI needs 200 billion in revenue entirely by taking existing share, then Google, Meta and Amazon all have to give up share, and burning that much money to carve out only a small pie isn't very meaningful. He argues total advertising has no reason to increase because of AI — US companies spend an average of 3% of revenue on advertising, and there's no reason for it to exceed 3% with AI. The real space is in America's 30 trillion GDP and 15 trillion in labour costs; customer service alone is a 300 billion market, and tearing one hole there could yield hundreds of billions in revenue.
— FredaIn their own words · checked verbatim
So it's a very simple plus 2 minus 10 and you become a -8. The large-model business model is essentially a negative snowball — every year it burns more cash than the year before.
所以就是一个很简单的你加2减10 你就变成了一个-8 大模型这个商业模式本质是一个负向滚雪球 每一年都会比上一年的现金流烧得更多
Freda13:07
The thing you burn money on only truly generates cash flow on the day you stop burning it. That is the essence of a model company.
你烧钱的东西 只有在不烧了那一天 才真正会有现金流 这个就是模型公司的本质
Freda15:09
I think investing is quite important — you have to be honest with yourself and know what money you're earning, beta or alpha.
我觉得投资挺重要的 就是要对自己诚实 知道自己赚的是什么钱 是beta还是alpha
Freda35:23
I think empty talk about whether it's a bubble is too subjective and maybe too hollow, but there are two levels on which you can think about this question.
我觉得空谈是不是泡沫 太主观可能比较空 但是有两个层面 可以去思考这个问题
Freda1:15:04
Figures
| Gemini MAU | 650 million | 21:13 |
| GPT MAU | over 1 billion | 21:13 |
| Anthropic 2B revenue share | over 80% | 28:25 |
| Time for Anthropic annualised revenue to go from 1 billion to 7 billion | under ten months | 29:19 |
| Number of Robinhood business lines | over 11 | 33:21 |
| Robinhood average assets per user | 10,000 dollars | 36:24 |
| Waymo fleet size | 2,500 vehicles | 46:32 |
Glossary
- Crossover fund
- A fund structure where the same team invests in both private and public markets.
- Scaling Law
- The rule that model performance improves as compute and data scale up.
- Single GM
- An organisational structure where each business unit has its own PM, developers, finance and HR.
- Prediction Market
- A platform where people bet real money on event outcomes, such as Kalshi and Polymarket.
- Agentic Commerce
- A shopping model where an AI agent completes selection and payment on the user's behalf.
- Neo Labs
- A batch of large-model startups founded in recent months with funding of a billion or more.
How to listen
Investors and founders watching AI business models, US tech stocks and private-market valuations — especially anyone trying to understand how model companies burn cash and where AI revenue comes from.
The first 3 minutes of self-introduction and the keyword recap of 2021-25 can be fast-forwarded.