The scarcest thing in the AI era isn't information, it's a conversation with a human feel
When 90% to 95% of informational content can be pulled straight from an AI, conversations whose purpose is exchanging information are being hollowed out, and what's left between people is emotional connection.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Token is the most misleading unit there is
Freda says Token itself is extremely misleading, because for the same task, different models or applications can differ by tens or even hundreds of times in Token consumption, and the key variable is Token per task. There are three reasons: output length can vary enormously, since a good model finishes in one or two hundred lines of tight code while a weak model writes thousands of lines of verbose code; many models do a large amount of reasoning tokens before answering that the user never sees; and the agentic workflow itself acts like an amplifier on consumption. So consuming more Tokens does not mean the result is better, and developers have fed back that Cursor is less energy-efficient for the same result.
— FredaTop companies' AI spend is only 1% to 2% of EBITDA
Freda ran the numbers: Meta is rumored to spend several billion a year on Cloud Code, and it announced 10% layoffs, which in order of magnitude just about corresponds to that expense, though he stresses this is not a direct causal relationship. Uber's CFO said the company burned through its entire planned annual AI Budget in three months on Coding alone, on the order of tens of millions. Pull out the top few companies, especially those doing Token Maxing, and everyone's AI spend currently corresponds to roughly 1% to 2% of EBITDA, which is still manageable, and that number is better than he had imagined.
— FredaCharging by Token has to become paying for outcomes
Freda has a fairly strong judgment: the industry will gradually become rational, and the charging model has to change, from charging by Token to paying for outcomes. He cites Sierra doing AI customer service, charging purely for outcomes: it only charges when the AI can resolve the issue without escalating to a human, and charges nothing if it escalates, with different unit prices by problem complexity and degree of resolution, so the customer and Sierra are very interest aligned, and both sides want to burn fewer Tokens. But this works in customer service because it can be quantified; in creative scenarios like writing, outcomes are hard to quantify, so long-term it may still be charged by Token.
— FredaOnly better AI can train better AI
Over the past two years the default assumption was that the SOTA model changes hands every few months: Gemini was hyped at the end of last year, before that OpenAI, now Anthropic, and the lead is on a rolling basis. But Freda says that in the past month or two, for the first time he has started to question whether this assumption will hold, and the core reason is that once coding agents matured, an important loop appeared: only better AI can train the next generation of better AI. Once that gets running it has a whiff of recursive self-improvement, and past a certain critical point the curve becomes very steep, and for those chasing later it may genuinely no longer matter.
— FredaThe model layer isn't SaaS, it encourages using several vendors
Freda reflects that he spent too much time over the past two years comparing OpenAI and Anthropic, when the more important logic is knowing that both will be very large. On business model, SaaS is a per-seat subscription with a capped price, cheaper the more you use it, and customers tend to use only one vendor; models, by contrast, are usage based and token based, and the choice a CFO makes is no longer which vendor to pick this year but which model to use for each query, so the business model in turn encourages everyone to use many vendors at once, comparing and switching at any time. So the relationship between OpenAI and Anthropic is not competition in the traditional sense.
— FredaCoding's TAM was calculated absurdly wrong at the time
Freda says that a year or two ago investors estimated Coding's TAM by multiplying price by volume: the volume was four or five million developers each, and on price some internally said 20 dollars and some said 200 dollars, since a company's total software subscriptions come to at most one or two thousand dollars, which works out to a market of roughly 10 billion dollars. But today Anthropic's revenue alone is far beyond that. He reflects that both the volume and the price estimates were wildly wrong, and that the real TAM is anything a computer can operate, and as Dario said, the white-collar TAM is thirty to forty trillion. The moment he realized this was when Cursor's coding revenue shot up to 1 billion dollars.
— FredaThe negative snowball has another way out
The negative snowball is a concept Dario raised, and Freda translates it into investor language: if training a model cost 1 last year and this year brings in three to four of revenue, at 50% gross margin that's less than two of gross profit, but this year's training cost is at least 3, so adding sales and headcount the profit is negative, and as long as training cost grows several-fold each year, the business rolls ever more negative. Dario's understanding at the time was that the only way out of this loop was for training scaling to slow down, but the past few months have proven there is another path: if the slope of revenue growth is steeper, no longer 3 against 3, the company suddenly becomes profitable too.
— FredaAnthropic does the revenue of tens of thousands of traditional software staff with three thousand people
Freda says software has always been a push model, needing sales and a whole go to market, and ServiceNow and Adobe each have thirty or forty thousand people, Salesforce has seventy or eighty thousand, with sales and marketing at forty or fifty percent of revenue. But Anthropic is also a to B company, 80% enterprise users, with only three thousand employees, and where a traditional software company has each employee corresponding to 500,000 dollars of revenue, Anthropic has each employee corresponding to tens of millions, an order of magnitude apart, and it genuinely doesn't even have a proper sales team. This made him reflect on whether users' demand for intelligence is unlimited, and whether he had previously overemphasized the importance of sales.
— FredaPutting the electric motor where the steam engine was
Freda uses Dario's distinction between technology diffusion and economic diffusion to separate progress in model capability from the speed at which it is actually absorbed into the economy, and studying organizational structure is essentially judging economic diffusion. The historical example is electricity: from the light bulb to a rise in social productivity took 40 years, and in the intervening twenty or thirty years productivity didn't rise and even dipped a little, because factories were still stacked vertically and designed for the steam engine, and everyone simply removed the steam engine and stuffed the electric motor into the same spot, and productivity only went up once the assembly line took shape. He thinks AI is right now at the stage of stuffing the electric motor into the steam engine.
— FredaHierarchy is essentially an information-carrying mechanism
Freda says hierarchy in a company is generally understood as a power structure, but its deeper function is information transfer: once an organization grows, no one can see the whole picture, and you need layers of managers to collect signals from the CEO, synthesize and distill them, and pass them down, and from the bottom up it's the same way of translating and decomposing. Meetings, syncing progress, aligning for a quarter, all are information-carrying mechanisms, because the cost of passing things between people is relatively high. In a tech company, a product goes from the PM writing the PRD, the designer visualizing it, the developer building it over weeks or months, QA over a few weeks, and then go to market, and end to end it can be six months, much of which is translation cost.
— FredaEvery step becomes the new bottleneck
Freda heard a founder describe it: at first what got hit was the developer's build time, which used to take two or three months and after vibecoding was written in two weeks, then QA became the bottleneck, and there was no reason to spend weeks on QA, so the QA people were laid off or their function changed, and then the PM and design in front became the new bottleneck, and after that was fixed go to market became the newest bottleneck, like whack-a-mole, solving problems nonstop. He thinks the whole process actually needs to be redesigned, from a relay race of one baton after another into a three-to-five-person basketball game, where the necessary skills are all inside the team, the team can make decisions directly on its own, and only very large problems get escalated.
— FredaThe hard part of AI trading systems is the data
Freda thinks that if you give an agent sufficiently clean and sufficiently large data and clearly tell it the trading objective, in theory it will do 100% better than a human, and the reason it hasn't been done today is only that people's thinking is also very muddled and they haven't fully deconstructed their own thought process. One difficulty is data: financial data sounds standardized but is in fact very fragmented, and he says wiring an agent to financial data requires connecting to a dozen or twenty vendors to piece together the data you normally use. The other is understanding how the different players behind the market make decisions: in US equity trading volume, quant accounts for over sixty or seventy percent, retail roughly thirty percent, and traditional institutions barely account for any trading volume at all.
— FredaModel companies are reaching into the application layer again
Freda says people have been asking this question since 2022 and the answer keeps changing. In 2023 everyone called them AI wrappers and worried the model companies would eat everything; in 2024 and 2025 people stopped saying this, because application companies' revenue was also growing fast. But perhaps just in the past two months, he has started to worry very much again about competition from model companies, because the models themselves have become very stateful, with skills, able to use tools, with connectors, and the boundary between the application layer and the model layer has once again been pushed toward the application layer. Anthropic has already come out and said it will eat programming first, and its second-largest business will go after the financial sector; OpenAI is about to launch its own audio model.
— FredaWhat worries him most is cloud vendors' cash flow turning negative
Freda says what he worries about most is that the cash flow of the big companies, especially the cloud vendors, will look very ugly going forward. Following today's capex curve, by 2027 the free cash flow of several big vendors will all turn negative, the whole industry has more than a trillion dollars of capex, and these are only the on-balance-sheet numbers, since every big vendor has signed a lot of off-balance-sheet capital expenditure, and big customers have signed long-term contracts with storage companies, so capex forecasts can almost only be revised up, never down. He doesn't know what state everyone will be in when they actually see all-negative cash flow.
— FredaThe cloud business itself has gotten worse
Freda says it's not that cloud vendors get no return: Google has 200 billion of capital expenditure this year and cloud revenue will be roughly 100 billion dollars, which may pay back in two years, and Amazon is similar. But the cloud business itself has gotten worse: first, there is more competition, and second, it has had a lot taken away from it in the compute stack value chain. Cloud used to sell a lot of upper-layer software, and that part had very high margin, but now much of that value has been taken by the model companies. The cloud vendors combined now already have 2 trillion dollars of revenue backlog, more than half of which is long-term compute orders from models like OpenAI and Anthropic.
— FredaThe essence of anxiety is fear of uncertainty
Freda says that honestly he is very anxious, and the essence of anxiety is fear of future uncertainty. His specific daily state is feeling he can't keep up: auto research came out, someone demoed a new workflow, there are new features on cloud every day, and software and hardware are changing too much too fast. In tech investing, when something new comes out you should try it yourself before making a judgment, but the fact is there is too much and it's too fast, and every day you open Twitter and it says I have to look at this, this is changing my life, and there's a feeling of here's something new again and I've fallen behind again. He also describes the pain of failing to install OpenCloud at home late at night, and feeling that if he can't even install this, is he going to be made obsolete by the times.
— FredaConversations whose purpose is exchanging information are being hollowed out
Freda says he used to love booking 30 minutes with industry experts and getting high-concentration knowledge from top people, and that feeling was very satisfying. But now for many work-related conversations, no matter who the other person is, 90% to 95% of the informational content he can get the answer to directly from an AI, and the answer may be the one he prefers. He stresses it's not that he has some sense of superiority and doesn't want to talk to people, but that he genuinely feels conversations whose purpose is exchanging information are rapidly being hollowed out. What's left between people is emotional connection: once he sat down with a friend with no plan to talk about anything big, and as they talked it drifted to the courage to do things, life's regrets, what you care about, and looking back it was the most substantive conversation he'd had in a long time.
— FredaIn their own words · checked verbatim
I think Token is also a very misleading unit, because the most important thing here is Token per task — for the same task, different models or different applications may differ by tens or even hundreds of times in Token consumption.
我觉得Token其实也是一个非常容易误导人的这么一个单位 因为这里面最重要的就是Token per task 就是说你同样一个任务 不同的模型或者说不同的应用 消耗的Token可能差几十倍甚至上百倍
Freda3:03
It's that only better AI can train the next generation of better AI, so once this loop gets running, it has a bit of that flavor the researchers call recursive self-improvement.
就是你更好的AI 才能去训练出 下一代更好的AI 所以就是这个循环 一旦跑起来 就有一点点 就是那些researcher说的 叫recursive self-improvement 这个味道
Freda12:05
But the model is definitely usage based and token based, so as a CFO the choice you make is no longer which vendor I pick this year, but you're optimizing which model I use for each query.
但是模型它 就肯定是一个usage based 然后是一个token based 所以你作为CFO 就是你做的这个选择 不再是说 我今年选择哪个vendor 而是你在去optimize 就是我每一个query 我用的是哪一个模型
Freda16:05
But the bigger shock is that Anthropic is also a to B company, 80% of its users are enterprises, and it has only three thousand employees.
但是就是比较大的震撼 就是entropic其实也是一个土币的公司 它80%都是企业的这个用户 员工就只有三千人
Freda28:13
The electric motor was actually invented very quickly, but it wasn't really used on the assembly line, it didn't really raise efficiency — everyone still just removed the steam engine, swapped it out, and stuffed the electric motor into the same spot.
就其实电机很快就被发明了 但是并没有就是真的用到流水线上 并没有真的提升效率 就是大家还是哦 把蒸汽机拆掉 然后换出去 然后把电机塞到同样一个位置
Freda37:15
I think what I worry about most is that going forward the cash flow of these big companies, especially the cloud vendors, will look very ugly.
我觉得我最担心的就是接下来 这些大厂 尤其是云厂商的现金流 会非常的难看
Freda1:01:28
Rather, I very genuinely feel that the meaning of conversations whose purpose is exchanging information is being rapidly hollowed out.
而是说 我 我很真实的觉得 以信息交换为目的的对话 这个意义是在 快速的被掏空的
Freda1:19:40
It didn't give me any informational increment, but it was more human.
就他没有给我 任何信息上的增量 但就是比较有人味吧
Freda1:20:40
Figures
| Anthropic Opus nominal price | 5 dollars per million input tokens, 25 dollars per million output tokens | 7:04 |
| Anthropic cache hit rate | possibly as high as 98%, 99% | 7:04 |
| Top companies' AI spend as share of EBITDA | 1% to 2% | 9:04 |
| Meta layoff percentage | 10% | 9:04 |
| Early estimated market size for Coding | about 10 billion dollars | 17:06 |
| White-collar TAM cited by Dario | thirty to forty trillion | 18:07 |
| Anthropic employee count | three thousand | 28:13 |
| Quant share of US equity trading volume | over sixty or seventy percent | 45:18 |
| Revenue backlog of several cloud vendors | 2 trillion dollars | 1:03:30 |
Glossary
- Tokenmaxxing
- The practice of taking pride in consuming more Tokens and treating usage as a measure of capability or investment.
- recursive self-improvement
- Better AI training better AI, forming a self-accelerating loop.
- economic diffusion
- The speed at which model capability is actually absorbed into businesses and the economy and converted into revenue.
- Neo Labs
- New model labs started by researchers who left AI Labs, currently a hot investment theme in Silicon Valley.
- Acquire Hire
- Big companies buying small teams to bring the people back in, giving strong founders an exit mechanism.
How to listen
Investors watching AI business models and valuations, people working at model companies, and founders thinking about how organizations should change in the AI era.
If you only want the investment substance, the anxiety and loneliness section after 1:14:17 can be skipped.