Freda's Investment Notes Ep. 2: Tokenmaxxing and Human Connection
Summary
Token consumption is not value; the real comparables are task effectiveness, token per task, and dollar per token. On the same coding task, an efficient model may write 100-200 lines while a weaker one writes several thousand, with hidden reasoning tokens and agentic workflows potentially creating differences of dozens of times or even more than 100x in consumption. Freda’s core reminder: “Tokens are not all the same.”
Tokenmaxxing contains enormous waste, but total AI spending will continue to surge—both statements can be true. User counts and tasks per user are expanding, while token per task and dollar per token will keep improving; Meta’s several-billion-dollar external-model bill and Uber exhausting its full-year coding AI budget in 3 months still amount to only roughly 1%-2% of EBITDA at leading companies. Over time, pricing will shift from tokens to outcomes, with Sierra’s model—charging only when customer service is resolved and not when a case is escalated—providing the clearest precedent.
Anthropic’s lead may, for the first time, stop being a rotating title because coding agents have created recursive self-improvement: better AI training the next generation of AI. Engineering productivity gains have risen from roughly 5% to 15%-20% by the end of last year and are now higher, while model iteration cycles have compressed from months to roughly 1 month; without a top-tier coding engine, challengers will develop at a lower cadence. Freda’s analogy is horse-drawn carriages versus cars: “Once the car can run reliably,” a fast horse is no longer a meaningful comparison.
The model layer is not traditional SaaS but a usage-based intelligence market, allowing OpenAI and Anthropic to both become enormous while customers route individual queries across multiple vendors. The early estimate of a roughly $10B coding TAM, based on 4M-5M US developers paying $20-$200 per seat, was “spectacularly wrong”; the real market is closer to Dario’s $30T-$40T in global white-collar time. The key investor skill is not never being wrong, but immediately discarding the old framework when Cursor’s revenue races toward $1B.
AI’s impact on software is not just feature replacement: it raises uncertainty, compresses valuations, and forces organizations to shift from a relay race to a basketball game. E-signature, project management, and BI products with high UI exposure and orderly data structures are most vulnerable; Excel and Office retain resilience because their semantics are messy and their outputs require human refinement. The real organizational redesign is to use autonomous 3-5-person teams to eliminate translation layers, rather than “putting the motor in the steam engine” while preserving an 80,000-person, multilayered hierarchy.
Markets will become both more efficient and more volatile as traditional information-arbitrage alpha is rapidly erased by agents. Event-driven trades may be priced within 1 second of an announcement, and earnings analysis will become systematized; crowded agents will also amplify the same signal. Future alpha will rely more on understanding long-term, cross-industry trends, while retail preferences for low PE, big dreams, charismatic founders, and low absolute share prices must become part of institutional models.
The biggest US equity risk is not a lack of AI demand, but cloud providers collectively turning free-cash-flow negative by 2027 under trillion-dollar capex. Google could still recoup roughly $200B of capex against about $100B of cloud revenue within 2 years, but off-balance-sheet commitments, long-term storage contracts, and capital rotation into new IPOs will continue to pressure the Mag 7; if SpaceX, Databricks, and OpenAI list at a combined valuation of roughly $4T, the market may absorb the issuance even if incumbent mega-cap weights cannot.
AI is hollowing out conversations whose purpose is information exchange while making emotional connection between people scarcer and more valuable. Freda admits she once cried late at night after failing to install OpenClaw, while also feeling the rush of “suddenly it seemed like anything was possible” beyond the anxiety. When roughly 95% of informational content can be queried from AI, her answer is simple: “What remains between people is emotional connection.”
Deep dive
1. The Token Explosion Is Hiding the Real Unit Economics of a Task
Freda compares dollar per token with the industrial-era dollar per kilowatt-hour: it measures input-output efficiency, but must be viewed alongside token per task; comparing token revenue or consumption without considering task outcomes can easily mistake inefficiency for demand.
Coding offers the clearest example: on the same task, a strong model may complete the job with 100-200 lines of concise code, while a weaker model writes several thousand lines of redundant code. Both appear to have “completed the task,” but their costs and downstream maintenance burdens are in completely different leagues.
Hidden reasoning tokens distort the comparison further; agentic workflows act like an amplifier by repeatedly calling models and tools. It is therefore not unusual for the same task to generate differences of dozens of times or even more than 100x in token usage across models and applications.
Freda’s decomposition of total spending is: users × tasks per user × token per task × dollar per token. The first 2 variables will surge as AI expands from programmers to ordinary white-collar workers, while the latter 2 will keep improving, so “waste falling” and “total usage exploding” can both be true.
2. Tokenmaxxing Will Fade, but It Will Not Take AI Spending Growth With It
Cursor exposes the early illusion: when the same outcome consumes more tokens, the CFO sees credits surging alongside higher employee productivity and keeps increasing the order. “Employees like using it” and “the product is efficient” are often conflated in today’s reporting.
Anthropic Opus’s list price is $5 per million input tokens and $25 per million output tokens. Freda cites third-party analysis suggesting its realized price may be closer to roughly $1, partly because cache-hit rates can approach 98%-99% and agent workflows have far more input than output.
Meta is reportedly spending several billion dollars a year on external models, a figure roughly matched in order of magnitude by the savings from a 10% workforce reduction; Uber used up its entire annual AI budget on coding alone within 3 months, at roughly tens of millions of dollars. Freda estimates that AI spend at the leading tokenmaxxing companies currently represents only about 1%-2% of EBITDA and remains manageable.
When 张小珺 asked whether the token explosion would naturally fade, Freda’s answer was “both yes and yes”: there is clearly waste today, but user and task growth will outpace efficiency gains. No one boasts about a light bulb consuming more electricity; the market ultimately rewards an “LED” that costs more but uses less power.
3. Outcome-Based Pricing Aligns Supply and Demand Better Than Token Pricing
Freda’s strong commercial view is that, as the industry rationalizes, pricing will gradually shift from tokens to outcomes because customers are buying problem resolution, not the amount of intermediate reasoning a model performs.
AI customer-service company Sierra is her preferred case study: it charges when AI resolves the issue without handing the customer to a human, and charges nothing once a case is escalated, with prices varying by complexity and degree of resolution. Customers and vendors are thereby aligned around higher resolution rates and less token waste.
The shift will come first in measurable settings such as customer-service resolution, sales conversion, loan collections, and insurance claims; creative work such as writing makes “success” harder to define and may remain token- or usage-priced over the long term. Outcome pricing is not a universal answer, but a path that will spread outward from industries where outcomes can be measured.
4. Coding Agents Are Turning Model Leadership Into a Recursive Advantage
For the past 2 years, investors have assumed that SOTA changes hands every few months, with OpenAI, Gemini, and Anthropic rotating through the lead. Freda has recently begun questioning that rolling-basis assumption for the first time because coding agents give the leader a loop in which better AI trains the next generation of better AI.
She is careful to say that recursive self-improvement has only just passed an early milestone, perhaps becoming visible around January this year. If the loop crosses a critical threshold, the curve could steepen suddenly and the chasers may lose the ability to catch up.
6 months ago, coding models may have lifted engineers’ total factor productivity by only about 5%; by the end of last year the gain had reached 15%-20%, and it is “definitely higher” now. Model release cycles have compressed in parallel, from several months or even a year to roughly 1 release per month.
Freda’s analogy is the transition from horse-drawn carriages to cars: early cars broke down constantly, so a fast horse could still compete; “once the car can run reliably,” a horse without an engine is no longer comparable. OpenAI’s restructuring of its GPT and coding teams, Sundar Pichai’s direct oversight of coding at Google, Meta’s SOTA push, and xAI’s public bid for Cursor all follow the same logic.
5. The Codex-versus-Claude Code Debate Actually Shows How Close They Are
On the debate over whether Codex or Claude Code is stronger, Freda’s practical judgment is that when 2 products are compared continuously and at scale, their real-world performance is usually not far apart. As a late entrant, Codex is using subsidies and opening up Claude-restricted use cases to wage a price war—a rational catch-up strategy.
She reflects that she spent too much time comparing OpenAI with Anthropic and overlooked the fact that “both will become very large.” Traditional SaaS is seat-based, with a price ceiling that encourages customers to pick a single vendor; models are usage-based, so the CFO can optimize model selection query by query, naturally supporting multiple vendors and constant switching.
The relationship is therefore not a traditional zero-sum contest: one model may be better for coding today, another for visualization or financial tasks tomorrow, and enterprises can call both. At the model layer, the fight is for a larger share of the “intelligence” in the compute stack, not a single annual vendor contract.
6. The Old Coding TAM Formula Underestimated Both Volume and Price
1-2 years ago, investors estimated a roughly $10B coding TAM by multiplying 4M-5M US developers by $20-$200 per seat. Some even believed a company could spend only $1,000-$2,000 on all software subscriptions, making $200 look aggressive. In hindsight, the framework was “spectacularly wrong.”
When Cursor’s coding revenue rapidly reached roughly $1B, a single company had already captured 10% of the old TAM; that should have triggered an immediate rewrite of the framework. Freda’s standard for investing is not to be right forever, but to recognize “very quickly” that the original investment framework and TAM were fundamentally wrong once the revenue curve disproves the assumptions.
The new framework is closer to Dario’s $30T-$40T in global white-collar time: any task that amounts to “if I could code, I would have a computer do this automatically” is potential coding TAM, including Excel, trading, and more digital workflows to come. Cowork being built by just 2 people is a snapshot of the changing software production function.
7. The “Negative Snowball” Now Has a Second Escape Route
Dario’s “negative snowball” can be simplified as follows: last year’s training cost was 1, while this year’s model generates 3-4 of revenue; at roughly a 50% gross margin, less than 2 remains as gross profit, but training the next generation will cost at least 3, before sales and headcount are added, leaving the company loss-making.
The old solution was simply to slow training scaling; the past few months have revealed a second path—revenue growth is no longer merely 3, but far above 3. Exploding demand is forcing companies to redirect more pre-committed compute from training to inference, and if training costs fall below inference gross profit, profitability can arrive suddenly.
Freda emphasizes that companies can manage inference gross margin, but demand ultimately determines the split between training and inference compute. Anthropic built its forecasts on “extremely first-principles” analysis and still failed to fully anticipate the demand explosion at the start of the year; under supply constraints, unexpectedly strong demand can actually push margins higher.
This also explains why model companies themselves may not understand the end state before the market does: detailed forecasting can estimate costs, but cannot tell you in advance how many white-collar tasks a new product will convert into real calls.
8. ARR Definitions Are a Mess, While the Revenue Ceiling per Gigawatt Remains Invisible
Media-reported ARR is often neither annual nor recurring: some companies multiply the past 4 weeks of revenue by 13, while others multiply 1 week by 53, producing materially different results. Anthropic reports gross revenue and OpenAI reports net revenue, so direct comparisons create false precision.
In Freda’s illustrative adjustment, if roughly 40% of Anthropic’s revenue comes from third-party APIs, converting to net could reduce the figure by about 20%; converting OpenAI from net back to gross could increase it by roughly 5%-10%. She considers EBITDA more comparable and admits she “wasted a lot of time” on these definition changes.
OpenAI and Anthropic may each have several gigawatts of compute today, reaching 5 GW by year-end and potentially 10 GW next year. The market once viewed $10B of revenue per GW as a strong outcome; it now appears that $10B or even $50B may be possible because nobody has actually identified the upper bound on revenue per unit of power.
Anthropic’s real investment inflection point was the jump from $1B to $5B of ARR in less than 6 months, faster than OpenAI over the same period. Gross margins looked ugly and government relationships were a concern, but conversations with infrastructure engineers suggested there was substantial room for optimization.
9. Model Giants Will Converge; Monetization Remains the Key for xAI and Meta
Freda once expected OpenAI, Anthropic, and Gemini to specialize rapidly; she now expects their business models to converge: “Whatever takes off, whatever generates revenue, everyone will do it.” She herself switched away from GPT projects containing accumulated tax and medical-checkup records after Anthropic added visualization.
She even considers combined revenue of roughly $200B for OpenAI and Anthropic by year-end “not particularly strange.” The point is to train investors to adapt to larger numbers and upside surprises, not to make a firm forecast.
On xAI, she gives the non-answer that there are “not enough data points to judge success or failure.” Leasing a roughly 300 MW H100 data center—about 20% of its total—to Anthropic, which is short of capacity, could generate an estimated $4B-$5B of revenue, but that would simply monetize idle inference assets and would not mean xAI has abandoned training to become a neocloud.
Meta’s year-end TBD should be an attempt to reach SOTA, with Gemini 2.5 perhaps the current benchmark. The real question is how to monetize the model once it is built. Meta AI, OpenClaw-style assistants, and internal Cody only become a new AI revenue stream independent of advertising if they can be sold externally.
10. AI Is Making Traditional Software Headcount and Sales Models Look Irrational
ServiceNow and Adobe each have 30,000-40,000 employees, while Salesforce has 70,000-80,000; sales and marketing may represent 40%-50% of revenue. Traditional enterprise software depends on push and a full go-to-market machine. Anthropic, by contrast, gets roughly 80% of its users from enterprises with only 3,000 employees and no traditional large sales force.
By Freda’s rough math, traditional software generates about $500K of revenue per employee, while Anthropic generates more than $10M—a full order-of-magnitude difference. Cowork, skills, and projects are not inherently easy to use, yet users still teach themselves, prompting her to ask: “Is demand for intelligence unlimited?”
Meta is already considered highly efficient, but still has roughly 80,000 employees and generates about $3M of revenue per employee. As a consumer platform without constant customer-acquisition pressure, why does it need so many people—and why are her friends still broadly so busy? Those are concrete questions for organizational redesign in the AI era.
Public software equities are broadly down more than 50%, while many private-market stars remain at pre-AI valuations, creating an inversion. If a stock price halves but the company wants to maintain the same level of employee equity compensation, it must issue twice as many shares, making SBC an even larger source of dilution and operating pressure.
11. Software Valuations Lost First to Uncertainty, Then to Product Replacement
Freda explains the selloff through DCF: the market does not need to know that a company will definitely be replaced; if uncertainty around future cash flows rises, the discount rate rises and present value falls. That is why software was re-rated before a clear revenue shock appeared.
Software with UI at the center of its value proposition and a narrow point solution is most exposed; e-signature, project management, and parts of BI may be hit first. These products primarily offer polished interfaces, while an agent can bypass the interface and complete the task directly.
The more orderly the data structure, the easier the product is to replace with an agent. Each row in Monday represents a fixed task object and each column has a clear meaning; Excel mixes names, numbers, formulas, and different semantics across arbitrary cells. “Everything is an Excel wrapper” does not mean every wrapper has the same defensibility.
When 张小珺 asked whether Office had become meaningless, Freda revised her earlier view: AI can generate an HTML presentation quickly, but as long as one person in the workflow still has to drag elements around and refine the output manually, PowerPoint remains more precise. Office’s value may shift from a creation tool to the standard editing surface for human-machine collaboration.
12. Even Data Warehouses Are Not Absolutely Safe; the Real Opportunity Is in Decision Residue
The market generally views data warehouses, data lakes, and storage as safer assets. Freda offers a counterexample: valuing 50 stocks once required storing the data in a warehouse and writing precise SQL; now a skill can be written in the cloud to call the Bloomberg API and finish the task in seconds with near-zero setup cost.
She believes AI CRM and AI ERP are not transformative if they merely replace manual entry with voice input. The larger opportunity is to capture the decision process that legacy systems never recorded, rather than storing only the final compressed result.
CRM may record that a customer received a 25% discount, but not why it was not 30%, who vetoed what, who persuaded the CFO, or what the customer was actually worried about. AI can process meetings, emails, negotiations, and other unstructured data directly, turning these “vanished processes” into searchable, inferable enterprise assets.
13. Software Must Rebuild Its Protocols for Agents, Not Just Add a Chat Window
Freda found that when connecting agents, Slack’s HTTP request-response model was adequate for human chat but ill-suited to agents that need persistent, real-time, always-on connectivity; Discord’s WebSocket architecture was much easier to work with. She even considers it unsurprising that a model company might acquire Discord.
The observation points to a full-stack rebuild: browser, identity, payment, compliance, email, and phone systems were all designed around human call frequencies and permission boundaries, while agent concurrency, continuous connectivity, and autonomous execution run into entirely different constraints.
AI-native software therefore does not embed a copilot in an old product; it redesigns interfaces, state, authorization, and auditability from the question of what an agent needs. Market trades around CPU shortages can also be viewed as challenges that the same new calling patterns are imposing on infrastructure.
14. Companies Are Still Putting the Motor in the Steam Engine, Not Completing Economic Diffusion
Dario distinguishes technology diffusion from economic diffusion: the former is the clear, steep curve of model capability gains; the latter is the absorption of that capability by businesses and the economy into revenue and productivity, which may lag by years or decades and is far harder to time.
It took roughly 40 years for the light bulb to produce a clear improvement in social productivity, with no improvement at all for 20-30 years in between. Early factories simply removed the central steam engine and put electric motors “in the same place,” preserving a vertical structure designed around belt transmission; electricity released its economic value only after the assembly line was redesigned.
The 20th century also produced its own productivity paradox: offices and banks were already using computers in the 80s, and individuals felt faster, but macro productivity did not move. Productivity jumped only in the mid-to-late 90s, when Walmart, Amazon, and others rebuilt their businesses around databases, early ERP, networks, and supply chains.
Freda believes AI is still in the “putting the motor in the steam engine” phase: everyone is adding models to existing workflows, but few are asking why the processes were built this way or why a company needs 80,000 people and so many layers.
15. Organizational Hierarchies Are Fundamentally Information Transport; AI Will Turn the Relay Race Into Basketball
In Freda’s first-principles breakdown, hierarchy is not merely a power structure; it is a mechanism for compressing and transmitting information. The CEO’s signal is synthesized by managers and passed down layer by layer, while frontline information is translated on the way up. Meetings, quarterly alignment, and progress updates are all costly forms of information transport.
A typical technology product flow has the PM writing a PRD, the designer visualizing it, developers building for 2-3 months, QA testing for several weeks, and sales handling go-to-market; end to end, it can take 6 months. Much of that time is not creation but one person translating what another person “actually wants to do.”
Once AI compresses development from 2-3 months to 2 weeks, QA immediately becomes the bottleneck; after QA is rebuilt, PM and design become the slowest steps, followed by go-to-market. This “constant whack-a-mole” of local optimization shows that the entire process must be redesigned together.
She envisions a future in which the “relay race, one baton at a time” becomes a basketball game played by 3-5-person teams: skills are concentrated in autonomous units, QA is embedded in development, PMs become more generalist, and most decisions are made locally, with only major issues escalated. Middle management will be compressed, though she has not yet seen a broadly recognized successful US-company template.
16. The Challenge of Investing Agents Is Not Intelligence but Messy Data and Messy Human Objectives
The investment industry spends enormous time finding information, cleaning data, comparing expectations, and assessing positioning. Freda believes that, with sufficiently clean data and explicit instructions on holding period, target return, maximum drawdown, and exit conditions, an agent would “theoretically be 100%” better than a human.
That has not happened primarily because financial data are far more granular and fragmented than they appear; when Freda actually connected the systems, she needed more than a dozen and as many as 20 vendors to assemble the data she uses every day. Human investors have also failed to fully deconstruct their own styles and decision processes, leaving them unable to describe the ideal trade to an agent.
In her rough breakdown of US equity trading volume, quant accounts for 60%-70% or more, retail about 30%, traditional directional institutional trading far less than imagined, and platform institutions only single digits. If high-frequency trading mainly reinforces momentum, the number of players actually determining market direction is smaller than people think.
Retail preferences must also enter the model: retail investors often look at low PE, favor big dreams, charismatic founders, and low absolute share prices; at the same market cap, an $8 stock is often more attractive than an $800 stock. Tesla, SpaceX, and Robinhood all carry this narrative; ultimately, “retail likes stocks whose prices go up.”
17. Markets Will Become More Efficient and More Volatile at the Same Time, With Event Alpha Evaporating Fast
Freda expects the information gap between retail and institutions to close rapidly, while fundamental research and medium-frequency quant strategies converge. Institutions will no longer simply “educate retail,” because in thematic, style-driven, narrative-driven markets, retail is itself part of the price-formation mechanism.
After OpenAI announced a partnership in the past, institutions might spend several days calling contacts and checking numbers before the stock moved to consensus. An agent can read historical cases alongside current information and may price the announcement within 1 second of its release. Event-driven trading will therefore become “increasingly almost unnecessary.”
Faster reactions will amplify the same signal, making trades more crowded and volatile; faster information incorporation will also make markets more efficient. Real alpha will shift from “I discovered it slightly earlier” to “I understand the larger trend better,” increasing the value of cross-industry research.
Alpha from covering overlooked small and mid-cap names may also be erased by AI’s low marginal cost of learning. Earnings analysis will gain something close to a god’s-eye view, and earnings days may no longer produce the large error-correction moves seen today.
18. Application Revenue Has Segmented, but the Model Layer Is Moving Up the Stack Again
In the revenue map, coding is the largest category, already generating several billion dollars; healthcare companies such as Abridge and OpenEvidence are around $2B, legal companies such as Harvey and Legora are above $1B, customer service is near $1B, and video generation generates several hundred million dollars.
AI inference infrastructure companies such as Together and Fireworks already generate several billion dollars; chip companies including Cerebras and the acquired Groq are also among the larger players. Beyond these categories, far fewer AI startups have reached the $1B scale.
In 2023, the market worried that AI wrappers would be swallowed by the models; application revenue growth and exits in 2024-2025 eased that concern. Freda has turned cautious again over the past 2 months because models are becoming stateful, with skills, tools, connectors, and memory, pushing the application boundary outward again.
Anthropic has said it will take coding first and then enter finance, while OpenAI’s audio model could pressure companies such as ElevenLabs. Legal applications emphasize compliance and hallucination, while customer-service companies emphasize the last mile; Freda’s reservation is that businesses focused only on regulation or edge work may capture neither the fattest value pool nor attractive long-term margins.
19. VC Is Using Portfolio Math to Bet on New Labs While Rebuilding Agent Infrastructure
Freda continues to invest in applications. One practical reason is that nearly every VC she knows has launched a growth fund to take late-stage assets such as OpenAI and Anthropic, leaving existing VC funds with more available capital. With IPOs, acquisitions, and exits also increasing, startup financing is not difficult.
Silicon Valley may already have close to 100 New Labs. Once OpenAI’s valuation approaches $1T, investors will ask how much more can be made by continuing to hold it; a single new lab producing a multibagger return makes the portfolio worth backing, while strong teams can often return to leading labs through acqui-hires, reducing exit risk.
This is not a business of accurately picking the one winner; it is fund-size math. If a fund invests $50M in each of 4-5 companies, total exposure is roughly $200M. Only funds with several hundred million dollars of capital are suited to cover the category, while smaller funds cannot diversify adequately.
Another opportunity is reducing inference costs and building agent-native infrastructure. Gmail’s API send limits are sufficient for humans but can easily trigger blocks for agents, creating companies such as AgentMail and AgentPhone. Browser, identity, payment, and compliance systems may all be rebuilt around machine users.
20. The Real Prize in Agentic Commerce Is Enterprise Procurement, Not Consumer Web Browsing
Compared with a few months ago, Freda is more convinced that To B is larger than To C: consumers often enjoy browsing products, so the incremental value of having AI place the order is limited. She sees OpenAI taking a more cautious approach as well, entering through travel rather than broad consumer commerce.
Cross-border enterprise procurement is entirely different. A US company buying office sofas remotely must simultaneously handle quantity, unit price, volume discounts, certifications, import-export compliance, payment terms, and early-payment discounts—a long chain involving multiple parties.
Her judgment is that “any scenario with high communication costs is naturally very well suited to agents.” The value is not clicking buttons on someone’s behalf, but having an agent negotiate across systems, verify conditions, handle exceptions, and complete the transaction loop.
21. Autonomous Driving and Robotics Are Both Slower Than Expected, and No Clean Scaling Law Has Emerged
Freda explicitly revises her view: autonomous driving and robotics overall are progressing more slowly than she expected. The most important autonomous-driving development is Nvidia’s open-source model Alpamayo; if it becomes the Android for automakers, capabilities could spread rapidly beyond a small group of leaders and rewrite the industry structure.
Autonomous driving is also beginning to emphasize reasoning, because simply adding data and using imitation learning cannot cover every edge case. The model must understand what is happening on the road, not merely reproduce driving actions in its training data.
In robotics, “more data produces improvement” does not amount to a scaling law. A true scaling law requires a clean, extrapolatable decline in pre-training loss, with generalization across tasks, environments, and embodiments. Robotics currently looks more like continuous feature improvement, without an LLM-style curve.
Hardware is another constraint, with US companies traveling to Shenzhen to learn over the past few months. Freda believes that if US-China relations were somewhat better, investing in some Chinese robotics companies would be a reasonable decision.
22. Trillion-Dollar Capex Can Produce Returns, but Free-Cash-Flow Pressure Matters More
Freda revisits last year’s calls: treating OpenAI and Anthropic revenue as a market barometer for this year was right, while expecting the market baton to pass from semiconductors to AI applications was “completely wrong.” Anthropic’s strong revenue has in fact supported the current AI narrative, with hardware still at the center.
Her biggest concern is that, on the current capex trajectory, several large cloud providers will all turn free-cash-flow negative by 2027. Industry balance-sheet capex has already exceeded $1T, with extensive off-balance-sheet commitments as well; long-term contracts with storage companies mean forecasts are “almost only going to be revised up, not down.”
The investment returns are not imaginary: if Google spends roughly $200B on capex this year against about $100B of cloud revenue, the investment can be understood as paying back in 2 years, with Amazon in a similar position. But competition in cloud is intensifying, while model companies are taking the high-margin value from software above the infrastructure layer, so the quality of the business is still deteriorating.
Cloud providers collectively have roughly $2T of revenue backlog, more than half of it in long-term compute orders from model companies such as OpenAI and Anthropic. At the same time, they are turning internally developed chips into external products: Google has begun selling TPU directly, a major turning point in a multiyear strategy, while Amazon’s Trainium has reached the scale of an independent chip company. Neoclouds and model companies are taking value from the compute stack.
23. Advertising, IPOs, and Layoffs Will Reallocate Market Weight
Freda revises her view that OpenAI advertising could only take share from Google and Meta’s existing pools. Amazon has quietly built an advertising business approaching $100B, much of it previously classified as trade and promotion—for example, brands paying for more prominent shelf placement at Walmart rather than allocating the spend to a traditional advertising budget.
SpaceX may list first, followed by Databricks and OpenAI. At a combined valuation of roughly $4T, a 5% financing would require about $200B and imply a float of roughly $400B. Global active mutual funds hold about $1T in cash, and adding retail, sovereign wealth funds, and hedge funds, the market can “absorb it.”
The bigger issue is rotation: investors will free up capital for new IPOs, putting existing Mag 7 holdings under pressure, especially as their free cash flow deteriorates at the same time. The original 7 giants will be diluted by a new generation of mega-cap companies rather than seeing the market simply expand with all large stocks rising together.
Layoffs have dramatically different earnings elasticity across companies: a 20% reduction at the Mag 7 might lift EPS by only about 5%-10%, while a 20% reduction at a labor-intensive SaaS company such as Salesforce could translate into 40%-50% EPS growth.
24. The Employment Shock Cannot Be Called Definitively; Diffusion Speed and Breadth Will Decide the Outcome
Asked whether AI must cause layoffs and deflation, Dario Amodei responded that holding this view would be “arrogant and not open-minded.” Freda uses the line to acknowledge that she has no basis for a definitive answer; “this time is different” is itself one of the most dangerous phrases in investing.
Historical shocks were not gentle: the power loom displaced hand weaving, agriculture’s share of the US labor force fell from roughly 40% to 1%, hundreds of thousands of telephone operators were automated away, and the Rust Belt suffered from trade restructuring. But the agricultural transition took roughly 100 years, and operator automation also unfolded over decades, leaving society time to adapt.
The counterexamples matter as well. In 2015-2016, Hinton believed deep learning would make radiologists unnecessary; to date, job openings and compensation in the field have instead risen. The “software is dead” debate is currently intense, while developer hiring data remain at historical highs. Freda therefore insists on taking it one step at a time.
张小珺 noted that, in a bearish scenario, governments could respond through money printing, universal basic income, or taxation, so mass unemployment would not necessarily produce long-term deflation. Freda expects AI diffusion to be faster, but its coverage of white-collar work may not exceed the breadth of the agricultural revolution; both portfolios and psychology should leave room to adjust to multiple paths.
25. Long-Term Productivity Optimism Does Not Mean Founders Should Wait for the End State
Freda remains unequivocally optimistic over the long term: technologies that materially increase productivity ultimately benefit the economy. Incumbents are no longer absolutely safe, financing remains available, coding sharply lowers product costs, new demand continues to emerge, and the field remains unsettled—making this a strong window for entrepreneurship.
Her counter-advice to founders is “don’t listen to anything I said above.” The preceding framework mainly applies to large companies with clear business models; early-stage projects cannot easily prove a moat in advance, and there is almost nothing a large incumbent truly wants to do that it absolutely cannot do.
Startups cannot be launched on end-state certainty. She would rather accept “a little irrational excitement, followed by serious preparation and hard work—if you want to do it, do it.” This does not deny the risks; it recognizes that before a new market forms, perfect analysis often eliminates the action itself.
26. AI Is Hollowing Out Information Exchange, Making Genuine Connection Scarce
Freda does not sugarcoat Bay Area anxiety: “Honestly, it’s extremely anxiety-inducing.” New models, workflows, software, and hardware appear every day, while Twitter repeats “you have to look at this” and “this is changing my life,” constantly producing the feeling of falling behind again.
She once cried late at night after failing to install OpenClaw: multiple components had to be connected, JSON had to be written, API keys had to remain private, and she could not even find the cursor. A single installation failure quickly expanded psychologically into: “I can’t even install this, I’m falling behind on everything—am I about to be left behind by the era?”
She uses “strong opinions loosely held” to navigate an uncontrollable future: she can hold strong views on layoffs, economic effects, and how long Anthropic’s lead will last, but must hold them loosely. Intelligence has long-term equalizing power, allowing a child in a small town to access tools close to those available to a Silicon Valley engineer; resource redistribution, however, may make that equalization “very painful.”
In work conversations, roughly 95% of informational content can already be asked of AI, with answers tailored more closely to the need. Freda believes that “conversation for the purpose of information exchange is rapidly being hollowed out,” so meeting people should shift from acquiring knowledge toward observation, inner journeys, and lived experience.
27. What Remains Between People Is Emotional Connection That Cannot Be Compressed Into an Answer
Freda recalls sitting casually with friends in Palo Alto with no agenda, then finding the conversation moving toward the courage to act, regrets in life, what matters, and what still creates a spark. The exchange added no information, yet it was one of the most substantive conversations she had had in a long time because it felt “more human.”
When 张小珺 asked what remains between people, Freda’s answer—while calling it somewhat sentimental—ran through the entire episode: “What remains is emotional connection.” Knowledge, facts, and analysis can be handed to AI; meeting people becomes simpler, centered on things so sincere that they might once have been embarrassing to say aloud.
On the title “Language Is the World,” she ultimately offers a new interpretation: language once carried much of the work of translating and transmitting information, and AI will take over that part; the part of language used to connect people will not disappear, but become increasingly valuable precisely because it is scarce.