Summary

The bubble question has two layers. Freda says asking whether something is a bubble is too subjective: today, this is definitely not a bubble, because GPT has existed for less than 3 years, user adoption has already reached the equivalent of a decade of internet progress, large companies’ ROIC from AI investment is improving quarter by quarter, and Gemini 3 has shown that pre-training has not hit a wall. Whether it later turns into a bubble depends on just 2 things: whether models keep improving and whether AI revenue keeps rising.
Large models are a negative snowball. Training costs rise 10x a year under the scaling law: if last year’s cost was 1, the second year’s 2 of revenue pays it back, but the next generation costs 10, leaving “plus 2 minus 10 equals negative 8.” There are only 2 ways to turn positive: increase the revenue multiple, or stop burning money on models 10x larger. “What you spend money on only produces real cash flow the day you stop spending it”; when the scaling law slows, the income statement will display “violent beauty”—which is good news for investors.
OpenAI’s non-consensus case is enterprise. The market treats GPT as purely To C, but individual and enterprise users are now close to evenly split, with millions of enterprise users; building a super-portal on the enterprise side “will develop very quickly.” Advertising is “too logical,” while cloud is a natural AWS-style extension. Watch dilution: valuation rose from $30B to $500B, but per-share value rose only 6x, not 20x. A $1T IPO in January 2027 would be reasonable at 10x revenue, and going public does not mean the peak—Netscape listed in 1995, while the bubble burst 5 years later.
Google and OpenAI pose real threats to each other. Google can use TPU vertical integration to wage an absolute price war and bundling to take share of GPT users’ time; the rational end state is Google taking the $20 low and mid-market while GPT keeps the high end. But insurance and law firms, Google’s biggest advertisers, have questions “too well suited to asking GPT,” and the landscape changes once GPT starts selling ads. Search moats have been fundamentally altered: models “painstakingly turn through dozens of pages in 1 second,” so Bing versus a self-built engine may not matter much.
Robinhood played a bad hand into the S&P 500’s top performer. Securities trading is a cyclical, structurally poor business, but diversification—11+ business lines and more than $100M in revenue—share gains, pricing power, and flat operating costs have produced strong alpha. Coinbase has had no alpha against Bitcoin since 2022, while Robinhood has alpha against both BTC and the Nasdaq. When Coinbase was valued at $100B and Robinhood at $30B early this year, Freda bought Robinhood repeatedly against the trend; it has risen nearly 3x this year. The long-term thesis is “the next Charles Schwab.”
The AI revenue ledger starts with a $400B “electronic tax.” America’s total pool is only about $400B—$260B in online advertising, $100B in e-commerce commissions, and $50B in subscriptions. If OpenAI’s $200B of revenue comes entirely from taking existing share, “burning money to build a small Google is not especially meaningful.” The real market is $15T in labor costs, including $300B in customer service alone. Visible AI revenue over the next 12-18 months is about $70B, with the bulk coming from roughly $30B for OpenAI and $15B for Anthropic; all other startups combined may generate only several billion dollars.
The math of concentrated positions is brutal. Of more than 2,000 VC funds tracked, only 80 have generated net returns above 5x, and the largest among them is just a $500M fund. A $20B fund would need $1T in exits to return 5x net to LPs—roughly the combined value of all US IPOs over the past 5-6 years, “simply impossible.” Druckenmiller’s counterintuitive view is that “the best investment is putting all your eggs in one basket.” Public markets offer unlimited capacity without a 10-year lockup; after dilution, Nvidia has actually delivered the highest return in this cycle.
2026 will be a hedge-fund market. The market’s anchor is shifting from Nvidia to OpenAI revenue; 90% of US GDP growth in the first half of the year came from AI investment. Go long AI beneficiaries in traditional industries—Walmart cut stockouts by 30%, while banks have lifted net profit by 15% through AI—and short IT outsourcing and content production. As cloud providers expand from 3 to 10, and Oracle enters with only a 10% margin requirement, margins will inevitably come under pressure. Tesla has only 1 question that matters: can it completely remove the safety driver? In a midterm-election year, “this administration cares deeply about stock-market performance as a KPI; knowing that is enough.”

Deep dive

1. Three 20% Drawdowns in 5 Years: The Market Itself Is No Longer Normal

Freda Duan comes from a Silicon Valley technology fund spanning public and private markets—referred to on the podcast as “Ultimate Capital,” the institution behind the BG2 podcast, with heavy private-market positions in OpenAI, Anthropic and ByteDance, and public-market positions in Meta, Nvidia, Robinhood and Snowflake. She is one of the few Chinese investors at a top-tier US fund and describes it as “one of the few US funds willing to make its research and thinking public.”
Her annual keywords: 2020 was the pandemic, rate cuts and a booming market; 2021 stayed hot, but showed “signs that the tide was going out”—small and mid-caps peaked in February, while Microsoft turned in November; the Nasdaq fell 35% and Meta 70% in 2022; AI took off in 2023 “before anyone had fully reacted”; 2024 was the best year for investing; and 2025 has been about tariffs, US-China relations and the continuation of AI.
A structural observation: since 2000, the Nasdaq has suffered only 6 drawdowns of more than 20%; since 2020, it has already had 3. “The frequency has clearly increased.”
This is Xiaojun’s third episode in the annual series. In the first 2, 朱啸虎 said he could not see a bubble for at least 3 years—“when everyone is talking about a bubble, it definitely hasn’t arrived”—while 戴雨森 called 2026 the “year of R (return)” but explicitly said he would not short.

2. The 2022 Warning Sign: VC Money Eventually Poured Into Nvidia

Freda’s retrospective: during the bleak market of 2022, an extremely abnormal financing pattern emerged. Inflection AI raised $200M in a Series A, while Anthropic, Cohere and Stability each raised several hundred million dollars in the same period. Three things did not make sense: VC money was being used for capex, the average Series A should have been around $10M, and the public-market collapse should have flowed through to VC.
After meeting the founders, she found the logic “extremely simple”: “They just repeated a few words: scaling law, I need to buy GPUs, Nvidia’s cards are simply this expensive.” Xiaojun summarized it as: VC money ultimately flowed into Nvidia. “Yes, you could put it that way.”
The advantage of the crossover structure was clear. At the start of 2023, it was “really hard to tell which” of the 20-30 companies building their own large models would reach the top tier. But a team covering both public and private markets could directly use Nvidia in the public market to express a view on models and AI. Over the past 2 years, most AI exposure has been public: cloud, power, and both beneficiaries and losers.

3. Three Interlocking US Equity Themes, Above Them a Government With Ideas

Freda’s framework for US equities today has 3 main themes: AI, reindustrialization and financial digitization. Reindustrialization includes rare earths, the return of domestic manufacturing and data centers. After the April tariff deal, Japan and South Korea together pledged more than $1T for US infrastructure; Japan has already invested more than $100B in energy projects to “power these data centers.” Any answer to “where does the money for AI come from?” has to include this.
The anchor for the third theme is the Genius Act stablecoin bill passed in July. Once Agent Commerce arrives and agents start making payments, stablecoins will be back in focus—creating another loop back to AI. Above the 3 themes sit “a government with very strong ideas of its own” and next year’s midterm elections.
US investors’ interest in China is real: the visit boom is genuine, and they have been “very impressed” by EVs and robotics. But actual investment amounts are not materially different from prior years. They are watching 2 things most closely: the outcome of US-China negotiations and Hong Kong IPO activity. “If more technology IPOs come later, US investment will flow back.”

4. Betting on OpenAI: A Product Company, Not Just a Model Company

The entry point was when Freda “first realized that OpenAI was a product company, not just a model company”—with roughly 100M-200M monthly active users, while “the market was extremely confused and could not tell which companies were truly in the first tier of models.” The conclusion sounds like consensus today, but in 2023 the market worried that To C products lacked retention, and the positioning of Perplexity and Cloud Rock was unclear. GPT had no Meta-style network effects; it was “one-to-one distribution,” winning retention through first-mover advantage and better user experience.
The analogy is about market structure. Content is non-exclusive, so streaming is naturally fragmented—YouTube, Netflix and Disney each hold single-digit shares. But “with a chatbot, you want one company to remember all your content.” GPT looks more like search, with a highly concentrated structure, where 2 companies capture the market.

5. The Large-Model Ledger: Plus 2 Minus 10 Equals Negative 8

The business model is “simple and brutal”: GPUs are the main cost, and training costs rise roughly 10x a year under the scaling law. If the previous year’s training cost was 1, the second year can recover it with 2 of revenue, but the next generation costs 10, so “you become a negative 8.” Cash burn rises every year. There are logically only 2 ways to turn positive: increase the revenue multiple or stop training a new model 10x larger. Dario then splits the latter into 2 possibilities: physics makes further training impossible or requires burning all the money in the world, or the scaling law slows enough that further spending is no longer worthwhile.
The moment the spending ends is “violent beauty”: add 2 without subtracting 10, and the income statement suddenly looks good. Xiaojun asks whether the disappearance of the scaling law would actually be good for investors. “Yes. At that moment, investors will see the margins they most want to see.”
The Netflix analogy: cash flow was at its worst in 2019, at negative $3B. In 2020, the pandemic stopped production and “there was no training cost,” suddenly turning cash flow positive at $2B. “What you spend money on only produces real cash flow the day you stop spending it. That is the essence of a model company.”
Xiaojun’s pushback is worth preserving: if Netflix stops investing in content, its brand advantage disappears, and Disney cannot stop making content either. Freda concedes that model companies will never stop training completely, “but if training costs stop increasing several times every year, that is already enough for margins.”

6. OpenAI’s Revenue Mix and 2 Non-Consensus Views

Media reports identify 4 lines: ChatGPT, including enterprise GPT, contributes more than 70%; API revenue is currently about half of Anthropic’s, though Anthropic expects to reach 5x or more of OpenAI’s API revenue over the medium to long term; Agent revenue includes the SoftBank partnership; and new products. The company expects the latter 2 lines to grow faster. On a pure GPT comparison with Netflix, 300M subscribers paying $30 a month implies a $100B-scale business; year-end annualized revenue is projected at $20B, versus $13B this year.
The first non-consensus view is that enterprise is dramatically underappreciated. Individual and enterprise users are close to evenly split, with several million enterprise users. Freda is one herself: “I connected GPT to company email, Slack and internal systems, and the collaboration is extremely useful.” Enterprise is the easier route to becoming a super-portal: software pricing is simple, and in the Google-OpenAI competition, office suites can be called directly. “They do not care whether they are at the very top of the customer-acquisition funnel.” Consumer use faces 2 practical obstacles: third-party websites block Agent logins, and merchants resist it—Amazon’s more than $70B in advertising revenue depends on users arriving so it can recommend toothpaste, toothbrushes and floss together.
Advertising is “too logical”: with more than 800M weekly active users and paid users accounting for “roughly 50%,” free users could see a Nike ad when asking about running shoes, while paid users get an absolutely neutral result. “Scaling will be very fast.” Cloud is “something you do almost incidentally.” AWS began by renting out excess capacity from self-built data centers, and “the same thing will happen with this GPU wave; the competitive structure of US cloud will change substantially.”

7. Google: A Perfect Trapezoid, With an Equally Real Threat

OpenAI internally sees Google—not Anthropic or xAI—as its biggest rival. Its position is “a perfect trapezoid”: horizontally, search, YouTube and Workspace; vertically, cloud through proprietary chips. Gemini has 650M monthly active users and GPT more than 1B. Google has 2 ways to wage a price war: TPU vertical integration gives it much higher margins on the same model, while classic bundling turns a $20 YouTube subscription into “give me $10 more and I’ll give you Gemini.” The rational medium- to long-term structure is Google taking the $20 low and mid-market while GPT keeps subscriptions costing hundreds or thousands of dollars.
Freda refuses to take sides: “The market’s negative sentiment toward OpenAI and positive sentiment toward Google are far too one-sided.” Search advertising has not fallen because people mostly use GPT as a giant Wikipedia, asking why the sky is blue—questions Google never monetized anyway. Insurance and law firms are Google’s largest advertisers, and “the question of buying car insurance is too well suited to asking GPT.” Once GPT starts advertising this month, the e-commerce landscape will change. The US advertising pool is finite; if OpenAI sells ads, they must come from Google and Meta. “OpenAI’s competition with Google is real, and the threat to Google is equally real.”
Google’s internal turnaround mechanism was to tie incentives to model leaderboards: whoever moves up the rankings gets promoted and paid more, which is “very compatible with an overachiever culture.” There is also an industry rumor that after Noam Shazeer returned to Google, he fixed a bug by hand and pre-training suddenly started working. Xiaojun responds: “That must have been a bug worth several hundred million dollars.” Gemini 3 also showed that pre-training has not hit a wall.
The search moat has “fundamentally changed.” Google used to win because people could always find the answer in the first 2 or 3 results. Now models are “extremely willing to do the hard work, turning through dozens of pages in 1 second,” so Bing versus a self-built engine “may not be much of a difference.” “This is a very big deal.”

8. Sam’s $1.4T and the Math of Going Public

Freda’s view of Sam is positive: employees trust him and are “very grateful that Sam led them to wealth.” He also deserves credit—“he was genuinely one of the earliest believers in AGI and led the market.” The $1.4T plan was smart in principle: AI burns cash, so bind the entire industry together. But “the timeline was too long and the number was too large; it scared the market. If it had guided only 1 or 2 years of spending, the effect would have been much better.”
Valuation and share price must be separated. The move from $30B to $500B refers to valuation, but per-share value rose only 6x, not nearly 20x; “equity dilution was extremely serious.” A $1T market cap at a 2027 IPO would correspond to roughly 10x projected revenue. “As long as the revenue arrives, the valuation is reasonable.” “There is a whole street in Boston full of large funds. Each puts in $10B, and a $100B IPO appears.” As for “going public means the peak”: “It is not that mechanical.” Netscape went public in 1995, and the bubble burst 5 years later. In the AI wave, capital is more concentrated and “will only be burned at a few large companies.”

9. Anthropic Is Not Spending Less; It Assumes a Higher Return

The businesses differ: more than 80% of Anthropic’s revenue is To B. Both companies show a “smile curve” in retention—retention drops first and then recovers, while new-user retention is higher than older cohorts at the same point. “There is no problem on that front.” Gross margin was “quite a shock”: within 1 year, Anthropic went from “losing 2 for every 1 collected” to near OpenAI’s level through model architecture, auto scaling and optimization. “Long term it will be 70%-80%; there is no reason to worry.”
Growth has passed the baton. OpenAI took less than 2 quarters to go from $100M to $1B annualized revenue; Anthropic took 4-5 quarters. Above $1B, the order reversed: Anthropic took less than 10 months to go from $1B to $7B. In recent months, the 2 companies have added roughly the same amount of annualized revenue each month. The product strategy is to target large, profitable verticals such as coding and finance, and build the best Agent model, with cloud skills as a miniature Agent system.
On the claim that Anthropic has better cash flow and more efficient training: “Cash flow means nothing if you strip away the business model.” Anthropic assumes total compute stops growing after 2028, while model ROI improves every year—$1 invested in year 1 returns $5 in year 2. “It is not spending less; it is assuming a higher rate of return. The fact that the 2 companies make 2 different assumptions is itself worth thinking about.”
Neo Labs are emerging: nearly 10 new model companies have appeared in recent months, with academic backgrounds and more than $1B in financing. Freda is candidly uncertain: “I have not fully thought it through either. I do not know whether there will be a point when everyone believes large models have converged around 1 or 2 leaders, and the rest stop training and embrace open source.” Ilya’s SSI is pursuing continuous learning; Mira’s Thinking Machines just released Thinker to let users customize training for open-source models. Investing in it means betting that the open-source market will keep growing. In valuation, every model company is benchmarked against “what percentage of OpenAI it is worth.” “This company is the most important benchmark in both public and private markets.”

10. Robinhood: A Bad Hand Played Into the S&P 500’s Top Performer

Freda’s core judgment is counterintuitive: this year’s top S&P 500 stock is “fundamentally a poor business model, essentially a bad hand.” Securities trading is highly cyclical: “In a bull market, everyone operates like a tiger; in a bear market, it lies flat.” The company can do 4 things: diversify into banking, prediction markets, wealth management and international markets, with more than 11 business lines and over $100M in revenue; gain share to smooth the cycle; exercise pricing power—crypto commissions rose from 10 basis points to 60 basis points in 3 years, allowing it to “control its own destiny”; and keep costs in line—operating costs have been a flat line since 2022. Meta in 2022 was the counterexample: revenue fell, costs surged and cash flow was cut in half.
This is not a “buy brokers in a bull market” thesis. “A key part of investing is being honest with yourself and knowing what money you are making—is it beta or alpha?” A regression of Coinbase’s share price against Bitcoin shows no alpha at all since 2022; if you are calling the crypto bull market, you are better off buying Bitcoin directly. Robinhood, by contrast, has “very strong alpha” against both Bitcoin and the Nasdaq and “has taken a different path.”

11. The Next Charles Schwab: The 34-Year-Old Account and the Wealth Divide

The growth logic is that “each generation has its own account.” Baby boomers used Schwab, people aged 45 used eTrade, and Robinhood users average 34 years old. Average assets per account are now $10,000, versus more than $150,000 at Schwab and IBKR. Age 35 is a wealth divide in America: average wealth rises roughly 3x from 35 to 40 and peaks at 55. “Even if Robinhood does nothing, its assets will grow naturally.” More than 60% of Schwab’s revenue comes from wealth management. The direction is clear: “Broadly speaking, Robinhood is the next Charles Schwab.”
The trade was built during a difficult market at the start of the year. Coinbase, valued above $100B, was “losing share and losing price,” while Robinhood, at $30B, was executing well and taking share everywhere; the market treated both as beta. “Under enormous pressure, thinking the year might be a bear market, I still bought it.” After nearly tripling this year, “it is no longer cheap,” but it has “the best chance to become a one-stop financial app.”
The upside is substantial. US alternative assets exceed $10T, yet ordinary people cannot buy into OpenAI or SpaceX. If Robinhood created a VC fund filled with star companies, charging 2%-3% like Cathie Wood’s ARK, “it would not be surprising if it reached tens of billions.” In Europe, tokenization could open markets in a single rollout, avoiding the 1-2 years required to obtain licenses country by country.

12. Bad Boys, the Single-GM Model and Retail Tribes

The founders are opposites. Coinbase’s founder is a “good kid” who emphasized compliance from day 1; “to this day, operating costs cannot be cut, and compliance is a large part of them.” Vlad is the “bad kid”—willing to imagine and act boldly, compelling in public communication and popular with retail. “He will sincerely ask me which companies he needs to learn from, especially those larger than him,” while remaining highly aggressive: “he will pursue any market he wants with extreme aggression.” Silicon Valley capital still prefers founder-led bad boys who dare to break convention.
Product speed rests on an organizational redesign. In 2022, Robinhood “essentially rebuilt the company”—moving from a centralized structure to a single-GM model, with each business unit having its own PMs, developers, CFO and HR. “The annual product output can be what others produce in 5-10 years; that is not an exaggeration.”
Retail investors are now an important part of the market. In some options markets, 80%-90% of activity is retail. Twitter KOLs form tribes—the “Robinhood guy,” the “Palantir guy.” On these 2 stocks, “retail research is overwhelmingly better than institutional research.” The stock-picking approach is to find names retail likes and institutions can also own: strong momentum and less zero-sum trading. Software is the counterexample—“almost no retail, just hedge funds trading against each other; a 30% move up or down is pure pain.”

13. The Common Traits of Disruptors and the Prediction Market With a Royal Flush

Disruptive companies share 3 traits: enormous TAM—Robinhood wants to become the super-app for all of finance; a product so good it needs no sales force—Google, Netflix and GPT; and a sense of inevitability created by timing, place and people. Smartphones and bandwidth enabled short video and TikTok, the end of Moore’s Law enabled GPUs and AI, and the blank space in lower-tier markets created Pinduoduo. “Each one carries a little sense of destiny.”
The most interesting new species this year is the prediction market—Kalshi and Polymarket, “not investment advice.” Kalshi’s latest trading volume exceeded $60B, up more than 100% quarter over quarter, a classic example of explosive growth in a new category. For $1 or $2, users can bet on everything: elections, whether Tesla earnings beat expectations, or whether Taylor Swift will become pregnant. On Polymarket, someone early on bet that Gemini 3 would arrive on November 18, accurate to the day. Bloomberg and Google Finance have already integrated its data.
The timing is favorable: US mainstream media is suffering a severe credibility crisis, the world is unstable, and “people desperately want to know what is actually true, while there always seems to be someone who knows something.” Prediction markets are legal across the US, while California and Texas do not even allow sports betting, and this administration is easing regulation. “They have been dealt a royal flush. The question is how big they can become.”

14. Waymo Is Already Profitable in San Francisco; Tesla Only Needs to Remove the Safety Driver

Waymo was the biggest surprise of the year. City launches are accelerating: roughly 5 years in San Francisco, under 2 years in Austin and 6 months in Silicon Valley. A new city costs “around tens of millions of dollars.” There are 2,500 vehicles on the road and annualized revenue “may be just under $800M.” Hyundai’s existing factory in Georgia can produce at a 100,000-vehicle scale, so “volume will ramp very, very quickly in the next 2 years.” San Francisco is already profitable: using a $170,000 vehicle depreciated over 4 years, and including remote safety-operator costs and insurance, “even today it is profitable.”
On the operating model, Waymo works with everyone—Uber, rental companies and fleet managers—but fundamentally needs cleaning, maintenance, charging and parking. “Avis or a professional fleet operator is the best fit.” Uber is not a labor-intensive company. Waymo must solve peak-to-trough demand balancing; Uber can adjust prices and fleets as an asset-light platform, while Waymo’s vehicles sitting idle during off-peak periods would weigh on profits.
The 2 companies are mirror images: “One has a hardware problem; the other has a software problem.” Waymo’s hardware costs—$70,000-$80,000 each for the vehicle and sensors, plus $20,000-$30,000 for chips—cannot be cut to Tesla’s $30,000 price. “Waymo’s long-term fate depends heavily on whether Tesla succeeds.” Tesla can mass-produce easily, but Elon’s pure-vision thesis from 10 years ago “has still not truly proven itself in Robotaxi.”
The market math is large. Americans drive about $3T miles per year; at $1 per mile, that is a $300B annual market. Uber charges roughly $3 per mile, and “if it is safe, there is no demand problem.” The top 10 cities need only 10% share and 10,000-20,000 Robotaxis; Uber has more than 3M vehicles today. Technically, autonomous driving still relies on imitation learning and roughly billion-parameter edge models. Put a stop sign on a person and some models will stop too: “They memorize; there is a lot of overfitting.” Every other vertical has shown general models outperforming specialized small models. “I would be curious to see how large-model companies approach autonomous driving, and whether latency can be solved.”

15. Robotics: VCs Are Investing With Their Eyes Closed, but Autonomous Driving Comes First

The blunt reality: “Many VCs are honestly investing with their eyes closed.” Locomotion, the lower half, is largely solved; manipulation, the upper half, “still has a long way to go.” There is no consensus on the path: humanoid versus quadruped, real data versus simulation versus a hybrid. Areas of convergence include joint hardware-software training, residential rather than industrial deployment and more first-principles support for humanoids—“of course, largely because 老黄 and presumably Elon are leading the charge.” But it is still early, just moving from research into engineering.
The analogy is autonomous driving 10 years ago: “When a model is extremely data-starved, any data you feed it looks highly helpful, but you cannot extrapolate linearly. Progress will be slower than people expect.” Investors face a further problem: unlike large language models, robotics has no leaderboard, and task completion remains low. “Apart from saying the market is genuinely huge, it is hard to make a large bet.”
The optimistic case is a robotics GPT moment within 2-3 years: manipulation achieves general actions such as folding clothes, putting things away and loading a dishwasher. “The completion rate may not be high, but this is where it starts.” One X and several others will sell products into customers’ homes next year. But Freda keeps expectations low: “The market has to solve autonomous driving first, then robotics.” Robotics has not even settled on a form factor, and the data gap is too large. China’s autonomous driving will develop well because its EV cycle produced camera and connectivity data, while US gasoline cars are not connected and have at most 2 cameras—“even if you collect it, you cannot use it.” In robotics hardware, “China is definitely the best in the world.”

16. Consensus and Non-Consensus, 3 Types of Founders and the Brutal Math of Concentration

The fundamental public-private distinction is that “the best public-market investments are non-consensus—Meta in 2022, Google this year—while VC is consensus investing.” Each round requires a great deal of capital and market acceptance to support the company for years. Experience in judging people helps: excellent founders rarely truly fail even after several pivots. When Ilya raised money recently, he reportedly had only 1 sheet of paper: “I have raised more than 100 rounds and never lost an investor money.” That is his confidence.
Funds tend to prefer 3 founder archetypes. Auto Venture, a major Roblox shareholder, likes the “hedgehog”—digging 1 hole for life and emulating Buffett, rather than the “serial entrepreneur” constantly chasing new trends. Ruby Capital likes the Rebel—a bad-boy temperament and a desire to fight the world; Robinhood is the classic example. Hummingbird likes the Nezha type—psychological scars, “my fate is mine to control,” and a refusal to sell the company for liquidity, as with Craftken and several game companies. “Overall, the differences in judging people are not that large; you can recognize excellent people.”
Druckenmiller’s counterintuitive point is that “the best investment is putting all your eggs in one basket.” The willingness to use 1 basket shows the research is complete; “the risk is actually much lower than making half-informed random bets.” Ribbit put $500M into Robinhood within 24 hours of its 2021 crisis and made more than 7x. Buying Carvana at $10 may have produced more than 30x. Greenoaks limits partners to no more than 5 meetings a week, encouraging them to make fewer judgments and place larger bets.
The VC return math is brutal. Of more than 2,000 VC funds tracked, only 200 have generated more than 3x and 80 more than 5x; among those 80, the largest fund is only $500M. A $20B fund seeking a 5x net return for LPs must return more than $100B; at 10% ownership, that requires $1T of total exits—“equivalent to the combined market value of all US IPOs over the past 5-6 years. It is simply impossible.” VC therefore needs to stay small and excellent. Public markets offer unlimited capacity without locking capital for 10 years. In 2023, instead of deciding which of dozens of model companies would win, an investor could simply buy Nvidia. “After accounting for all equity dilution, Nvidia’s return was the highest of the lot.” Freda’s fund has fewer than 30 people, with the same team covering public and private markets, creating “a lot of synergies.”

17. Sector Scan: Coding’s Iron Law, Video’s New Logic and the Nightmare of Agent Commerce

An “open secret” in investing: “If a sector can create $1B of annualized revenue within 1 year, whether or not you understood it before, you must invest—this is the market telling you a world-class company will emerge here.” Coding generated more than $5B of annualized revenue in 1 year, with 4 companies above $600M. Claude Code and Codex “rose in a straight line and came from behind” after launch. But the competition has barely begun: Google and Winserve have made coding free. “What does that mean for Cursor?” Coding is also the easiest area for model companies to do themselves: no domain knowledge needs to be collected and no sales force is needed for GTM. “That is very different from finance and law.”
Startups are now left with verticals. In the last software cycle, companies started at the general layer and moved into verticals; the general layer has now been eaten by large models. Software companies are “like chicken breast—very consistent in flavor,” with 80% gross margins and net retention above 120%. AI application margins are counterintuitive: “the more people use it, the lower the gross margin becomes.” Valuation should focus on absolute profit: contracts are larger and gross profit per user is several times traditional software. AWS was dismissed for 50%-60% margins when it launched, but ultimately became bigger than all software.
Freda is “very bullish” on video. The US media industry is worth more than $800B, and “video is the highest-bandwidth input format for the brain.” The logic has fundamentally changed. TikTok optimizes matching—the matching function itself. “For the first time, video has become an object that can be directly optimized; user engagement or ad click-through can be used as an objective function and aggressively optimized. There is so much that can be done.” Companies worth tens of billions will emerge; Google, Meta and Elon are all working heavily on it.
Agent Commerce, OpenAI’s next direction and something “we may see this month,” will be “a nightmare for most merchants.” Half of Booking.com’s transaction volume comes through Google, and many people do not know that this part is effectively zero-margin. Amazon’s nearly $100B in advertising is “eyeball revenue”—if users do not arrive, it cannot collect. The only clear beneficiaries are long-tail Shopify merchants, whose niche products an Agent can find after searching dozens of pages. Payments are even more disruptive: “The Agent does not mind the hassle; GPT is my electronic wallet. Do we still need PayPal as an intermediary? That is a serious question.”

18. Bubble Accounting: A $400B Electronic Tax or $15T in Labor

Xiaojun’s key question is whether the enormous AI revenue pool is existing spend or incremental spend. Freda’s “electronic tax” framework puts total US digital revenue at roughly $400B: $260B in online advertising, about $100B in e-commerce commissions and roughly $50B in subscriptions, including Netflix and YouTube. If OpenAI reaches $200B of revenue entirely by taking existing share, Google, Meta and Amazon would all have to give up share. “If all that spending merely burns money to build a small Google, the significance really is not that great.” She also rejects the idea that AI will expand the advertising pool: US advertising has grown only 5%-6% annually for 20 years, and companies spend about 3% of revenue on advertising. “AI gives no reason to exceed 3%; I would be happy if it does not fall.”
The real market is much larger: US GDP is $30T and labor costs are $15T, with customer service alone worth $300B. Michael Dell and Jensen’s logic is that a 10% increase in global productivity from AI would create more than $10T in additional GDP, making massive annual compute investment rational. Freda’s path is that models first take electronic revenue while capability is limited and the advertising landscape deteriorates. Given time, genuine Agents and AI for Science emerge, “and AI can enter the $15T labor market. Then the math works.”
Xiaojun pushes back socially: if major productivity gains cause layoffs, who consumes? Freda’s example is 100 people producing 100 units becoming 80 people producing 110 units through AI—cheaper and better, with real GDP up 10%, while the remaining 80 workers’ income and company profits both rise, creating a positive loop. “The 20 people laid off will indeed suffer for a period.” A California legislator—referred to verbally as Rukana, likely Ro Khanna—is already proposing an AI tax or layoff tax. “This is not alarmism.” Freda remains optimistic: every productivity wave in history brought layoffs, but society absorbed the shock through reemployment and redistribution.
Visible AI revenue over the next 12-18 months: roughly $30B for OpenAI and $15B for Anthropic; 2 verticals near $2B each, coding and video generation; several $500M-scale verticals, including audio, customer service, legal and healthcare; plus AI revenue at listed companies, for a total of about $70B. “The most important point is that the bulk of trackable AI revenue comes from OpenAI and Anthropic. All other startups combined generate at most several billion.” The bubble question has 2 layers: today is definitely not a bubble—GPT has reached the adoption equivalent of a decade of internet progress in under 3 years, while large companies’ ROIC is improving quarter by quarter. Whether it later becomes a bubble depends on whether models keep improving—Gemini 3 showed pre-training has not hit a wall, and a Blackwell-trained model expected in Q2 next year will be worth watching—and whether AI revenue keeps rising.

19. 2026 Outlook: A Hedge-Fund Market Anchored to OpenAI Revenue

Running through the large names, with no investment advice: Google is the consensus leader, with SOTA models and the best positioning, but next year the focus will be intensifying advertising competition—OpenAI scaling ads and “TikTok, which everyone is close to forgetting,” with US users roughly equal to Instagram, higher time spent and less than one-third the revenue. Meta and Tesla have similar setups: “fundamentals are not especially good; what moves the stock is AI progress.” Meta has fallen only twice in 15 years as a public company, and both years had negative cash-flow growth; “cash flow will definitely fall next year.” For Tesla, only 1 thing matters: whether it can completely remove the safety driver, plus the structure and dilution of its xAI investment. Apple is “very puzzling”: the stock performs well without AI, and its culture of waiting for a product to be absolutely perfect before launching is unique. Tim Cook should leave next year; “recently he is often seen buying coffee, looking extremely relaxed.”
She is not particularly fond of the cloud landscape. GCP may grow more than 50% quarter over quarter, but the market is expanding from 3 clouds with 35%-40% margins to 10 clouds. Oracle only needs a 10% margin to enter, CoreWeave and other neo-clouds keep appearing, and model companies are entering themselves. GPU rental is too commoditized; “in the long run, every company will have to develop its own chips.” Customer concentration is also extreme: Azure’s incremental growth is 70% from OpenAI. Microsoft’s CEO has said profits will not come from renting GPUs, but from databases, storage, security and other add-ons. For Nvidia, watch whether proprietary chips can deliver and whether TPU capacity can be rented out. “In a world with severe power shortages, the price gap between GPUs and proprietary chips becomes less important.”
The long-short framework is clear: next year’s economy will diverge sharply, making it “a classic hedge-fund market.” Go long AI beneficiaries and short AI losers, including IT outsourcing and content production. Excess returns will come from AI applications in traditional industries. In Q3, S&P 500 companies for the first time gave specific data on AI productivity gains—it took 5 years for the cloud wave to begin migrating workloads. Walmart used AI for supply-chain forecasting, cutting stockouts by 30% and saving 30M truck miles a year. Financial companies cut processing time by more than 80% and halved default rates; one bank reported a 15% increase in net profit from AI. The world’s largest logistics company, CH Robinson, has AI handling half of bookings and has lifted productivity 35% in 2 years.
“Today, the US equity market is the AI market.” AI investment accounted for 90% of US GDP growth in the first half of the year. The market’s bellwether is gradually shifting away from Nvidia: “the market cares more about AI revenue than AI investment.” The largest and most measurable number is OpenAI revenue—watch user growth, revenue growth and whether the US government participates in its infrastructure. Consumption is healthy, the deficit has improved, and Q1 still brings stimulus from tax refunds and tax-free tips. Most important is the midterm-election year: “This administration cares deeply about stock-market performance as a KPI; knowing that is enough.”

145: 孙宽 of Hypershell: The Birth of the First “Consumer Exoskeleton”Mustafa Suleyman: The AGI Race Is Fake, Building Safe Superintelligence & the Agentic Economy | #216