Vol. 76 How We Somehow Ended Up Here — A 170-Slide PPT Explaining the 2025 AI Industry
Summary
- 庄明浩的核心判断不是“AI 有没有泡沫”，而是行业已经把技术、资本与基础设施同时推到了历史级极限。 The market has moved from “is this a bubble?” to “what kind of bubble is it?” For now, it looks more like a productive, equity-driven bubble, but Oracle, Meta and others are starting to lever up, shifting risk toward debt. “We somehow ended up here” is no longer just rhetoric.
- OpenAI 是整个资本循环的发动机，也是最难估值的风险节点。 It is now valued at roughly $500B, with 800M weekly active users and projected 2025 revenue of $12B, yet it may not turn cash-flow positive until 2029 or 2030 and still expects to burn $115B. Sam Altman was already saying in 2024 that AI would need $7T; OpenAI has now committed to more than 30GW of data centers and roughly $1.4T in spending. “Whoever gets the five-year expectations takes off.”
- NVIDIA 已不只是卖芯片，而是在充当“AI行业的央行”。 It has surpassed a $5T market cap, generated roughly $26.4B in quarterly profit and holds about $100B in cash, while using $100B to bind itself to OpenAI and investing in Intel, Nokia and Neocloud. The loop in which OpenAI buys compute and NVIDIA reinvests in OpenAI is now in place. 庄明浩’s view is that Jensen Huang is using capital allocation to manage the entire industry.
- 技术主线仍是从推理模型走向Agent，但真正的瓶颈已经从模型能力扩展转向数据获取、奖励、记忆和评估。 Math and coding are advancing fastest because their answers are verifiable, while creative tasks lack clear reward signals; reward hacking, process rewards, long- and short-term memory conflicts and obsolete benchmarks remain unresolved. Agents may therefore still need 5–10 years. “This year is only year one of a decade-long R&D cycle.”
- 产品端的赢家尚未锁定，第一方模型厂商正在吞回应用层价值。 ChatGPT is going all in on an All in One product with 800M weekly active users, while Gemini continues to catch up at 60%–70% weekly growth and, in some weeks, 95%. AI coding once set the fastest ARR records; now standalone tools are broadly seeing double-digit traffic declines, while Claude Code and Codex are growing rapidly. More broadly, 60% of AI Apps and 75% of Web applications declined in Q3. “User value and active scale” are replacing model narratives.
- 中国模型的技术存在感与资本定价出现了巨大背离。 DeepSeek, Qwen, MiniMax, Kimi and GLM continue to push open source; in developer surveys, DeepSeek rose from 0 to 53% and Qwen from 0 to 25%, while Airbnb publicly said it chose Qwen after comparing it with GPT. Yet leading Chinese model companies are valued at only a few billion dollars, more than 100x below OpenAI’s $500B. “This isn’t dividing by 3, 7 or 10. It’s dividing by 100.”
- 二级市场的AI交易正从MAG7向整条数据中心供应链扩散。 In 庄明浩’s 2025 grouping, the MAG7 rose roughly 16%, while energy, chips and software gained 53%, 41% and 33%, respectively. Even storage, which accounts for only about 1% of data-center costs, is now in shortage, making Seagate and Western Digital two of the S&P 500’s strongest performers. Amazon, Microsoft, Meta and Google are each expected to spend more than $100B in CapEx in 2026, while Neocloud is expected to grow 69% annually over the next five years.
- 对投资者最重要的不是抢着证明泡沫，而是同时防止“过早下车”和低估尾部风险。 Leading AI companies currently trade at an average P/E of roughly 28x, well below the internet-bubble peak of about 89x, and Big Tech cash flows can still cover most investment. But GPUs may have only about a three-year economic depreciation life, while Oracle’s CapEx has reached 135% of operating cash flow. Once shadow banking and debt take over, the risk structure changes. “For many people, a bubble is simply a bull market they did not participate in.”
Deep dive
1. A Seven-Month Summer and Trillion-Dollar AI Are the Same Kind of Extremes
庄明浩 opens with the story of a Hangzhou women’s-clothing retailer who said, “I sold short-sleeved shirts for seven months this year.” Shanghai was still warm enough for T-shirts in early November; 影视飓风 sold 200,000 basic T-shirts, and Tmall’s fashion team even went to Lingyin Temple to pray for cooler weather so winter clothing could sell during Singles’ Day.
He connects this absurdity to VCs “holding rituals to pray for rain”: in conditions of extreme uncertainty, people repeatedly perform rituals, hoping one extreme will resolve another. Meanwhile, tech companies are making trillion-dollar bets on a future nobody understands. “Two extremes, the same absurdity.”
That is why this 170-slide deck is called How We Somehow Ended Up Here, with the even more direct subtitle: “How the fuck did things end up like this?” The question is not merely whether AI is being overbuilt, but how humanity moved from seeking moderation and balance into an “inevitable age of extremes.”
2. The Industry Is Moving Faster Than the Deck, but Parts of the Old Framework Still Hold
From the 139-page 2024 year-end review, to the 74-page DeepSeek breakout deck, the 70-page Mistral deck, the 58-page Agent deck and September’s 154-page Reasoning Wants to Fight Monsters, Scenarios Want to Open the Map, Investors Want to Spend, 庄明浩 finds that simply reading the past tables of contents reveals how the industry’s attention has accumulated over time.
This time, he compresses the still-valid material into a “review”: OpenAI is still trying to solve the problem of facing simultaneous pressure from investors, Microsoft, NVIDIA, governments, its own team and competitors. L1 Chatbot, L2 reasoning and L3 Agent remain the mainline, but they are advancing in parallel rather than replacing one another linearly.
The money narrative has not changed either: raise more capital, build more and better data centers, train stronger models, and produce better results. The question is whether this narrative, defined largely by the United States, can be adopted by China—or whether China needs to develop a different capital and open-source structure.
The most painful part is that “the industry is developing faster than you can make the PPT”: in the final week of October alone, OpenAI and Microsoft reset their partnership, NVIDIA held its Washington GTC, and Apple, Microsoft, Google, Amazon and Meta all released highly information-dense earnings reports.
3. Reinforcement Learning Raises the Ceiling for Reasoning—and Exposes the Data Ceiling
In 2025, reasoning models remain the main highway for natural-language models, a line that can be traced back to OpenAI o1 in September 2024 and was reinforced by DeepSeek R1 early this year. Pretraining produces L1; post-training and reinforcement learning drive L2 and even L3. The “technology plus ruthlessness” route is now fairly clear.
Math and coding are progressing fastest not because models naturally prefer them, but because their answers are easy to verify: correct answers receive rewards, wrong answers receive penalties. Creative, preference-based and ambiguous tasks lack equally clear verifiers, so “how do we define useful?” becomes a training problem in its own right.
Rewarding outcomes alone can also induce reward hacking: like people, models may stop at nothing to hit a metric. As the marginal benefit of outcome rewards declines, process rewards become important again. Ultimately, all of these problems return to the data-acquisition leg of the compute, algorithm and data triangle.
4. Memory and Benchmarks Are Losing Their Old Measuring Sticks
Longer context windows are only the starting point for memory. Models still need to balance long- and short-term memory, handle semantic drift, memory conflicts and control over specific memories, and improve retrieval efficiency. A model’s own memory also cannot simply be conflated with the external memory required to complete a task.
Old benchmarks can barely measure the strongest models anymore: test sets may leak, models can train specifically against them, and models iterate monthly while evaluations update only every six months. Optimizing for a high score is also different from optimizing to be useful in real-world settings. A benchmark that remains valid over time and updates dynamically is therefore “very, very difficult.”
Go player 柯洁 offered the sharpest analogy. Asked whether he worried about AI learning from professional players’ data, he replied: “Our level is so low—why would AI need our data? Wouldn’t that just contaminate its database?” AI Go may not be moving from 90 to 100; it may already be at “100 million,” leaving humans without even a coordinate system.
5. America’s Main Model Table Is Down to Four, and Meta Has Fallen Behind
庄明浩 reduces the U.S. closed-source model table to OpenAI GPT, Anthropic Claude, Google Gemini and xAI Grok. Meta’s Llama, he says, “quietly fell behind” in 2025: its reputation, update cadence and leaderboard position no longer put it alongside the top four.
Once the technical route and objective are set, the competition naturally becomes an all-out arms race. Every company calls its release the most advanced; the leaders keep changing, while Chatbot, reasoning and Agent capabilities are being contested simultaneously. No single leaderboard can provide a stable answer.
Developer surveys are not absolute data, but they show the direction: Gemini usage rose from 31% to 80%, Anthropic from 46% to 67% and xAI from 0 to 31%; Llama fell from 49% to 43%, while Europe’s Mistral declined from 37% to 22%.
Google deserves a table of its own: it has TPUs, cloud infrastructure and Anthropic as a critical partner. Faced with the OpenAI–NVIDIA–Oracle loop, Google’s position is essentially: “Why should I play with you? I can just set up my own table.”
6. Chinese Open-Source Models Have Reached the Table, but Valuations Are More Than 100x Lower
The “Six Little Dragons” narrative has run its course, but DeepSeek, Alibaba Qwen, MiniMax, Kimi and GLM are still developing rapidly, while Tencent and ByteDance are also in the game. 庄明浩 highlights MiniMax’s newly released model that week, which received strong reviews in both China and overseas.
Open source has become a kind of globally correct narrative advanced by Chinese companies, and it has been written into policy opinions related to “AI Plus.” Airbnb’s CEO publicly said the company compared Qwen with the latest GPT and chose Qwen; in surveys, DeepSeek rose from 0 to 53% and Qwen from 0 to 25%.
Capital markets have not applied the mobile-internet-era conversion rule of dividing U.S. valuations by 3, 7 or 10. OpenAI is worth roughly $500B, while leading Chinese model companies are worth only a few billion dollars. Even Mistral, Europe’s lagging sole survivor, has secured investment from ASML on scarcity value and is valued at roughly $14B.
7. DeepSeek V3.2 Matters Most Because It Deferred Everyone’s Expectations
Model cost per token is falling exponentially; 庄明浩 cites the rule of thumb that costs decline roughly 10x each year. But the leading model’s list price does not fall, because the number one model can always command a premium. When an old champion is displaced, its price drops; the new number one remains in the same price band.
This also explains why falling model costs have not translated into lower gross margins for many model companies: users and applications keep wanting to call “the best model,” and the model in first place has not cut its price.
DeepSeek R1 was built on the V3 base model, so the market had expected a cross-generation V4 to give rise to R2. Instead, after V3.1-0528 came V3.1 Ultimate on September 22 and V3.2 on September 29—not V4. “In the end, I was the one who got hurt.” Whether R2 will be built directly on V3.2 remains an open question.
8. Multimodality Is Entering the Production Pipeline, and World Models Have Reawakened the Metaverse
Images and video have moved from isolated battlefields into media production pipelines. Google’s Veo 3 remains ahead, while Kuaishou, ByteDance and Chinese startups are moving quickly. The question is no longer just whether a model can generate a single image, but whether it can fit into a continuous, controllable content-production workflow.
World models extend the front line to Google DeepMind’s Genie 3, 李飞飞’s World Labs and Hunyuan World Model 1.1. 庄明浩 cites the idea that world models could become “the womb that gives birth to AGI.” If world models are eventually perfected, the metaverse narrative may become viable again.
DeepSeek and GLM releasing OCR models at the same time points to another data route: pure text tokens may be approaching some kind of limit, while humans learn text, images and spatial relationships through vision. Encoding information as images and then recognizing and compressing it might break through text-length constraints, but the results still need to be validated by model companies.
9. Agent Is Not a One-Year Product but Year One of a Decade-Long Systems Project
“2025 is the first year of Agents” has been repeated by Sam Altman, Ilya and countless reports, but 庄明浩 has long believed that a genuinely useful general-purpose Agent may still require 5–10 years. A core Kimi researcher likewise judged that an Agent capable of replacing 80% of white-collar work would take a long time, absent a completely disruptive technology.
Karpathy’s analogy comes from autonomous driving: when he saw a demo in 2012, he also thought the technology was ready. Twelve years later, it still has not reached the ideal state. So 2025 is not the year Agents mature; it is year one of a decade-long cycle of continuous R&D and iteration.
Today’s Agents resemble a group of players kicking frantically around the goal but never scoring. The problem is not just the model: the web, accounts, permissions, databases, security, user systems, memory, execution costs, wait times and outcome evaluation can each become the weak link that drains the entire barrel.
Deep Research, Computer Use and Agent Model describe both technologies and products, and the boundaries remain confused. The spotlight has dimmed, but Agents and their infrastructure remain core financing categories. Once companies actually start building, the industry talks less about definitions and runs into engineering reality more often.
10. OpenAI Is No Longer a Company but at Least Three Pillars, with More Businesses Potentially to Come
In 2025, OpenAI released GPT-4.1, GPT-5, DALL·E 3 and Sora 2, while also launching Codex, the Sora App, a browser and in-ChatGPT applications. It can no longer be defined merely as a lab, nor summarized simply as ChatGPT.
Organizationally, technical R&D has been separated from product and application operations, and OpenAI brought in a female former Facebook mobile VP born in 1986 to run the product and applications teams. Of roughly 3,000 employees, 600 reportedly came from Facebook. That has improved product delivery while creating cultural and integration challenges.
In a livestream on October 29, Sam divided the architecture into three pillars: research, product and infrastructure. Underneath are chips, racks, data centers and energy. 庄明浩 therefore believes OpenAI may include other businesses, including hardware; the Microsoft partnership specifically excludes hardware projects.
11. ChatGPT Wants to Be All in One
ChatGPT can already call Booking to reserve hotels, Canva to create designs, Figma to build UI, Spotify to play music and Zillow to search property. What OpenAI wants is an All in One product like WeChat, Alipay and Douyin.
It can attempt this first because weekly active users rose from 400M at the start of the year to 800M. It was still at 700M in August, then added another 100M in just over two months. More importantly, retention has formed a “smile curve”: the longer users stay, the more likely they are to remain, creating self-reinforcing platform scale.
If a user report allocates demand among search, recommendations, copywriting, images, data analysis and coding, the threat looks limited. Put the corresponding Google, Excel, Canva, Photoshop and translation tools back into the picture, however, and ChatGPT’s potential target is an entire row of established software gateways.
12. Chatbot Has Not Been Decided, but AI Coding Has Entered the First-Party Harvest
ChatGPT’s share remains high and may still be growing 3%–5% weekly, but Gemini, starting from a smaller base, once grew 60%–70% week over week for several consecutive weeks, with some weeks reaching 95%. If that pace continues, the gap will close quickly. The Chatbot battle may not be over.
AI coding has penetrated architecture, coding, testing, version releases, databases, operations and QA, with an extremely granular division of labor across the stack. The fastest companies to reach $100M and $500M in ARR were almost all coding companies. “Today it’s six months; tomorrow I’ll do it in three.”
Starting in Q3, however, traffic to standalone coding products broadly declined. Even Cursor was not spared, with some weeks down 15%–30%. The issue is not necessarily disappearing demand; traffic may instead be shifting to products built by the model companies themselves.
Anthropic’s Claude Code and OpenAI’s Codex are growing extremely quickly, and Claude Code now holds the latest record for reaching $500M in ARR. Third parties can still win, but when model companies are valued at tens or hundreds of billions of dollars, they will not readily surrender the clearest monetization opportunities.
13. Browsers, Search, Chatbots and Agents Have Become One Chaotic Battle for the Gateway
Chinese companies are especially active in general-purpose Agents, with Manus and Genspark as representative examples. 庄明浩 suspects that “general AI” itself is a more Chinese narrative: the United States seems to assume AI will exist, but places less emphasis on the phrase “general AI.”
Browsers are natural Agent environments. The boundaries among OpenAI’s and Perplexity’s browsers, Chrome, Quark, Felo, Doubao, ChatGPT and ima are becoming increasingly thin. Quark retains a search box but can also switch into an assistant mode resembling Doubao.
The Manus route runs a sandboxed browser inside an Agent; OpenAI runs an Agent inside a browser. “There is no right or wrong answer between running an Agent in a browser and letting AI run the browser.” The real prize is control of users’ tasks, accounts and web execution.
14. Total AI Applications Are Still Growing, but Most Individual Products Have Already Declined
A Q3 report showed overseas AI applications growing roughly 19% quarter over quarter and China’s growing about 7%. Its conclusion was that “the frenzy and imagination surrounding models are gradually receding, and the market is beginning to use the most basic yardsticks—user value and active scale—to measure the true winners.”
The underlying data are colder: roughly 60% of native Apps declined, while 75% of PC Web applications fell. Plugins that were still growing 75% in the first half were growing only 48% by Q3. In other words, more than half of the AI applications in the market are declining.
Average token consumption per user per session is also falling. That could mean stronger models are producing answers in fewer rounds, or weaker user stickiness; the program does not force a single explanation. Time spent is rising, but leading Chinese AI applications still typically hold users for only a few to a dozen minutes, not yet enough to support a classic mobile-internet platform opportunity.
15. ARR Is “Neither Annual, nor Recurring, nor Even Revenue”
ARR is supposed to mean annual recurring revenue: subscription and add-on revenue after churn. But many AI products have been live for less than a year and simply multiply one month’s revenue by 12. Reported figures may also include promotions, campaigns and upgrades, while it is unclear whether churn has been deducted.
Borrowing Voltaire’s description of the Holy Roman Empire, 庄明浩 says today’s ARR is “neither annual, nor recurring, nor even revenue.” Even so, $100M ARR has become the threshold for a star AI company: reaching it in roughly 18 months makes a company a superstar; taking four years makes it an ordinary star.
a16z therefore partnered with a financial platform covering payments for tens of thousands of startups and ranked companies by the actual corporate cash being spent on AI. OpenAI ranked first, Anthropic second and Replit third, followed by ElevenLabs, Cursor, Notion, Perplexity, Canva, Lovable, Gamma and Midjourney.
Only 12 of the top 50 companies also appeared in the traffic rankings. The other 38 were not well known but could still collect cash. General-purpose tools accounted for roughly 60%, while vertical applications in customer service, sales and HR accounted for about 40%. Traffic, ARR narratives and actual enterprise spending are three coordinate systems that do not overlap.
16. The U.S. Bull Market Is Highly Concentrated but Has Started Looking for Narratives Beyond the MAG7
The MAG7’s market-cap trajectory almost tracks the S&P 500 because the seven companies now account for close to 30%–40% of the index. Over the roughly three years since ChatGPT launched, 庄明浩 cites gains of 165% for AI companies, 24% for non-AI companies and 68% for the S&P 500.
Of America’s 10 largest companies by market cap, all but Walmart and Berkshire Hathaway are considered AI plays. Their combined market value equals 77% of U.S. GDP, versus roughly 34% for the top 10 during the internet bubble. The concentration is in an entirely different league.
NVIDIA took 180 days to move from its first $1T to $2T market cap, then 66 days to reach $3T, 273 days to reach $4T and just 78 days to reach $5T. 庄明浩 calls it the biggest of “those freakishly powerful and glamorous beasts.”
17. The Same Earnings Season Showed Why Strategic Position Matters More Than Being an “AI Giant”
Google posted quarterly revenue above $100B for the first time, with Google Cloud growing rapidly and signing large cloud and TPU orders with Anthropic. The market called it Google’s “NVIDIA moment,” and the stock rose about 7.6% after hours.
Meta also posted record revenue, but profit, rising CapEx and Llama’s slippage all came under pressure. One explanation for the stock’s 7.3% after-hours decline, 15% intraday drop at the open and 11% loss at the close was that Meta is neither Google nor Microsoft: it does not have a cloud business.
Microsoft’s cloud business grew roughly 40% as it continued to chase AWS. AWS, meanwhile, lifted its growth rate from the low-to-mid teens for more than two years back to roughly 20%, a two-year high; Amazon rose about 12% after earnings. These are multi-trillion-dollar companies that can still move by double digits overnight.
庄明浩 acknowledges that every post-earnings explanation is “all hindsight.” But the differences still show that the seven sisters’ respective mixes of cloud, chips, models, advertising and capital expenditure ultimately translate into very different stock-market reactions.
18. The Trading Theme Has Spread from Models to Every Layer of the Data-Center Supply Chain
In 2023, the MAG7 rose roughly 76% and other AI companies 45%. By 2025, the MAG7 was up about 16%, while energy, chips and software gained 53%, 41% and 33%, respectively. The market needed a new narrative, so capital spread from the seven giants into power, chips, storage and cloud.
Chips remain the absolute center of gravity. Beyond NVIDIA, AMD, Broadcom, TSMC, Intel and even Qualcomm are competing for pretraining and inference share, while Google and AWS have joined with in-house chips. Chinese GPU companies receiving rapid approval to list also reflects this strategic position in the industry.
The formerly simple chain—funding, data centers, model training, revenue—has lengthened to include GPUs, networking, land and buildings, energy, memory, hard drives and even silicon supply. Many of the hot themes in China’s A-share market can also be located on this global supply-chain map.
19. Even Storage, at 1% of Costs, Is in Shortage—Showing Demand Is Nearing the Physical Limit
One data-center cost breakdown puts GPUs at roughly 40%, networking at 13%, land and buildings at 11%, energy at 10% and storage at just 1%. Yet even that 1% is now in shortage. Seagate and Western Digital were once treated as sunset businesses, but became the two best-performing stocks in the S&P 500 through Q3.
After the Redmi K90 launched at a price that drew user backlash, Xiaomi cut the K90 series by RMB300 the next day. Lei Jun explained: “Memory prices have risen so much recently. We hope everyone can understand.” 庄明浩’s reaction was: “Who would have thought that both memory and hard drives would run short?”
20. Neocloud Is Tearing Open the Public-Cloud Market as a “Compute Specialty Store”
AWS, Microsoft Azure and Google Cloud remain the top three, but once the installed base stabilizes, the focus shifts to incremental demand. Microsoft temporarily surpassed AWS in quarterly additions through its OpenAI tie-up, while Oracle used OpenAI orders to raise expectations for its cloud business.
Oracle Cloud had revenue of roughly $10B at the time, yet projected growth to $144B over the next five years, an almost exponential curve. Its primary support was also OpenAI.
The Neocloud market is expected to grow 69% annually over the next five years, with CoreWeave and Nebius as representative players. AWS is like a department store offering more than 200 services; Neocloud is a specialty shop, with its website stating exactly how much one GPU costs per hour or month. Customers buy only cards, connections and compute services.
NVIDIA has built a network of “sons” through equity stakes and supply agreements. On the program’s cited figures, CoreWeave accounts for the largest share of NVIDIA’s external equity holdings. When chips are scarce, priority access to the newest GPUs and discounts naturally turn equity relationships into competitive moats.
21. There Is More Money in Private Markets, but It Is Flowing Almost Entirely to Fewer and More Expensive AI Deals
Global financing in the first three quarters of 2025 had already surpassed the full-year totals for 2023 and 2024, with quarterly funding above $90B for four consecutive quarters. The number of deals, however, continued to fall, meaning individual rounds were getting larger and valuations more expensive.
AI projects accounted for about 51% of deal volume but took 85% of the capital. 庄明浩’s literal summary is “the money is all being taken away”: capital is not embracing AI evenly; AI, especially late-stage leaders, is absorbing the overwhelming majority of liquidity.
Average financing per company has reached roughly $24M, above the $21.7M recorded in 2021. Rounds above $500M accounted for just 18% of market value in 2021 but roughly 63% today. Conventional VC funds may have only a few hundred million dollars in total, yet face companies raising hundreds of millions in a single round.
In less than three years since ChatGPT launched, more than 100 new AI unicorns have appeared. Another count puts the existing total at 203, with the average AI company taking two years to become a unicorn versus nine years for non-AI companies. The top 30 funds also took roughly 75% of market capital, showing heavy concentration on the fund side as well.
22. Companies Are Going Public Later, and Exit Markets Are Growing “Wild” New Routes
IPOs remain cold, while M&A has clearly recovered, and buyers are no longer limited to public companies. OpenAI has used its $500B equity valuation as leverage to acquire companies, with the largest deal this year involving a hardware company co-founded by former Apple designers.
The average time from startup to IPO has lengthened from 12.2 years a decade ago to 15.9 years today. Of six U.S. private companies valued above $100B, five are related to AI. The larger and less urgent a company is about listing, the stronger the exit pressure on early investors.
The first route is secondary shares: even a tiny holding is enough for someone to call themselves an “investor” in OpenAI, Anthropic, xAI, SpaceX or ByteDance. The more aggressive route is Robinhood’s move to put private-equity interests on-chain as stock tokens, allowing ordinary users to trade them through Web3.
23. Sam Cleared the Old Liabilities First, Then Maxed Out the Next Five Years of Expectations
Sam divided his recent work into “not dwelling on the past” and “going all in on the future.” The first category included protecting technology, users, revenue and the core team; dismantling the old cap on investor returns; completing the restructuring of the for-profit entity; securing government approval; and renegotiating the Microsoft agreement.
The result: OpenAI is valued at roughly $500B, ChatGPT has 800M weekly active users, projected 2025 revenue is $12B, SoftBank’s final roughly $22.5B tranche has arrived, and many employees became billionaires by selling old shares. Sam again confirmed that he holds no OpenAI equity and receives only a salary.
But as early as February 2024, when OpenAI was valued at roughly $86B, with $1.6B in revenue the previous year and about 100M weekly active users, Sam had said AI would need $7T. Everyone thought he was insane. Now the company’s valuation is roughly 6x higher and weekly users roughly 8x higher, while that number has started to enter the realm of discussion.
Looking forward, OpenAI expects revenue to reach $200B around 2030, but may not generate positive cash flow until 2029 or 2030 and will burn roughly $115B in the meantime. Unable to fill the gap on its own, it can only transmit those future revenue expectations to suppliers and capital markets.
24. OpenAI and NVIDIA Have Built a Self-Reinforcing Capital Loop
Stargate first pushed infrastructure expectations for the next five years to $500B. OpenAI then promised Oracle roughly $300B in cloud contracts over five years. Desperate for orders after falling behind in the cloud era, Oracle saw its stock surge once the contract landed, while its revenue curve was rewritten.
OpenAI also signed with AMD for roughly 6GW of compute, with OpenAI able to receive up to 10% of AMD’s shares; it partnered with Broadcom on roughly 10GW; then it signed with NVIDIA for 10GW, while NVIDIA in turn committed to invest $100B in OpenAI. The “golden finger” was complete: whoever receives the five-year expectations sees its valuation and revenue expectations take off.
The loop works like this: capital goes to OpenAI, OpenAI orders from cloud and chip companies, supplier revenue and share prices rise, and investment, equity or financing capacity flows back into OpenAI. The meme of power strips connected to one another but with no external power source captures the market’s mockery of circular financing.
NVIDIA is happy to facilitate it. It generates roughly $26.4B in quarterly profit and holds about $100B in cash; beyond binding itself to OpenAI, it invested $5B in Intel and $1B in Nokia. It is no longer merely a supplier but is allocating credit and liquidity across the entire industry.
25. From $100M to $1T, Every Threshold Is Breaking an Old Record
$100M was once close to the transfer fee for a Cristiano Ronaldo-level football star. Today it can also be the “transfer fee” for a top AI scientist or the ARR threshold for a star startup. AI labs increasingly resemble sports clubs: owned by billionaires or tech companies, with star employees earning more than $100M a year and huge internal pay gaps.
$1B is the unicorn threshold; $10B is now the entry ticket for an AI star company’s valuation and roughly the scale of OpenAI’s $12B and Anthropic’s $9B in annual revenue. The Meta–CoreWeave deal at about $14.2B, Microsoft–Nebius at about $17.4B and OpenAI–CoreWeave at about $11.9B also sit in this bracket.
Historically, Uber burning roughly $18B, Tesla $9.3B and Netflix $11B were considered extreme. OpenAI expects to burn another $115B. Google’s acquisition of Wiz cost about $32B, while NVIDIA’s quarterly revenue was roughly $46.7B—slightly above the market cap of Baidu, which 庄明浩 checked that day at about $43.8B.
A 1GW data center is estimated to require roughly $50B in investment. OpenAI has committed to building more than 30GW, with roughly $1.4T in spending commitments. If it eventually lists at a $1T valuation and sells 10% of the company, it would need to raise $100B. The largest IPO in history, Saudi Aramco’s, raised only about $26.2B.
26. The Bubble Is Now the Default Assumption; the Debate Is Only About Type and Whether “This Time Is Different”
Sam himself acknowledged publicly in August or September that AI investment contained a bubble. By Q4, Substack was running an “AI bubble” headline almost every day. 庄明浩 observes that the discussion no longer asks “is there a bubble?” but “what kind of bubble is it?”
Bezos calls it a “good bubble,” meaning something like the internet rather than tulips. Another four-quadrant framework classifies bubbles by productive versus unproductive and equity versus debt; AI currently looks more like a productive, equity-driven bubble. Looking back over 400 years, Coatue further divides bubbles into stock markets, infrastructure construction, real estate, credit, collectibles and other categories.
Supplier financing, government support, enormous capital requirements and overbuilding are all already visible. Collapse, consolidation and eventual value creation remain to be seen. Every cycle produces someone saying “this time is different,” and financial history even contains multiple books titled This Time Is Different.
The evidence for “different this time” is not nonexistent. The average P/E of leading companies during the internet bubble was roughly 89x; today’s AI leaders average about 28x. Cisco once reached 101x, while current multiples are about 31x for NVIDIA, Microsoft and Apple; 22x for Google; 20x for Meta; 24x for Amazon; and 37x for Broadcom.
27. The Real Turning Point Is Debt and GPU Depreciation, Not a Single Claim That Valuations Are Too High
CapEx as a share of operating cash flow is currently about 54% at Microsoft, 55% at Google, 65% at Meta and 80% at Amazon, but 135% at Oracle. The first four can still fund investment through their own businesses; Oracle has to borrow. Meta’s roughly $27B of data-center bonds were still four times oversubscribed, and easy financing may tempt CEOs to add more leverage.
Railways and fiber can depreciate over 25–30 years; even if the builders fail, the assets can still serve later users. GPUs, by contrast, may have a reasonable economic depreciation life of only about three years because new cards iterate so quickly. If the largest cost item in a data center must be depreciated rapidly, the AI revenue required to cover it is far higher than for historical infrastructure.
The practical case for continued investment is that Big Tech is first using AI to transform its own businesses. Google disclosed monthly token consumption across the company of roughly 1.3 quadrillion, up from 980 trillion and 480 trillion previously, with earlier levels lower still. The main consumer is not an external startup but Google itself; ByteDance and Volcano Engine are similar.
A five-indicator model has not yet turned fully red: AI investment is about 0.9% of GDP, near the 1% danger line; investment and industry revenue are severely mismatched, but revenue growth remains fast, valuations sit between green and yellow, and funding quality is still considered healthy under that framework. 庄明浩 retains his objection: “Others already think this is extremely dangerous.”
28. The Hardest Investment Move Is Not Identifying the Bubble but Avoiding Being Too Early in Proving Yourself Right
The crisis framework divides the process into three stages: bubble formation, asset misallocation and collapse. First come the people making blind linear extrapolations, alongside those who know prices are too high but fear selling too early. Capital then flows toward unsuitable companies in hot industries, and the eventual decline may be triggered by shadow-bank debt and runs.
But “don’t be the genius who gets off early” matters just as much. Greenspan warned about the internet bubble as early as 1997; even after the bubble burst, the index never returned to the level at which he issued the warning. Peter Lynch said the money lost trying to predict a correction is often far greater than the money lost in the correction itself. “A bubble is simply a bull market they did not participate in.”
The final data set brings the grand narrative back to ordinary life. Among 101 conversations at the May podcast festival, 12 involved AI, or 12%; among 128 this time, only 8 did, or 6%. Attention had been cut in half. 庄明浩’s closing conclusion was the most straightforward: “AI or not, it doesn’t matter. Not that many people care about this stuff. Enjoying life may be the most important thing.”