Summary

庄明浩’s keyword for 2025 is “inflection point,” because the turn could be either up or down. If construction proceeds as expected, data-center demand could explode in 2026; if the bubble breaks, the turn may already have begun by the end of 2025. The key is not choosing long or short, but recognizing that technology, products, users, capital and even talent have all “unconsciously reached their limits,” while five years of capital expectations have been compressed into today’s prices.
DeepSeek R1 set a million-dollar-scale final training cost against the United States’ billion- and tens-of-billions-dollar spending—and the $500B “Stargate” narrative—redefining the year’s competitive landscape. It triggered a sharp selloff in NVIDIA and the “Magnificent Seven,” while quickly generating the counter-narrative that greater efficiency ultimately expands compute demand. Its more durable impact was to put reasoning models, open versus closed source, MoE, cost, and competing US-China technology paths on the main agenda at once. “Everything after that became engineering optimization and improvement.”
Pure language models are still improving, but the AGI leap the market once priced into GPT-5 never arrived; the ChatBot distribution war ended first. Sam Altman’s own post-mortem is that humanity may already have “quietly walked past” the Turing test without any earth-shaking change. 庄明浩 therefore describes the past three years as close to linear evolution. Meanwhile, ChatGPT’s weekly active users reached 800M; even if Gemini rose from a few tenths of a percentage point to a few percentage points, it remained in catch-up mode. “The pure ChatBot battle is over.”
Multimodal monetization may arrive faster than language-model monetization, because images, video, marketing, e-commerce and creator tools already have distribution and demand chains that need no fresh proof. Veo 3 was the first product to give mass users a sense of the leap from silent to sound-synchronized video generation. Chinese companies caught up in roughly 3 months; 庄明浩 estimates that an indigenous equivalent of Sora 2 could appear in about 2 months. DeepMind sees world models as “another main table” on the road to AGI, but Genie, domestic Hunyuan 0.1 and 3D interaction remain largely at the demo stage.
The “year of the Agent” may last 5 years, because L3 crosses from generating language into executing actions; the real difficulty lies beneath the surface in protocols, browsers, APIs, memory, infra and workflows that have yet to be standardized. Manus matters not only for its features, but because it gave users the mental model that “this is roughly what an Agent should look like,” and produced roughly $90M in ARR. Meanwhile, every company is pursuing a different memory architecture, leaving cloud vendors with only one answer: “Only children make choices—we’re all in.” The lack of standards raises implementation costs but also leaves a window open for startups.
Open source has evolved from idealism into a technical, commercial and geopolitical strategy for Chinese model companies. A model’s survival depends ultimately on users, token consumption and ecosystem coverage—not on which benchmark it won; open source does not preclude charging for APIs and To B services, while enabling local deployment, data trust and sovereign-AI choice. DeepSeek and Qwen have therefore become the central Chinese storyline, while Kimi K2 and GLM-4.5/4.6 continue to establish positions in coding, Agents and open source.
The obstacle for China’s application layer is not a shortage of models, but the fact that entry points are occupied by “castles” such as WeChat, Douyin and Xiaohongshu, while To B payment and exit paths remain weaker than in the US. The program cited QuestMobile data showing roughly three-quarters of standalone App and Web AI applications suffering user declines, while plugins grew by roughly two-thirds—but the question is, “Who is willing to settle for being a plugin?” This is pushing investors toward AI hardware, overseas teams and founders with industrial backgrounds at DJI, Dreame, Roborock and Insta360, while temporarily setting aside exits that cannot yet be calculated.
OpenAI is turning itself from a model company into a system variable for the capital markets: a headline can send AMD up roughly 40%, Broadcom up roughly 10%, and even let NVIDIA act like a “central bank,” using cash flow to support the entire cycle. OpenAI is now valued at $500B, yet it is writing 5-year and 10-year contracts for an industry whose technology changes monthly, allowing partners to fully price in distant expectations at the same time. The real fragility lies in debt, data-center equipment that may be obsolete within 2 or 3 years, storage and power bottlenecks. This is not simple, deliberate “collusion”; it is a system in which everyone can no longer afford to sit out and watch the consequences compound.

Deep dive

1. 2025 Is an “Inflection Point” Because Technology and Capital Are Hitting Their Limits

庄明浩 did not simply call 2025 the year of the bubble or the year of the turn. He chose a word that can be read in both directions: “The benefit of an inflection point is that it can turn upward, and it can turn downward.” If data-center construction is delivered, 2026 could bring an upside explosion; if the bubble turns into a break, the decline may already have begun by the end of 2025.
The sense of limits 庄明浩 describes mainly concerns technology, products, money, capital and even the broader AI debate—all of which have reached their extremes. Few people still repeat the opening line “one AI year equals 10 human years,” but the experience of all 3 participants working in the industry remains that it feels “like someone pressed 8x speed.”
庄明浩 repeatedly stresses that the point is not that the outcome is already known, but that “we have unconsciously reached the limits of many things.” In his framing, AI may even have pushed the question to “the limits of humanity.”

2. DeepSeek R1 Punctured the US Compute Narrative Through a Cost Mismatch

R1 was released roughly 2 days before the Lunar New Year holiday. 冯骥 called it potentially a “national-fortune-level” event. US financial and technology media covered it heavily at the time, focusing on the more-than-$1M final training cost listed in the R1 paper and its sharp contrast with the $1B- and $10B-scale infrastructure narrative taking shape in the US.
The “Stargate” plan was discussed at a scale of $500B, making R1 a case of “Chinese people being very good at accomplishing big things with small amounts of money.” Markets immediately questioned the logic of existing infrastructure spending, and NVIDIA and the “Magnificent Seven” suffered enormous declines: the narrative once treated as the only path suddenly had an alternative solution.
The counter-narrative that followed borrowed from the history of coal: greater efficiency may cause short-term turbulence while expanding total demand, ultimately supporting longer-term compute growth. The program did not decide which narrative would win, only noting that the debate was still unresolved by year-end.

3. R1 Drew the Year’s Reasoning-Model Roadmap in Advance

OpenAI proposed the L1-to-L5 framework in 2024: L1 is ChatBot, L2 is reasoning, and L3 is Agent. o1 became the representative reasoning model in September 2024; from then on, reproducing and engineering this path became the common task for leading companies.
R1’s lasting significance lies not only in cost, but in making reasoning models a question every company had to answer. Roughly 2 to 3 months after its release, leading companies had largely delivered reasoning models of their own. Competition then shifted toward foundation models, reinforcement learning, reasoning efficiency and engineering optimization.
庄明浩’s annual judgment is clear: “If you look only at pure language models, R1 has basically described the competitive path for the year and the battles to be fought.” Technology routes, open versus closed source, parameter scale, MoE and US-China comparisons in product deployment were all pulled onto the main stage by that release.

4. GPT-5 Did Not Create an AGI Moment, While Google Returned Forcefully to the Table

OpenAI had led the industry to expect GPT-5 to “cross the AGI threshold,” but the actual release did not produce a commensurate shock. Sam Altman said on a podcast that major thresholds need not come with dramatic moments: GPT-3.5 and ChatGPT had, in some sense, already crossed the Turing test, but humanity merely “quietly walked past it and kept moving.”
庄明浩 therefore characterizes the progress of pure language models since November 2022 as linear or close to linear. Models continue to improve, but the marginal excitement from benchmark wins and version updates is diminishing, while the industry still struggles to define what the arrival of true AGI would look like.
Google was the year’s strongest technical comeback, with Gemini, Veo 3, Nano Banana and world models forming a coherent signal. Anthropic chose To B and coding more firmly, catching up through use-case differentiation; its revenue curve may even be steeper than OpenAI’s, albeit from a smaller base.
xAI has built a distinctive position through its integration with X, virtual companionship and a more “untamed” generation boundary. Microsoft, amid its complicated relationship with OpenAI, began releasing proprietary models and chose voice for its first move. Meta became quieter on model releases, with its most conspicuous action being “writing big checks to recruit people.”

5. China’s Model Story Has Converged on DeepSeek, Qwen and Open Source

The “six little tigers” are no longer an effective annual narrative. The stronger consensus among Chinese companies has shifted to open source. DeepSeek and Qwen sit in the front rank: the former has built global influence through technology and papers, while the latter is tied to Alibaba Cloud, investment and a full-stack AI narrative spanning language, multimodality and programming.
DeepSeek V3 was released at the end of 2024, while R1 came from reinforcement training on top of the foundation model. After continued updates through V3.1, the market expected V4 but got V3.2 instead. 庄明浩 therefore believes 2025 “may not be the year we get V4,” and R2 may not arrive on schedule either.
Kimi K2 and 智谱’s GLM-4.5 and GLM-4.6 remain on the coding, Agent and open-source tracks after adjusting their strategies. Europe also has companies such as Make Fun remaining at the main table for natural-language models. Compared with the US, China is pursuing a path that combines open source, engineering ingenuity and resource deployment—not a single myth of low cost.

6. “World’s Best” Is Losing Its Value; Voice SOTA Looks More Like an Organizational Credential

The host cited 可灵 and 元宝, which successively claimed to be “world’s best.” Each selected a different leaderboard and ranked first on a particular dimension, yet neither made the top 10 on the leaderboard chosen by the other. 庄明浩’s response is therefore to ask first, “World’s best where?” and then examine the blind tests, question sets and scoring methods.
When every model can find a leaderboard on which to publish an SOTA result, the external impact of the score keeps shrinking. Some releases look more like “managing upward.” The metrics with real industrial significance are users, token consumption, developer coverage and actual revenue—not a one-off ranking detached from the use case.
Voice was once viewed as a peripheral battlefield with low investment and high cost-effectiveness. Xiaomi, Bilibili and Xiaohongshu all used results from particular voice leaderboards for PR. OpenAI pushed into Realtime, while Microsoft’s proprietary models also started with voice. But once a “depression” is discovered, every company enters, and the lead is quickly erased.

7. The ChatBot Battle Is Over; Moats Are Shifting to Habit, Memory and Distribution

庄明浩 offers a stark conclusion: “The pure ChatBot battle is over. ChatGPT has already won.” ChatGPT’s latest disclosed weekly active users reached 800M; a year earlier, the figure may have been only a few hundred million, meaning it doubled—or even doubled again—in less than a year. Even if Gemini rose from a few tenths of a percentage point to a few percentage points, it was still starting from a very small share.
Whether users pay increasingly depends on usage habits, brand understanding, context and long-term memory rather than a single benchmark. 庄明浩 was initially more hesitant, but by year-end he leaned toward the view that the flywheel “is beginning to work”: models are accumulating an understanding and memory of users.
OpenAI is also turning ChatGPT into an all-in-one entry point, directly calling services such as Spotify, Zillow, Canva and Figma inside conversations in a form resembling Chinese internet mini-programs. Pulse is intended to convert weekly active users into daily active users, with an interaction that is “very smooth, very natural.”

8. Sam Altman Runs OpenAI Like an Investment Portfolio

The 3 participants repeatedly returned to a background the market has forgotten: Sam Altman was originally an investor and taught entrepreneurship at Y Combinator. He has said, “I’m not good at management; I’m an investor.” Inside OpenAI, he behaves more like someone supporting individual teams, continuing to allocate resources to ideas and scaling the projects that work.
That ability shows up in both product and communications. After launches, partner companies display token-use medals in different colors; viral Sora 2 content then parodies Sam or puts users in the frame with him. Every interaction binds users more tightly to the OpenAI brand.
The host compares Sam to Elon Musk: “I don’t care how you parody me.” If 张一鸣 were willing to become a similar public IP, ByteDance might gain the same viral distribution lever, but the host believes that is nearly impossible in practice.
OpenAI can no longer be described as a single company. It simultaneously resembles a product company, a technology lab, an infrastructure company and an “other company” carrying new projects such as hardware. 庄明浩 therefore sees OpenAI, at a $500B valuation, as a multi-entity portfolio managed by an investor.

9. “Accomplishing Big Things with Small Amounts” Did Not End Spending; It Forced the Two Paths to Converge

Frontier-model development increasingly looks like an engineering problem involving technology selection, trade-offs and implementation rather than pure zero-to-one inspiration. The US tends to believe in “throwing enough force at the problem,” while China looks for more resource-efficient paths; once one side finds an optimization, the other learns quickly, and once one side stacks up performance, the other tries to replicate it.
Of algorithms, data and compute, compute is the easiest to solve directly with money; talent can also be attracted with money. The program cited the judgment that “no company has gone bankrupt by paying a genius employee a high salary.” When the biggest companies discover that annual spending has reached tens of billions of dollars and they still cannot exhaust their budgets, paying extraordinary sums to poach talent becomes rational.
The program discussed an extreme example: a company co-founded by one of Scale AI’s founders was valued at roughly $12B, and he held at least 10% of it, yet Meta still lured him away with $3.5B. The host had also cited 广明 and 小俊’s formulation that US financial power combined with Chinese engineers is a defining feature of this AI cycle. 庄明浩 believes the arms race could continue rolling forward for years.

10. Multimodality Is the Closest US-China Red Ocean—and the Fastest Productivity Payoff

Images and video are now difficult to discuss separately, and Chinese companies have shown strong parallel-running capability in this arena. Douyin, Kuaishou and numerous video startups have continued iterating; after Veo 3 defined synchronized sound and picture, Kling and ByteDance-related models caught up in less than 3 months.
After Sora 2 launched, the market immediately asked when a domestic version would appear. 庄明浩 estimates that “2 months might be more or less enough.” His basis is not that capabilities must be identical, but that once the product boundary, technology choices and route are clearly defined, engineering teams in China and the US can catch up extremely quickly.
Unlike language applications in law, finance and HR, images and video already have mature advertising, e-commerce, short-video and marketing chains. Once a model crosses a threshold in controllability, quality or cost, it can immediately enter existing production workflows. The host views Meitu’s stock performance over the past year as supporting evidence of this speed of monetization.
China’s resource advantage lies in its short-video, e-commerce, marketing and tool ecosystems. The challenge is whether companies choose To B, To C or a tool platform. Technical capability can converge; the real difficulty is “cutting to the right boundary and packaging it properly.”

11. Sora 2’s Value Is Not “AI Douyin,” but the Productization Mechanism

MiniMax’s Hailuo had already shown world-leading motion reconstruction with a video of “a kitten competing in Olympic diving.” Jimeng had operated the image community “离谱,” and other teams had built image or video communities, but only Sora 2 truly broke out beyond the AI circle. The host’s question is that the difference comes not only from the model, but also from OpenAI’s brand, product packaging and distribution resources.
Before the “离谱” team shut down, the community’s early retention and engagement metrics were not poor, but it still failed to continue. The market’s optimism about such communities may previously have been only 30 points; after Sora 2 appeared, confidence in the same direction rose to roughly 60 points.
“The world does not need an AI version of Douyin” can be true, because Douyin’s content ecosystem, retention, monetization and user reach are already mature. But the program’s reversal is: “OpenAI needs Sora 2.” It needs a standalone product to carry the technology and an organizational mechanism for repeatedly packaging research into products.
The host stresses that OpenAI may not actually want to copy Douyin; “AI Douyin” is a proposition imposed by outsiders. The more important zero-to-one goal is to show the world that a new capability exists and let internal teams validate a complete SOP spanning model, interaction and distribution.

12. World Models May Be Another Main Table on the Road to AGI

庄明浩 lays out the technology path layer by layer. Natural-language models rest on the belief that language carries human civilization; coding may be a subset of language or a tool for building new worlds. Voice, images, video and 3D then fill in the sensory experience language cannot fully describe.
Once these local battlefields are packaged together, the endpoint is a world model. DeepMind’s view is that it is not merely a branch of the language-model main table, but may become “another main table” on the road to AGI—likened even to “the womb of the world,” generating environments that obey physical laws and support continuous interaction.
The true inflection point requires generation to be faster than human perception and action: think it and it appears. This is not watching a one-off video, but changing and exploring a world in real time. Genie creates a feeling of being “electrified from head to toe” because of this possibility, though the program acknowledges that commercial deployment remains far away.
The domestic Hunyuan world model is roughly at the 0.1 stage, able only to convert a photo into a relatively primitive interactive space, with limited image quality and clarity. Game companies have a natural use case; users of Bambu Lab machines also call models such as Hunyuan to generate 3D objects for printing, suggesting that 3D printing, alongside games, could become an early landing zone.

13. Creator Tools Are Not Afraid of Homogenization Because New Demand Keeps Arriving

Asked how to differentiate when there are already so many video tools, one entrepreneur answered, “Differentiation is not that important; the market is huge.” The claim is that TikTok adds roughly 1M users globally every day, 12% of whom click the plus sign in the middle—equivalent to 120,000 new video creators every day.
All of these creators need video tools, regardless of their level of professionalism. The program cited data showing roughly 20 companies in video tools with ARR above $20M; many solve only one specific problem or serve one particular country, and outsiders may never even have heard of them.
New products such as Higgsfield and Polo, created by its founder, can still acquire users and revenue quickly. That shows multimodality is not a single-point market reserved only for foundation models. Photography has been one of the most fiercely competitive categories since the early App Store: “It is still being fought today” because demand is hard and use cases are numerous.

14. Agents Cross from Language into Behavior; the First Year May Last 5 Years

In the L1-to-L5 framework, L2 reasoning still delivers an answer through dialogue, while an L3 Agent must actually “get work done.” It may operate a computer, website, database or enterprise system; the output is no longer just a few words, making reliability, security and the execution environment more complex at the same time.
庄明浩 agrees that 2025 is the year of the Agent, but adds a note of caution: “Maybe the next 5 years will all be the year of the Agent.” General-purpose versus vertical is only the first fork. Permissions, scenarios, tools, workflows and degrees of automation come next, much as autonomous driving may remain at one level for a long time.
Manus’s greatest significance after its March launch was showing users what an Agent’s interface, interaction and output “should roughly look like.” Even if later entrants claimed they could reproduce it in 3 days or 3 weeks, its brand, timing, details and position in users’ minds could not be replicated with code.
Manus later disclosed roughly $90M in ARR and introduced RRR, a metric better suited to observing revenue trends. The host believes its growth curve exceeds that of many widely recognized high-growth companies, proving that Agents have moved from a viral concept into real revenue.

15. There Is Still No Standard Beneath the Agent Iceberg, So the Startup Window Remains Open

Agents offer non-model startups a new paradigm. What companies could do in the past was often dismissed as “wrapping a model,” but “wrapping is not a derogatory term”; it simply left limited room to maneuver. Today, protocols, browsers, APIs, scaffolding, infra, memory and the execution layer can all become independent companies.
One of the biggest divides is whether to remain compatible with an internet designed for humans or rebuild systems callable by machines. Even choosing compatibility requires deciding between browsers, pure APIs and other paths. Protocols remain unstandardized, and large companies and startups are competing for the entry points.
Memory is only one corner of the iceberg. Before interviewing relevant founders, the host found that every company was pursuing a completely different route. 庄明浩’s team could not determine whether to build in-house, adopt an open-source solution or buy a mature API, because no unified evaluation standard exists.
Cloud vendors can only go all in on the same problem. Asked how to prioritize the components of an Agent base, Alibaba Cloud’s top executive replied: “Only children make choices; we cloud vendors are all in.” Large companies will cover the full chain, but cannot clear every battlefield in the short term; the standards vacuum still leaves startups room to compete.

16. Agents Have Not “Cooled Off”; Attention Has Simply Returned to Model Releases

July 2025 was China’s “crazy July” for open-source models, with launches appearing almost every day. In August, GPT-5, Claude and other leading releases pulled attention back to foundation models. Agent companies entered the phase of building users, revenue and implementation, so their exposure naturally became less concentrated than during Manus’s debut.
A16Z Speedrun showcased 58 startups in Los Angeles. The first category identified on site was Agent as a Service: teams building Agents for different vertical industries and specific tasks. It does not shock the market every day, but it has already entered the startup mainstream.
The A16Z-related revenue rankings mentioned companies including 剪映, Kling, Midas and Jasper. 庄明浩’s conclusion is that the ChatBot brand battle is over, but the broader AI war continues; Manus has simply secured a relatively solid first-mover position.

17. Open Source Has Become a Technical, Commercial and Geopolitical Strategy

Before DeepSeek R1, Sam Altman believed the leading model would necessarily be closed source because it required enormous resources. After R1, the market began accepting that open-source models could quickly follow the frontier. The program cited a State of AI forecast that an open-source model will certainly hold the top spot for a period of time in 2026.
As capability gaps narrow and cost gaps widen, open source can bring developers, usage, feedback and ecosystem coverage. 庄明浩 stresses: “No matter how powerful a model is, if nobody uses it, it means nothing.” Open source also does not mean a company cannot make money; DeepSeek can still charge for APIs and To B services.
Open source is also a trust mechanism. Enterprises can deploy locally, reducing concerns about data leaks and secondary use. The host observed a ROG laptop with substantial memory and VRAM selling out quickly at a relatively reasonable price, precisely because it was suitable for running models locally.
The larger variable is sovereign AI. Countries can build data centers, but still have to decide whose models to use. How Japan, Southeast Asia, the Middle East and Africa choose outside the US-China framework could make open source a “weapon,” in quotation marks. Shanghai has even proposed cash rewards of RMB5M for influential open-source products or communities.

18. Coding, Vertical Agents and Companion Apps Form Three Underlying Application-Layer Tracks

AI Coding was an unavoidable application breakthrough throughout the year and has already split into multiple battlefields, including frontend, backend, databases and IDEs. Large companies and startups in China and the US are all releasing products and competing for funding; coding is also Anthropic’s clearest core use case.
General-purpose Agents are the visible theme, while vertical Agents are rapidly absorbing capital in law, finance and marketing. Law requires industry information, workflows, privacy and data security. Finance can be split into primary markets, secondary markets, banking and insurance; marketing is further divided into search, email, images, video and online and offline channels.
These vertical companies may not become especially large, but the segment is showing “abnormal prosperity” because every narrow problem can generate revenue. The large number of Agent as a Service startups at A16Z Demo Day is a concentrated expression of this fragmented demand.
Social and companion applications are less discussed but continue to evolve. Among the latest top 50 on H6Z, roughly 10 are Web products and roughly 12 to 13 are Apps. 庄明浩 describes the product generations as pure conversation, then immersive interaction through images and voice, and finally contextualized experiences.

19. China’s AI Applications Are Blocked by a “Castle-Style Internet”

庄明浩 believes it is “quite difficult” for China to produce an application like Cursor that defines a new category and interaction paradigm. Some products near the top of rankings are “90% traditional App and 10% AI features.” They can be placed on an AI leaderboard, but are difficult to regard as native new species.
China cannot talk as freely as US startups about achieving a certain ARR in a certain period, so it has to focus more on user counts. But mobile internet and Web entry points are already saturated. The program cited QuestMobile data showing roughly three-quarters of standalone App and Web AI applications suffering user declines, while plugins grew by roughly two-thirds.
Plugins can grow through existing entry points such as WeChat, Douyin and Xiaohongshu, at the cost of losing their status as independent products: “Who is willing to settle for being a plugin?” Some teams can only use Android phones or RPA plus AI to cross platform restrictions and improve productivity in the gaps between the major “castles,” which do not welcome them.
The US has a more open Web, stronger enterprise payment capacity and better tool coordination, allowing To B companies to raise astonishing rounds week after week. China faces constraints from SME willingness to pay, closed platforms and exit paths. New trends cannot simply copy the US because the underlying foundations are different.

20. The Primary Market Can Only Search for an Answer Through Conviction, Hardware and Cross-Border Arbitrage

Faced with concentrated model financing rounds, difficult domestic application exits and the risks facing overseas teams, 庄明浩 believes dollar VC firms struggle to calculate “what exactly they are expecting.” In practice, they can only look first at the founder, product and background, decide whether to believe, and set aside exit questions that currently have no answer.
One fund managing partner reduced the question to its foundation: do we still believe technology companies need patient capital to complete the journey from zero to one? If so, we should keep doing it. 庄明浩 agrees with the logic but warns that it is “too idealistic,” with many constraints when applied in China.
Capital preferences are also shifting: from researchers, to big-tech product managers, to AI hardware, and then to founders with industrial backgrounds at DJI, Dreame, Laifen, Roborock and Insta360. Hardware is visible and tangible, generates direct revenue and benefits from China’s supply chain, making it more likely to provide investors with trust they can “sleep on.”
Overseas VCs are even revisiting early gaming teams that have been nearly 10 years outside the reach of Chinese domestic VC. The logic is not to discuss exits first, but that Chinese teams are “cheaper across the board” than overseas teams, while overseas gaming M&A markets remain active. A partner at Thrive Capital once spent 18 months studying a single company, a time horizon that also highlights the difference between the 2 capital environments.

21. OpenAI Has Turned the Market into a Systemic Bet—and Spread the Risk Across the Entire Chain

庄明浩 used his own PPTs to track the rising discussion of the bubble: one page appeared in the fourth deck, roughly 2 pages in the fifth, and around 6 pages in the September version; the next deck will contain even more. News involving OpenAI and AMD sent AMD up roughly 40%, while news involving Broadcom pushed the latter up roughly 10%, showing that OpenAI now functions like a “magic finger” for the secondary market.
The leading companies in the Magnificent Seven have reached market caps of $3T to $4T and roughly $100B in annual discretionary operating cash flow, while OpenAI is valued at $500B. The recently emerging 5-year and 10-year contracts allow partners to fully price in distant expectations at the same time. Technology changes monthly, yet capital is being asked to believe in 5 years.
NVIDIA can also support the ecosystem like a “central bank” through high gross margins, cash flow and market position. OpenAI, NVIDIA and their partners may not have coordinated in advance, but they are all “happy to see it happen.” Other than Google, which owns chips, cloud, models and products across the stack, almost no company can afford the consequences of staying out. AI has therefore evolved from an industry into “the system itself.”
Risk is concentrated at the bottom of the chain. CoreWeave, xAI and others have begun using debt, which is more rigid than equity. The fiber left behind by the internet bubble was still usable 15 to 20 years later; today’s data-center equipment and cards may be obsolete in 2 or 3 years. By mid-October 2025, the program said the S&P 500’s best performers included Seagate and Western Digital, showing that the boom had spread from GPUs to power, cooling, transformers and storage components: “Everything has been bundled together.”
The scale of capital has also broken historical reference points. The program cited Saudi Aramco’s roughly $22.6B global IPO as the largest financing record, already below a single financing round for OpenAI or Anthropic. Amazon had burned roughly $2B before its IPO; Uber roughly $40B. Extrapolate that by 20x, and the prospect of the next generation of companies burning $800B before listing no longer sounds absurd. 庄明浩’s closing formulation is that this is “an unintentional conspiracy”: by the time everyone realizes what is happening, retreat is already too difficult, leaving only the option to keep moving forward.

Is the Cycle Over, Portfolio Psychology, & Prediction Market Alpha Why Creativity Will Matter More Than Code | Kevin Rose and Anish Acharya