2025 AI in the Field: Our Year of Witnessing and Fantasizing
Summary
The most accurate keyword for 2025 was neither “bubble” nor “breakthrough,” but the directionally ambiguous “inflection point.” Zhang Minghao believes the curve will turn upward if data centers are built out in earnest in 2026; if the market truly enters a bust, the end of 2025 could also mark the beginning of the downturn. Technology, products, funding, capital—and the larger questions humanity faces around AI—have all “quietly reached the limits of many things.”
DeepSeek R1 challenged the single-track assumption that capability can only be bought with massive capex, with a final training cost in the million-dollar range, but it did not end the compute arms race. Released just before the Lunar New Year, R1 was called a “national-fortune-level representative” by 冯骥 and briefly sent Nvidia and the rest of the “Magnificent Seven” tumbling; the mainstream explanation soon shifted to the idea that efficiency gains would expand usage. China and the US are converging: China is looking for engineering tricks, while the US is “paying to stack,” but algorithms and data remain the things money cannot easily buy.
The battle for pure-Chatbot users may already be over, with ChatGPT occupying the high ground through 800M WAU, accumulated memory and an All in One product strategy. Even if Gemini rises from a few tenths of a percentage point to a few percentage points, Zhang still places it among “the rest”; OpenAI, meanwhile, is bringing Spotify, Zillow, Canva, Figma and other services into the chat box, attempting to build the unified entry point that Meta and Google failed to complete but WeChat once did. Sam Altman’s advantage lies not only in models, but in turning products, marketing and capital relationships into platform momentum where everyone wins or loses together.
The investment case for multimodality has shifted from leaderboard rankings to the speed at which model capability becomes productized and enters mature industrial chains. Veo 3 moved video from silent to voiced for the first time, and Chinese models caught up in roughly 3 months; Zhang expects a domestic product similar to Sora 2 could appear in “2 months.” China’s short-video, e-commerce and marketing ecosystems mean that once image and video capabilities clear the hurdles of cost and controllability, they can enter production immediately, while world models may become “another main table” alongside language models.
The core leap for Agents is from “language” to “action,” but this may be an Agent year lasting 5 years. Manus gave users a rough sense of “what an Agent should look like” and disclosed approximately $90M in ARR; the real startup opportunity lies beneath the iceberg—in protocols, browsers and APIs, memory, frameworks and permissions. Vertical Agents, AI coding and mobile OS Agents have already begun to monetize; the market has not cooled, it has simply stopped manufacturing a feed-filling moment every day.
Open source and HarmonyOS have become China’s 2 realistic paths to competing for the global model ecosystem and the domestic device entry point. Open source trades lower cost and local deployment for adoption, Token consumption and some degree of trust, and may become a “weapon” in competing for third countries outside Europe; HarmonyOS uses HMAF to let 小艺 identify intent and invoke Agents inside apps such as JD.com and Meituan. For developers, the new ecosystem still has open slots: the episode cites one person running approximately 10 apps on a single cross-platform codebase, generating RMB70,000 in monthly revenue.
The biggest systemic risk is that technology is still advancing month by month while capital markets have already priced in 5 to 10 years of expectations. OpenAI is valued at approximately $500B, and partnership announcements once lifted AMD by approximately 40% and Broadcom by approximately 10%; trillion-dollar infrastructure commitments, debt financing and GPUs that may be obsolete in 2 years mean today’s data centers cannot simply be compared with fiber that remains usable for 15 to 20 years. AI is no longer just a company or an industry; in Zhang’s words, it has become “the market itself.”
Deep dive
1. The 2025 “Inflection Point” Holds Both Boom and Bust
Zhang Minghao’s keyword for the year was “inflection point”: “It can turn upward, and it can also turn downward.” If data-center construction explodes as scheduled in 2026, 2025 was the eve of the upswing; if the bubble turns into a bust, the inflection point could begin bending downward from year-end.
His recurring observation was that the industry had “quietly reached the limits of many things”—not only technology and products, but also money, capital and even the larger questions raised by how humanity confronts AI.
The 2 hosts did not try to construct a complete chronology of events; they acknowledged that the review was closer to the annual memories of 3 people who lived through it. What ran through the episode was not which launch won, but the increasingly tight tension between accelerating technology and the system’s ability to absorb it.
2. DeepSeek R1 Redrew the Starting Line for the Reasoning-Model Race
R1 was formally released in late January, roughly 2 days before the Lunar New Year, after which 冯骥 called it a possible “national-fortune-level representative.” US mainstream financial and technology media covered it extensively, and the event quickly escalated from a model launch into a contest between US-China technology paths and capital efficiency.
The sharpest numerical contrast came from the final training cost disclosed in R1’s paper: roughly in the million-dollar range. At the time, the US infrastructure narrative had already reached the $1B and $10B scale, followed by the approximately $500B “Stargate” plan. The gap briefly sent Nvidia and the “Magnificent Seven” into a major selloff.
The market quickly repaired the narrative using a logic similar to coal-efficiency gains: lower unit costs do not necessarily reduce total demand and may instead expand use cases. Zhang believes that explanation has continued to hold; R1 did not stop compute investment, it merely proved that “there are other ways.”
Its longer-term significance was the technology path. OpenAI’s o1 appeared in September 2024 and demonstrated L2 reasoning; within 2 or 3 months of R1, around Q1, nearly every leading vendor had launched a reasoning model. Zhang’s conclusion is that R1 clarified the battles language models had to fight in 2025, leaving engineering optimization as the main work afterward.
3. GPT-5 Did Not Deliver the Imagined AGI Moment
OpenAI had allowed the market to build enormous expectations that GPT-5 would “cross the AGI threshold,” but the actual release did not deliver a shock on the same scale. GPT-4o, GPT-5 and subsequent versions together form a process of continuous evolution, not a clean historical break.
Zhang relayed Sam Altman’s view from the H6Z [?] podcast: even if ChatGPT has in some sense passed the Turing test, the world merely “stepped over it lightly, then kept moving forward.” True AGI may therefore not arrive as a bell that rings on a single day.
Restricting the discussion to pure language models, Zhang’s assessment of the nearly 3 years since ChatGPT launched is pessimistic: “It’s still moving in a linear or near-linear state.” Models continue to improve, but the news value of being “the world’s best” is increasingly difficult to reproduce.
This gradualism does not mean competition is weakening; it means the dimensions of competition are expanding—from model capability to distribution, memory, developer ecosystems, enterprise workflows, product stickiness and capex. A single benchmark increasingly fails to represent a company’s true position.
4. The Global Model Table Has Split into Specialized Battlefields
Google was one of the most aggressive technology performers in Zhang’s view of 2025, advancing Gemini, Veo 3 and world models on multiple fronts. Anthropic has committed more firmly to To B and coding, with revenue growth even faster than OpenAI’s, although from a smaller base.
xAI’s Grok uses X to access real-time content and is experimenting with virtual companionship and video models; its path was described as “very wild,” with fewer generation restrictions. Microsoft, meanwhile, began preparing a substitute for its supply relationship with OpenAI by launching its own model around August.
Meta’s voice in model launches has faded relatively, with its annual theme becoming “writing big checks to recruit people.” The episode noted that other things are happening there as well.
5. China’s Model Race Has Narrowed from the “Six Little Dragons” to an Open-Source Mainline
The old “Six Little Dragons” narrative has faded. Zhang believes China’s leading vendors have formed a clearer consensus around open source. DeepSeek and Qwen are at the front, and both are growing rapidly.
DeepSeek V3 launched at the end of 2024 and became R1 through reinforcement learning. After continuous V3.1 updates, the market was waiting for V4 but saw V3.2 first. Zhang therefore speculates that 2025 may pass without V4—and perhaps without the next R-series release either.
Qwen is best understood within Alibaba’s broader AI narrative: GPU investment is proceeding alongside full coverage of language, multimodal and coding capabilities. Kimi K2 and GLM-4.5/4.6 have remained on the mainline through coding, Agents and open source, while Europe still has Mistral at the table.
6. First Place on the Leaderboard Is Giving Way to Adoption and User Habits
When faced with “another world’s first,” Zhang’s first question is: first on which leaderboard? Blind tests, human ratings, fixed question sets and legacy benchmarks use different evaluation methods; “first” and SOTA have already seen a clear decline in marginal excitement among both the public and industry insiders.
He even believes some leaderboard results look more like “managing upward.” Voice models were especially crowded in 2025, partly because end-to-end voice and deployment demand strengthened, and partly because the investment required was relatively small—making the category a cost-effective way for late-arriving AI Labs to build attention.
Payment for Chatbots increasingly depends on habit, brand and long-term memory, not merely the quality of a single answer. Once context and “its understanding of you” begin forming a flywheel, later entrants are no longer facing a gap that one benchmark can reverse.
Zhang therefore offers a very clear judgment: “The pure-Chatbot battle is over. ChatGPT has won.” Its latest WAU was reported at 800M, having doubled from “a few hundred million” in less than a year; even if Gemini multiplies its share several times, it still has only a few percentage points.
7. OpenAI Is Turning the Chat Box into an All in One Entry Point
OpenAI demonstrated ChatGPT directly calling services such as Spotify, Zillow, Canva and Figma. Zhang compared it to “mini-programs”: users stay inside the chat box, where a unified entry point understands the request and calls existing services to complete the task.
This path is attempting to achieve the All in One that Meta and Google failed to build but WeChat did. It depends on ChatGPT’s brand and 800M WAU, while using memory to strengthen stickiness and gradually push WAU toward DAU.
The hosts specifically discussed Pulse, released roughly 4 weeks before Sora 2. Zhang believes that although it is a new product, it works “very harmoniously, very naturally,” strengthening retention and enabling users to put their accumulated memory to work.
When Sam was asked at the beginning of the year whether he wanted the strongest AGI or a platform with 1B users, his answer was of course “both.” Zhang believes he is in practice more biased toward the latter. Technical leadership still matters, but platform scale, entry-point status and user relationships have become equally core assets.
8. Sam Altman Runs Products, Organizations and Markets Like an Investor
Zhang relayed Sam’s self-assessment: “I’m not good at management. I’m an investor.” Inside OpenAI, he supports teams, allocates resources and lets ideas grow, acting more like an early-stage investor; an OpenAI insider gave the hosts a similar impression.
This connects to Sam’s time at YC: one of his best-known contributions was a startup course, not a manual for running a large company. OpenAI now contains a product company, technology lab, infrastructure company and entities that may handle new businesses such as hardware, making it easier to understand as an investment portfolio than as a single-product company.
OpenAI also gave partners such as Figma and Deliveroo colorful medals representing Token usage, while turning founders’ IP into promotional material for Sora 2. The hosts’ assessment was that “learning marketing from Sam Altman” could itself be a worthwhile topic.
This ability makes OpenAI look like a “magic finger of the secondary market”: point somewhere and that asset rises. Even if the bubble ultimately bursts, Zhang still believes OpenAI is very likely to be the one left standing, because Sam has turned partners into a community where everyone rises and falls together.
9. “Doing More with Less” Has Not Eliminated the Engineering Reality of “Paying to Stack”
When R1 launched, the market used it to mock Sam, who had just signed the $500B “Stargate” plan: why could a Chinese team do more with less while the US could only burn money? 6 months later, US investment had not contracted; it had continued expanding.
Zhang believes frontier-model R&D increasingly resembles an engineering problem: the key is trade-offs, paths and execution, not pure zero-to-one inspiration. The US chose “brute force produces miracles,” while China searched for tricks and detours; once either side finds an effective method, the other learns from it, and the paths ultimately converge.
Of algorithms, data and compute, compute is the easiest to stack directly with money; talent can also be attracted with high prices, but algorithms are especially difficult to buy. He cites the line: “No company has gone bankrupt by paying its genius employees a high salary.”
Meta invests tens of billions of dollars each year and has found that even hiring out Silicon Valley may not exhaust the budget, so it has pushed prices to extremes. The episode relayed a roughly $3.5B recruiting offer for the co-founder of a company valued at approximately $12B, whose personal wealth was already in the low billions.
10. The Unit of Multimodal Competition Is Now the Entire Industrial Chain
Zhang believes images and video can no longer be discussed separately, and that US-China competition is tighter here than in pure language models. Douyin, Kuaishou and startups such as Vidu and Pix Works are advancing together; once a cross-generational capability is defined, follow-on speed becomes extremely fast.
Veo 3 moved silent video into the era of sound for the first time, and Kling and ByteDance models caught up in roughly 3 months. After Sora 2 launched, Zhang estimated that a domestic equivalent might take only “2 months,” because the product boundary, technical choices and route had already been defined.
Multimodality does not need to prove the use case from scratch: Meitu’s share price, marketing-video Agents and existing content supply chains all show that once a model clears a threshold in controllability, quality or cost, it will immediately penetrate production workflows.
China’s resource advantages lie in short video, e-commerce, marketing and tools ecosystems; competition is therefore also more systemic. Photography has been one of the most crowded categories since the App Store launched and can still produce an app of the year today. “It’s just too hard,” but vendors must ultimately answer whether they are building To B or To C.
11. Sora 2’s Value Is Building a Productization Machine, Not Recreating Douyin
MiniMax’s Hailuo went viral with videos of a kitten performing Olympic diving and gymnastics, proving that model capability is not exclusive to OpenAI. But the product that truly escaped the AI circle and entered mass awareness was still Sora 2, reflecting a major difference in brand and distribution.
Before that, image communities such as “捏它” and “离谱” had made similar attempts. They achieved some early retention and activity but failed to generalize or persist, remaining largely within anime and subculture circles rather than reaching the broader public.
Responding to the criticism that “the world does not need another AI Douyin,” Zhang’s distinction is that the world does not need a copy of Douyin, but OpenAI needs Sora 2. It needs an independent product to carry the technology and an SOP, organization and staffing model for continuously turning research into products.
OpenAI can no longer be explained by a single startup org chart. It is at least simultaneously a product company, technology lab, infrastructure company and “other company” exploring areas such as hardware. Sora 2 is a clearly bounded experiment in how one product entity can use the resources of the entire company for distribution.
12. World Models May Be Another Main Table on the Road to AGI
The foundational belief behind large language models is that if machines master language unique to humans, they may reach superintelligence. Coding was then split into 2 interpretations: a subset of language, or an independent capability for creating new worlds. Multimodality exposes the limits of language in fully expressing visual and sensory experience.
Zhang calls voice, images, video and 3D local battlefields, and the end state of bundling them together a world model. A senior DeepMind executive’s metaphor is “the womb of the world”: the model obeys physical laws and can generate interactive worlds after a wait shorter than real time.
Under this framework, language is one main table and world models may be another, rather than an auxiliary function. Google’s Genie electrified the hosts because it demonstrated another possible route to AGI, even though it remains far from commercialization.
13. World Models Are Early, but 3D Has Found Entrances through Games and Printing
The domestic example discussed was Hunyuan World Model 0.1, which currently can only turn a single photo into an interactive scene; clarity and image quality remain very early-stage. Zhang’s view is not that the capability is mature, but that after images and video become crowded, 3D and world models will become the next competitive surface.
Game companies naturally need generatable, controllable 3D worlds, while 李飞飞’s World Labs is also working on world models. Until the technology is truly ready, the market will mostly see demos rather than a GPT-3.5-style inflection point.
The host added a real-world use case through the Bambu Lab platform: users can call models such as Hunyuan and Meshy to generate 3D objects and then print them. Hunyuan and Meshy may already be among the leading sources of calls. World models remain distant, but 3D generation already connects game content with physical manufacturing.
14. The Video-Creation Tools Market Can Support Many Non-Leading Companies
The host mentioned Higgsfield, whose founder had previously built Polo; this class of multimodal startups is showing strong data and revenue growth. When he asked about the differentiation of a Sequoia-backed video tool, the answer was that “differentiation is not that important,” because the demand pool itself is large enough.
The entrepreneur’s calculation was that TikTok adds approximately 1M users globally each day, 12% of whom click the plus sign in the middle, producing roughly 120,000 new video creators daily. No matter how grassroots they are, these users all need production tools.
The same entrepreneur said that approximately 20 companies in video tools already have ARR above $20M. Many are unfamiliar to the public; they may solve a narrow problem or focus deeply on a single country, yet still build substantial businesses.
15. The Agent’s L3 Leap Pushes Models from Talking to Doing
In OpenAI’s L1-to-L5 path, L1 is Chatbot, L2 is reasoning and L3 is Agent. L1 to L2 remains a language problem; L3 becomes behavior, requiring the system to execute tasks in a computer, webpage, database or other system.
Zhang therefore warns that “this is the first year of Agents” does not mean maturity will arrive within one year: “Maybe the next 5 years will all be the first year of Agents.” Like autonomous driving, the category may pass through multiple stages and reliability thresholds within the same level.
General-purpose versus vertical is only the first fork. Further choices include browser versus API, using the existing internet versus rebuilding machine-readable systems, and how to manage permissions and feedback. Each combination produces a completely different company.
More importantly, Agents give non-model founders a new deployment paradigm. Startups once had to build around a model plus an interface; the behavior layer now creates room for protocols, tools, implementation, memory, security and vertical workflows, allowing the application ecosystem to truly open up.
16. Most of the Startup Value in Agents Is Beneath the Iceberg
The ecosystem has not even fully standardized its protocols, and large companies are competing for de facto standards. Vendors such as Amap and Google are building protocols, and OpenAI could lose the opportunity to early movers even if it wants to participate. Once language becomes behavior, large amounts of scaffolding are also needed to make the human-designed internet compatible.
Memory is only one example. Before interviewing Mem0, the host asked the founder of Memobase for advice and found that every company followed a different technical path. His own AI social team could not decide whether to build in-house, adopt an open-source solution or buy a mature API.
Asked about the selection criteria, the answer was that “the answer is floating in the wind.” 赵纯 decided to build a memory system first for a new IM because he believed the available market solutions were either unusable or unsuitable for his project.
Ant Group’s open-source Agent ecosystem map lays out a vast landscape spanning frameworks, Memory, Inference and more; these were also among the fastest-growing areas in the open-source ecosystem over the past year. Zhang’s metaphor is that users see only the Agent above the waterline: “Under the iceberg is an enormous system.”
17. Manus First Defined the Agent Product Mental Model, Then Proved Demand with Revenue
The core significance of Manus after its March launch was not merely technical leadership, but giving users a rough sense of “what an Agent should look like.” Becoming synonymous with a category creates a first-mover advantage that is difficult to replicate.
In response to dismissals that “a Manus can be copied in 3 days or 3 weeks,” the host stressed that successful businesses have never depended on R&D alone. Timing, details, boundaries, brand and distribution all determine the outcome. Copying the interface does not mean copying the user mental model.
Manus subsequently disclosed approximately $90M in ARR and introduced revenue-scale metrics such as RRR. Zhang believes the growth curve has surpassed that of most companies regarded as high-growth, proving that Agents have not remained a concept.
Asked whether “Agents cooled off in the summer,” he said attention had shifted: Chinese open-source models crowded the spotlight in July, while GPT-5 and new Claude versions took it in August. Agents no longer jolt people awake every day, but they have entered the phase of revenue, implementation and startup diffusion.
18. Vertical Agents, Coding and Companionship Are Forming a Multilayered Application Market
AI coding is the year’s unavoidable breakthrough, and it has already fragmented from a single keyword into front-end, back-end, databases, IDEs and other battlefields. Each link can produce an independent product; the general-purpose coding assistant is only the most visible layer.
a16z Speedrun showcased 58 startups in Los Angeles, roughly 2 minutes per company. After having someone organize the现场 recordings, Zhang concluded that the companies fell into 3 categories, one of which was Agent as a Service: executable-task Agents delivered to different vertical industries.
Legal was one of the most concentrated verticals for financing, requiring industry data, workflows, privacy and security. Finance can be split further into primary markets, secondary markets, insurance and banking, while marketing divides into search, video, images, text, email and online/offline channels.
Social and companionship rarely enter mainstream discussion. H U Z’s latest rankings had 10 companies on the web, 10 in the top 50 and 12 or 13 on apps, yet almost nobody discussed them apart from Character.AI. Zhang believes the products have already gone through 3 generations. AI anime also attracted concentrated financing, but its attention cycle “will not be particularly long, so you have to move quickly,” although he has invested in one company as well.
19. Control of Mobile Agents Is Better Held by the Operating System
End-to-end voice allows devices to understand users more naturally, while L3 behavior requires control of apps and data on the phone. Given permissions, security and privacy, Zhang believes this should be handled by the handset maker itself rather than an ordinary third-party app.
HarmonyOS’s 小艺 first interprets the intent behind a command, then uses HMAF to let each app register its own Agent API. The operating system does not need to complete every task itself; it can route the request to existing services such as Meituan, Ctrip and JD.com.
The host’s example was: “小艺小艺, recommend a rice cooker suitable for an elderly person.” 小艺 identifies the intent and invokes the JD.com Agent, which completes the task inside JD.com’s own app. The intelligent-agent platform opened in June also lets developers assemble more Agents from components.
The logic is the same as ChatGPT calling Spotify or Canva: the unified entry point handles understanding, while the specialized app handles fulfillment. China’s mature mobile internet provides a rich service layer, and the OS Agent opportunity is built on those existing capabilities.
20. Open Source Has Become a Business Model, Trust Mechanism and Global Competitive Tool
China pushed open source from the top down this year. Shanghai introduced policies encouraging companies to build open-source products and communities, with prizes of up to RMB5M. Zhang recalled Sam Altman once believing that the leading model would necessarily be closed source because it required enormous money and resources; after R1, the industry expectation became that open-source models would also appear among the leading camp. The State of AI Report even predicts that an open-source model will take first place for a period in 2026.
He further said that after GPT-5 launched, OpenAI also open-sourced several of its series. The important measures are no longer just benchmarks, but the number of adopting companies, developer scale and Token consumption: “No matter how powerful a model is, if nobody uses it, it has no meaning.”
Open source does not mean noncommercial. DeepSeek is open source while selling APIs and To B services; even if its price is only one-thirtieth or one-hundredth of OpenAI’s top model, a viable cost structure can still turn usage scale into a revenue and improvement flywheel.
Outside China and the US, markets such as Japan, Southeast Asia and Africa must decide whose models to adopt. Open source has a natural appeal through low cost, local deployment and openness, which is why Zhang calls it a “weapon” in competing for third countries.
21. A Shared Technical Language Can Still Pierce Geopolitical Division
Open source first answers concerns over whether data will leak or be repurposed by the platform. The host also mentioned that during the Singles’ Day period, a laptop with unusually large memory and VRAM, priced relatively affordably, sold out rapidly because it was naturally suited to local model deployment.
梁文锋 rarely gives interviews, but communicates with the technical world through papers and personally serves as first author. The host’s impression was that when many forces beyond anyone’s control obstruct normal communication, the technical community still retains a unique language.
After DeepSeek launched, Lex Fridman produced a roughly 5-hour episode analyzing why reinforcement learning worked. The host believed the discussion was “basically not pulled off course by a political perspective.” If a product is good enough, the technical community remains willing to evaluate it objectively and acknowledge its achievements.
22. HarmonyOS Has Reopened Spaces Sealed Off by the App Store
HarmonyOS has expanded beyond phones into laptops and cars, and the convergence of mobile devices with traditional PCs is viewed as inevitable. With pure HarmonyOS devices and HarmonyOS laptops appearing, mature mass-market apps have gradually adapted; the next stage requires a long-tail developer ecosystem as rich as a rainforest.
A Chongqing developer couple interviewed by the host runs the couples app Suki. Although the product also has versions on other platforms, it grew rapidly soon after launching on HarmonyOS and received early traffic and resource support. This is the direct first-mover dividend offered by a new ecosystem.
Another developer uses a single cross-platform, cross-device codebase to run approximately 10 apps, including Daily Cocktail, Daily Coffee and sunrise/sunset times. A new app can be launched every 2 to 4 weeks, generating approximately RMB70,000 in monthly revenue. The same keywords are already occupied on iOS and Android, while HarmonyOS still has “potholes” to fill.
23. The Problem for Chinese AI Apps Is Not Functionality, but Entry Points, Metrics and Exit Paths
Asked whether China had produced an app similar to Cursor that could define a new category, Zhang answered, “It’s quite difficult.” The US can repeatedly tell stories about reaching a certain ARR in a certain amount of time; Chinese apps more often compare user counts, while the mature mobile-internet environment makes new users extremely difficult to acquire.
The host cited QuestMobile’s first-half data: roughly three-quarters of users for standalone apps and web AI apps were declining, while only one-quarter were growing; plugins, by contrast, had growth of roughly two-thirds. The practical question that followed was: “Who is willing to just be a plugin?”
The reason is that entry points are occupied by the “castles” of WeChat, Douyin and Xiaohongshu. Even a standalone app with the product right can struggle to create distribution from zero; some so-called AI apps on the rankings may effectively be “90% traditional app plus 10% AI feature.”
Several AI image communities raised 2 or 3, even 3 or 4 rounds of financing, and their founders were not weak, yet they could not recreate Changba’s story of reaching approximately 1M DAU in a week. Even an occasional breakout today is more likely to “flash like a meteor.”
24. Dollar VC Still Believes in Innovation, but the Room to Execute Has Narrowed Sharply
In the US, an unfamiliar To B company can raise approximately $50M almost every week. China cannot directly copy the US Agent SaaS path because of SMEs’ ability and willingness to pay and the structure of closed platforms.
For dollar-focused VCs, after investing several rounds in a model company, all they can do is wait. Taking an application overseas creates another set of complex problems; staying domestic makes it difficult to answer “what exit are you expecting?” Zhang’s practical response is to set aside questions without answers and return to the founder and the product.
The underlying judgment from one fund managing partner was that if you still believe technology companies need patient capital to complete the zero-to-one journey, you should continue doing VC. The episode also invoked research associated with that year’s Nobel Prize in Economics to stress that truly rapid growth still comes from disruptive technological innovation; translating that answer into domestic execution, however, runs into a long list of constraints.
Capital has therefore shifted toward more tangible AI hardware, moving from researchers and big-tech product managers to entrepreneurs with backgrounds at DJI, Dreame, Roborock and Insta360. The logic is that software struggles to establish trust, while hardware at least offers revenue, supply chains and physical products as validation.
25. Overseas Capital Is Looking at China Again Because the Cost Gap Connects to Mature Exit Markets
The host relayed a case from the Uncapped podcast, in which Jack Altman interviewed Thrive Capital partner Wings Pancker about a team that spent 18 months researching a company. When Zhang heard it, he thought: “In China, 4 cycles have already passed in 18 months,” illustrating the completely different pace of decision-making on the 2 sides.
The partner had also visited China recently, following an extremely dense schedule and meeting many of the companies mentioned on the show. Zhang viewed this as a signal that “the world is still very interested in China,” rather than evidence that overseas capital had abandoned Chinese technology assets.
An even more unusual case was overseas VCs revisiting early-stage Chinese game teams—a sector that had been outside the normal range of domestic VC since around 2015. Their rationale was not to discuss an IPO first, but that Chinese teams were “cheaper across the board” than overseas teams while overseas game M&A remained active.
26. OpenAI Has Compressed 5 Years of Expectations into Today’s Market Prices
Zhang tracked the change in sentiment through his own PPT: the fourth deck had only 1 page on bubbles, the fifth had 2 pages, the Q3 summary in September had approximately 6 pages, and the next one would give the topic a standalone chapter. The rising page count itself is a thermometer for growing market anxiety.
Among the top “Magnificent Seven,” several companies already have market caps of $3T to $4T and roughly $100B in annual discretionary operating cash flow; OpenAI is valued at approximately $500B. At this scale, even before listing, it is large enough to influence global capital allocation through orders, equity and partnerships.
Zhang uses “Cloud Ladder Palm” to describe the full stack: Google owns cloud, models, products and infra at once. If AI becomes the foundation of everything, Google is worth $3T—so what should OpenAI be worth? Expectations that once covered 6 months or 1 year have consequently been extended into 5-year and 10-year contracts.
The contradiction is that “technology advances month by month,” and almost nobody can write a credible 5-year technology report, yet the market demands that 5 years of demand be fully priced in today. Sam discussed roughly $7T in construction demand in 2024; now, as the commitments of various companies are added together, the trillion-dollar scale has entered real-world discussion.
27. AI Has Upgraded from an Industry Theme to “the Market Itself”
After news of OpenAI’s partnership with AMD, AMD rose by approximately 40%; news of its partnership with Broadcom then pushed the latter up by approximately 10%. OpenAI’s long-term expectations flowed through partnership announcements into related companies, which immediately gained in share price.
No leading company can afford the responsibility of “not participating in AI,” so models, cloud, chips, energy, storage and capital markets have all been pulled aboard. Zhang’s summary is: “Today, AI is the market itself.”
This goes beyond the single-entity logic of Too Big to Fail. OpenAI is not merely a company, and AI is not merely an industrial chain, but a system too complex to separate. Even without any participant actively designing it, everyone’s incentives continue pushing the system in the same direction.
Zhang calls it “a huge conspiracy, but the conspiracy is not intentional”—nobody designed the bubble; scale, competition and the pressure not to fall behind rolled together until the system reached this point. By the time the industry recognized it, reversing course had become extremely difficult.
28. IPO Records, Burn Rates and Debt Risk Are All Breaking Old Reference Points
Zhang mentioned that the global record for a single-company IPO is held by Saudi Aramco. The episode cited both approximately $26.2B and $22.6B. Whichever figure is used, it is already smaller than the possible size of a single financing round for OpenAI or Anthropic.
He then offered a deliberately exaggerated generational extrapolation: Amazon burned approximately $2B before listing in the web era, while Uber burned approximately $40B in the mobile-internet era. If that figure is multiplied by 20 again, OpenAI might need to burn $800B before listing—and today that no longer seems entirely unimaginable.
The risk is not limited to equity. Meta, Google, Microsoft and Nvidia can support capex with strong cash flow, but cloud companies and Elon Musk’s xAI are relying more heavily on debt. Zhang stresses that “a bubble brought by equity is not so troublesome, but a bubble brought by debt is very troublesome,” because repayment is a hard constraint.
The fiber left behind by the internet bubble remains usable 15 to 20 years later, while today’s data-center GPUs may not be needed even 3 years from now and may be obsolete in 2. The same narrative that “infrastructure is ultimately useful” cannot eliminate mismatches in depreciation speed, residual value and debt maturity.
29. The Compute Frenzy Has Spread from GPUs to Hard Drives, Power and Cooling
As of mid-October 2025, according to the episode, the 2 best-performing companies in the S&P 500 were not Palantir, GE Vernova or gold miners, but Seagate and Western Digital. Capital had moved outward from the most direct beneficiary, Nvidia, searching layer by layer for the next data-center bottleneck.
After cloud, transformers, cooling and power were traded, the market realized that enormous quantities of data still require hard drives and storage components, whose supply is highly concentrated. Sam’s recent trip to South Korea and Japan to meet SK Hynix and Samsung was also understood by the hosts as part of the effort to secure storage supply.
The trade shows that AI infrastructure is not merely a question of GPU count. Every link can become a constraint when the entire system expands, and every link can become a secondary-market beneficiary. The problem is that bottlenecks are being priced one after another, which also means the system has less and less redundancy.
30. China’s Compute and Robotics Investment Is Pushing AI into the Physical World
The host recalled that tickets to the 2025 World Artificial Intelligence Conference and World Robot Conference were “impossible to get.” Compared with the compute narrative in Western capital markets, China’s investment is more directly connected to national-scale networks, manufacturing capacity, robotics applications and developer ecosystems.
Citing AI-organized materials, she said the “East Data, West Computing” project began in 2022 and had built an intelligent-compute network exceeding 30B FLOPS in 2025, using the analogy of “300 trillion phones each computing 100 times per second for 1 second.” These figures are retained as stated in the episode.
China has been the world’s largest industrial-robot market for the past 12 years. The episode cited the World Robotics Report as saying that in 2024, China ranked first in both new installations and operational stock; domestic vendors’ share rose from approximately 28% 10 years ago to 57%, surpassing foreign suppliers for the first time.
One guest compared robot competitions to F1 in car culture: competition tests technology but also builds cultural soil. Watching children interact with robots at the conference made the host realize they would be “a generation that grows up alongside robots.” That may be the era-defining feeling of looking back on this year from a much more distant point in time.