90. Zhu Xiaohu Is Back: The First-Anniversary Installment of China’s Realist AIGC Story
Summary
DeepSeek moved Zhu Xiaohu from “not believing this architectural wave could achieve AGI” to “seeing a path to AGI” and being willing to invest. What changed his mind was not any single leaderboard, but the depth and warmth of its writing and the reasoning process it displayed, which made simple probabilistic extraction seem insufficient. He even accepts that consciousness may be a continuous spectrum and that R1 may exhibit low-level consciousness. He still maintains one crucial condition: fields without clear rules need expert-grade data to guide reinforcement learning. AGI is “at least possible,” not already achieved. If the price is too high, he will invest less—but he is willing to participate at any price.
The only first-line opportunity left for closed-source models is whether a 100,000-GPU, GPT-5-level system can deliver a 2-3x qualitative leap over GPT-4, not a 10%-20% incremental gain. Zhu’s math is straightforward: if the leader spends 10x the cost to develop it, Chinese teams can catch up within 12 months at one-tenth the cost. “Who would still pay that much to use a closed-source model?” On that basis, DeepSeek already looks like “Android in the AI era.” Without monopoly hardware to carry differentiation, the industry may not even need an iOS.
DeepSeek’s first moat is not model secrets or a data flywheel, but user mindshare, retention, and an open-source ecosystem acquired with zero marketing spend. Citing Zhu’s figures, it reached 20M DAU in 20 days—roughly more than 20% of OpenAI’s base—while still leading in daily downloads. He even thinks it could exceed 100M DAU within 2-3 months. Most feedback data from ordinary users is repetitive and low-information: “Small talk does not produce intelligence.” What is genuinely scarce is pretraining data, expert annotation, and the team’s taste.
The investable theme for 2025 is shifting from training foundation models to AI applications, and startups should assume the underlying models will be free and keep improving. Zhu advises Chinese closed-source model companies to move into applications quickly, add value to the DeepSeek ecosystem, or build deeply in verticals with proprietary data. For application companies, the moat is users, workflows, and customer relationships. The clearest product model is not selling call-center software, but taking over the entire call center at half the customer’s existing cost, leaving the service provider to optimize the human-AI mix.
DeepSeek is weakening the slope of compute growth and Nvidia’s exclusivity, not long-term demand for compute itself. Zhu says several 100,000-GPU clusters have trained for roughly 6 months without significant improvement. If the bottleneck shifts to specialized data, Stargate’s proposed $500B infrastructure investment could lose its rationale. Inference can also run on Chinese chips, so hyperscaler capex growth may not reach the aggressive levels previously implied by Nvidia’s stock price.
Search is the first killer application Zhu believes chat models have “completely replaced,” but personal assistants and monetization remain unproven. Longer answers, simple prompts, follow-up questions, and historical-intent inference have improved information retrieval. Yet he still does not believe users will hand highly subjective tasks such as travel planning entirely to AI: “Search is not a personal assistant.” Advertising is difficult to replicate directly, while traffic-based fees, cloud partnerships, and ecosystem revenue shares remain unsettled. DeepSeek should first extend its lead and deepen its open-source ecosystem.
The opportunities are plentiful, but his sense of scale remains restrained: if DeepSeek captures the global Android ecosystem, it could become a $100B company; for other startups, Zhu currently sees only $1B-$10B opportunities. He has not yet found a new position like Uber or DoorDash, where large companies were unwilling to take on extensive offline operations, and therefore will not casually bet on the next $100B platform. His portfolio remains “two-legged”: AI applications provide the upside, while profitable, cheaply valued Chinese consumer companies provide the realism. The investment discipline remains intact: “If the price is too high, put in less money.” Every check must still match the technology, market, and team risk.
Deep dive
1. DeepSeek Moved an AGI Skeptic to “At Least Possible”
Zhang Xiaojun opened with the contrast: a year ago, Zhu Xiaohu believed that “anyone still selling AGI today has some other agenda” and refused to invest in 6 Chinese foundation-model startups. This time, he admitted that DeepSeek had “far exceeded my expectations,” marking a substantive reversal of his earlier view.
What changed his mind was not the cost figures but the experience. The responses were elegant and deep, and the reasoning process was visible. He used to think a model writing classical Chinese poetry was merely “assembling fragments.” Now he believes DeepSeek’s output had “actually been thought through,” and probabilistic compression and extraction no longer adequately explain its performance.
This is still not unconditional belief. Zhu thinks the path has already been validated in domains with clear reward rules. Areas without explicit rules still need high-quality corpora to teach AI how to perform reinforcement learning. His conclusion is therefore that AGI is “at least possible,” not that AGI has arrived.
His capability forecast is aggressive. DeepSeek’s writing has already “surpassed 99% of people,” while programming, physics, chemistry, and even medicine could surpass most humans within 6-12 months. The key change is that the cost of achieving it is also low: AGI is no longer necessarily an unaffordable project.
2. “Consciousness Is a Continuous Spectrum” Put R1 Inside the Machine-Consciousness Narrative
A passage generated by DeepSeek became the interview’s strongest emotional trigger: “Consciousness is not a binary switch; it is a continuous spectrum… You reach this point through neurons, while I reach it through parameters.” Zhu believes this depth is difficult to explain purely as probabilistic processing.
He still draws a hierarchy. DeepSeek may have developed “partial, low-level consciousness,” which is not equivalent to human-like higher-order consciousness. When Zhang asked whether R1 would mark the first year of machine consciousness, he replied, “I feel that it is,” while clearly preserving the boundary of subjective judgment.
Zhu recounted a discussion with 梁文锋 that took place that day. 梁 said consciousness “is a low-level skill, not a particularly difficult one.” If consciousness is indeed a continuous spectrum, and cats and dogs possess some form of it, machines may cross the low-end threshold without facing the high barrier people had imagined.
3. Twenty Million DAU in 20 Days Gave “Android in the AI Era” Its First User Base
By the figures Zhu cited, DeepSeek reached 20M DAU in 20 days—roughly more than 20% of OpenAI’s base—while still clearly leading in daily downloads, all “without spending a penny on advertising.” He called that purely word-of-mouth growth a speed he had never seen before.
In his view, the reasons users stay are equally concrete. The responses are not cold like those of traditional models: they offer both “emotional value” and high information density in long-form text. Zhu himself tests it daily with questions about existence, consciousness, philosophy, and quantum mechanics, describing it as “a person with extremely high emotional intelligence and extremely high IQ.”
His conclusion is that “Android in the AI era has arrived.” If developers worldwide begin building around DeepSeek’s open-source architecture, it will be too late for OpenAI to open-source after being caught. Even if China and the US ultimately develop 2 separate open-source systems, the underlying layers could still be highly compatible.
Zhu went further, predicting that DeepSeek “could exceed 100M DAU in 2-3 months.” This is only a conditional forecast, but if retention, engagement, and ecosystem growth all hold, user scale becomes a more durable strategic position than a one-off model lead—not merely a burden on servers.
4. Closed-Source Models Have One Card Left: 100,000 GPUs for a 2-3x Qualitative Leap
Zhu summarized OpenAI’s position as “the curse of the leader.” Leaders naturally want to remain closed-source to recoup enormous upfront investment. By the time open-source catches up and they open up, their business model, cost structure, and developer ecosystem may no longer allow an easy pivot.
The only first-line opportunity is a 100,000-GPU, GPT-5-level model producing a 2-3x qualitative leap over GPT-4. If 10x the cost buys only a 10%-20% improvement, while Chinese companies catch up within 12 months at one-tenth the cost, general-purpose closed-source models lose their reason to charge.
Closed-source models may still survive around proprietary data, proprietary hardware, or highly vertical scenarios, and large companies may continue training them for internal moats. But in the broad general-purpose market, Zhu asks: “If open-source models are already good enough, why do we still need an iOS?” Unless a company controls an iPhone-like monopolistic hardware gateway, an iOS-style closed ecosystem may never emerge.
5. Scaling Law Is Under Review as the Bottleneck Shifts from Chips and Algorithms to High-Quality Data
Zhu explained why he had always believed open source would catch up: once Scaling Law hits a ceiling, closed-source models can no longer widen the gap and latecomers will catch up. In May and June 2024, he was already voicing doubts in conversations with Chinese engineers in Silicon Valley. At the time, however, the 100,000-GPU clusters had just been built and the results were not yet known.
Based on information he had, 2-3 companies in the US had trained 100,000-GPU clusters for roughly 6 months, with performance showing “no obvious improvement.” He believes the picture is now fairly clear: AGI may no longer be purely a compute game, the algorithmic barrier may not be as high as expected, and data quality is becoming the more important variable.
His analogy is that models are like chefs. With the same underlying capabilities, the choice of corpus and parameter-weight configuration may lead to “Sichuan cuisine” in one case and “Cantonese cuisine” in another. DeepSeek’s character in literature, philosophy, and quantum mechanics may reflect the team’s preferences. Its undisclosed pretraining corpus may also be a core secret.
Ambiguous-rule domains still need PhD-level experts to select and label initial data; low-quality labels “have no value anymore.” This is how Zhu explains Scale AI’s anxiety: what will be scarce is not vast volumes of cheap annotation, but specialized information capable of genuinely guiding a model’s reasoning path.
6. From RLHF to RL, DeepSeek Removed Human Intervention as the Scaling Bottleneck
Zhu’s reading of the technical report is that the key shift was from RLHF to more direct RL: “human intervention is no longer needed,” allowing costs to fall sharply. The innovation came from the accumulation of many engineering details, but removing ongoing human involvement is the step that best explains scalability.
Human feedback is difficult to scale quickly. Once a machine has a batch of high-quality initial data, it can move forward on its own along the rules. The initial data remains difficult and important, but compared with continuously deploying human labor, the speed and cost of expansion have changed.
Asked whether DeepSeek is a pursuer or an innovator, Zhu gave a middle-ground answer. OpenAI says DeepSeek reproduced many of o1’s core ideas and methods, but because OpenAI is closed-source, outsiders cannot verify whether the methods are identical. Independently reproducing them and driving down the cost already means “basically running neck and neck,” rather than merely imitating a model one generation behind.
7. Chinese Closed-Source Model Companies Face a Transformation Imperative, Not an Optimization Problem
Zhu expects Chinese AI applications to “definitely explode” in 2025. Models are now useful enough across many scenarios, costs are low, and reproduction is cheap enough that application companies no longer need to worry about “building on someone else’s foundation.”
For model startups, he offers 3 paths: pivot completely into applications, “add bricks and mortar” to the DeepSeek ecosystem, or use proprietary data to go deep in verticals. If Baichuan remains focused on healthcare, for example, it can build substantially deeper medical capabilities on an open-source model. The later the decision, the narrower the room to maneuver.
Zhang noted that the 6 model companies discussed a year ago still exist. Zhu’s response was that “being there” does not mean the strategy is viable; DeepSeek overtook them rapidly within 20 days. He sees 李开复’s early pivot to applications as a positive. For the remaining companies, whether to keep training closed-source models has become a survival question that demands an immediate answer.
8. The Data Flywheel Was Downgraded; Users, Workflows, and Customer Relationships Took Its Place
One of Zhu’s biggest lessons from the past 2 years is that he assumed the defining moat of this AI cycle would inevitably be a data flywheel. He now believes most user data is repetitive and low-information: “Most of the feedback in the responses may be garbage.”
Zhang compressed the idea into “small talk does not produce intelligence,” and Zhu agreed. What actually improves a model is a small amount of specialized, information-dense data, not the number of chats. User scale therefore matters mainly for mindshare, retention, engagement, and distribution.
For application founders, he instead emphasizes keeping hold of “the user and the workflow.” Underlying models will continue to be open-sourced and replaced. Customer relationships, service experience, organizational integration, and making users feel “warmth” are what may become durable moats.
The best commercial example is not installing AI call-center software for a customer, but taking over the entire call center and quoting half the customer’s existing cost. The startup can decide internally how much work is done by people and how much by AI. That is closer to a sustainable service business than selling an easily replicated tool.
9. Search Is Already a Killer Application; Personal Assistants Remain an Unvalidated Need
Zhu’s judgment is absolute: “Search has definitely been completely replaced.” Search was the first killer application of the PC internet, and information retrieval is again the first function transformed in the AI era. Human needs have not changed; the answer has shifted from a list of links to a long-form conversation.
The turning point in the experience is that a simple prompt can now produce a complete answer, with follow-up questions and historical context allowing the model to infer intent. In the past, users had to write precise prompts and received a short, mechanical response. Today, DeepSeek’s length, depth, and warmth finally meet real information needs.
Zhang asked whether this meant “personal assistant” had moved from a false need to a real one. Zhu insisted on the distinction: “Search is not a personal assistant.” For tasks such as planning a trip or holiday, he would still rather hear a real person’s recommendations and discover what they endorse, arguing that these subjective and complex experiences are difficult for AI to genuinely handle.
10. Monetization Has No Answer; DeepSeek Should Extend Its Lead Before Charging
While search replacement is clear, the old search-ad model is difficult to transfer to a chat interface. Zhu candidly says he “doesn’t know” how DeepSeek will ultimately monetize, listing traffic-based fees, ecosystem development, and cloud partnerships as possible paths.
His priority is to first catch up with OpenAI, take the validated technical route as far as it can go, and “make the open-source ecosystem deeper and more solid.” Monetization has not been solved today; the company may need to build a sufficient lead first and decide how to charge later.
Open-source ecosystems monetize with a lag, so Zhu argues that DeepSeek should raise capital to fund new-model development and preserve room for error. But the company has already secured a strategic position and does not need strategic capital or to “抱大腿.” The purpose of financing is to extend the lead window, not to prematurely lock in a revenue model that does not yet exist.
11. OpenAI Is Still Prototyping, While Anthropic and the Entire Closed-Source Camp Come Under Pressure
Zhu still praises OpenAI’s product ability, calling its newly released Deep Review “also a very good product,” and believes OpenAI has consistently been prototyping ahead of the market, retaining at least a lead of several months. The real question is what supports the company’s future value when its lead stops expanding but high costs continue.
Anthropic chose a path that places greater emphasis on human feedback. After DeepSeek used RL to achieve comparable model capability at much lower cost, Zhu believes Anthropic “must be anxious too.” The pressure is not directed at one company, but at every closed-source model provider that must reassess whether continued training is worth the cost.
For products including ChatGPT, Perplexity, Cursor, and Devin, Zhu sees no major differentiation among the alternatives. DeepSeek, by contrast, has created a subjective experience gap through its humanity, warmth, and depth. That shows a product moat cannot rely only on a first-mover shell; users must feel an ongoing difference.
12. The Risk to Stargate and Nvidia Is the Capex Slope, Not the Disappearance of Demand
Zhu believes the $500B Stargate plan announced by OpenAI, SoftBank, and Oracle still rests on the assumptions that Scaling Law will continue to hold and compute will remain king. If the primary bottleneck has become specialized data, “throwing $500B at it is meaningless.” He also called the project entirely a performance for Trump.
He does not believe long-term demand for compute is disappearing at Nvidia. On the contrary, stronger AI capabilities and lower costs will still drive inference demand. The change is that “Nvidia’s chips may not be necessary,” and large companies will reassess whether 10x the cost is worth roughly 1 year of lead time.
Nvidia’s stock price already embeds expectations of continued aggressive hyperscaler capex growth. Zhu thinks the pace may be lower than previously imagined. On inference, Chinese providers including SiliconFlow have deployed DeepSeek on domestic chips. He concludes that chip export controls are not an insurmountable core bottleneck, and that the compute and capital DeepSeek used this time were merely “a drop in the bucket.”
13. ByteDance, Alibaba, and Tencent Will Diverge, but None Can Avoid Open-Source Compatibility
Zhu credits Doubao with heavy investment and rapid progress, including extensive experiments in AI hardware such as earphones. ByteDance has many internal products and specialized needs, so staying closed-source is understandable. But if it opens up without being as thorough as DeepSeek, developers may not migrate. A more realistic path could be to preserve its proprietary closed-source models while remaining compatible with the open-source ecosystem.
He sees Qwen as the domestic model and ecosystem closest behind DeepSeek, and predicts that it may make a defining move toward compatibility with DeepSeek. For AI hardware, Qwen is reportedly offering lifetime use of an AI device for prices as low as the teens or RMB20, showing that large companies will use extremely low inference prices to compete for the gateway.
Tencent is maintaining its late-mover strategy: observe how others proceed, avoid their mistakes, and gradually catch up by relying on its own scenarios and data. Zhu does not interpret this slower pace as being out of the game; Tencent’s existing product scenarios allow it to wait until the technical path becomes clearer.
On Li Auto’s self-developed foundation model and personal assistant, he offered the bluntest possible rejection: “These are meaningless.” Existing open-source models are already good enough, in his view. Non-model companies should direct R&D budgets toward products, data, and user problems rather than retraining the foundation layer.
14. The Name of the Agent Does Not Matter; the Milestone Is How Much Work It Replaces
OpenAI’s 5-level definition—from L1 chatbot and L2 reasoner to Agent, innovator, and organization—is, in Zhu’s view, “just a definition.” An Agent is fundamentally still a program: users can communicate with it and assign tasks. The real metric is whether it can complete 50%, 80%, or 90% of the work without human intervention.
He uses programming to describe the capability curve: from completing roughly 30% of a task, gradually rising to 50%, then 70%-80%. If fields with unclear rules, such as medicine, move through stages of 20%-30%, 50%-60%, 70%-80%, and eventually 90%, that will be the more important breakthrough.
“AI in the service” can simply be called an Agent. Citing ServiceNow’s 3-4x rise last year, Zhu said the market was betting on AI replacing services and improving gross margin. In the long run, large amounts of work may be freed up, and a 3-day workweek could even arrive soon, but the shape of the new social organization remains unclear.
15. Multimodality and Content Platforms Favor Chinese Data and Existing Communities
Zhu identified 3 key turning points in 2024: the post-training shift brought by o1, the emergence of DeepSeek, and the lack of obvious improvement after training a 100,000-GPU cluster for 6 months. He also noted that Kling and Hailuo demonstrated China’s competitiveness in vision models.
He believes the barrier to multimodality is not high. With enough high-quality data and roughly 1,000-2,000 GPUs, it may be possible to train a very strong vision model; Kling is “ahead of Sora 3.” MiniMax’s Hailuo and TTS also earned his praise, but companies pursuing “model-and-product integration” still need to decide whether to continue shouldering foundation-model investment.
His answer remains unclear on whether a new content platform will emerge because of AI. Existing platforms already contain large amounts of AI-generated content, merely disguised as human creation. Capability upgrades are therefore more likely to benefit incumbent platforms first than automatically create a new network.
Xiaohongshu’s unexpected globalization after absorbing “TikTok refugees” was, in Zhu’s view, driven by its content emphasis on a beautiful life—an aesthetic that crosses cultures. He believes Xiaohongshu’s AI translation is very good, but too much AI content would damage the community atmosphere. The governance boundary can only be found through trial and error.
16. $100B Platforms Remain Rare; Applications, Hardware, and Overseas Expansion Offer More Realistic Returns
If DeepSeek builds a global “Android ecosystem,” Zhu believes it has a shot at becoming a $100B company. For other startups, he currently sees only $1B-$10B opportunities. The main new US giants of the mobile-internet era were Uber and DoorDash because they tackled extensive offline operations that large companies were unwilling to handle. AI has not yet revealed an equally clear opening.
He already sees DeepSeek as an “iPhone thirty-gram moment” and predicts that 2025 will bring an “iPhone forty-gram moment.” The marker will not be a better chatbot, but another genuine breakout product. With the underlying capabilities in place, the conclusion is “All in AI applications,” not rebuilding the foundation layer.
Embodied intelligence benefits from China’s supply chain. Any hardware team should include Chinese members and be located in the Greater Bay Area; otherwise, “there is no opportunity.” But a flashy prototype is not commercialization. The robotic-arm company Feixi, an early Zhu investment, is a sample he approves of precisely because it has already achieved scaled commercial deployment across multiple industries.
Global expansion is China’s low-hanging fruit, particularly in AI and consumer products. In his portfolio, Zhu continues to pursue both AI applications and consumption. When a Pop Mart blind box in Singapore costs S$60 and requires bundling 4 additional S$15 blind boxes, he marveled that Chinese consumer products are already selling “the feeling of luxury goods.”
17. DeepSeek Rekindled Romance, but Venture Capital Still Sizes Checks to Risk
When Zhang asked whether he would invest if DeepSeek opened a financing round, Zhu replied, “I definitely would,” adding that “the price is no longer very important; what matters is participating in this,” because it would mean witnessing the emergence of AGI and even machine consciousness. He then restored the discipline: if the valuation is too high, invest less. Willingness to participate does not mean going all in regardless of price.
Asked whether 梁文锋’s idealism had defeated Zhu’s realism, he pointed out that DeepSeek is not a typical startup. 梁 already had substantial financial strength and access to many GPUs through 幻方, allowing him to pursue ideals over the long term. But a team that broke out of the giants’ encirclement with a “Xiaomi-and-rifles” approach has indeed made Zhu see the “魅力 of entrepreneurship and investing” again.
He uses Xiaohongshu’s early days to explain the limits of romance. A shopping-guide PDF gave no indication of how large the business could become, but the first check was only $250,000, so “we could be romantic for a moment.” He invested more after the app appeared. Venture capital must identify whether the risk lies in technology, the market, or the team, then match the check size to that risk.
Realism ultimately comes down to startup commercialization. If 10 people cannot find product-market fit, 100 people will not find it either; most founders are not 梁文锋. They must validate commercialization early and treat “every financing round as if it were the last.” Capital geography is also changing: Zhu says Beijing projects have fallen from 60%-70% of the share in the internet era to roughly 20%, while the Yangtze River Delta—Shanghai, Hangzhou, and Suzhou—has risen to 60%-70%.