Entrepreneurs
3 Curated Dialogues
张鹏
Key Views & Dialogues
Vol.233 Media, Communities and Investing: Talking with 张鹏 About Geek Park’s History and Its Unitree Investment
张鹏 frames mobile internet as “filling potholes” and AI and hard tech as “climbing uphill,” with paradigms unsettled and mental models potentially changing every six months. The greatest risk is becoming the “loading program.” Model companies first need to stay in the game; globally, there may be only 10 or perhaps 20, while embodied intelligence still lacks a converged data paradigm. The key question is whether intelligence can become real productivity.
View summary (8)
- 张鹏’s core read on the terrain of this AI startup cycle: mobile internet was “filling potholes,” while AI and hard tech are “climbing uphill.” The previous generation maxed out connectivity through smartphones: “Water flows down from the mountain; find a hole, dig a channel, and fill it—that is your value.” In this generation, the paradigm has not converged and the mountain is still growing. “The biggest risk is becoming the loading program”—the mental model can change every 6 months, so iteration speed must far exceed your imagination of the end state.
- AI should be compared with the personal-computing revolution of the PC era, not with mobile internet. “We can now mass-produce intelligence; this is a computing revolution.” The first question for model companies is whether they can stay in the game—there will not be 100 model companies globally; “I think there will be 10, perhaps even 20.” If others have “broken through the atmosphere and entered the AI-improving-AI phase” while you have not caught up, the gap could keep widening. Embodied intelligence has not even converged on a data paradigm; the bet is on “mass-producing real productivity.”
- Unitree validated Geek Park’s playbook of putting the community ahead of capital. 张鹏 spotted 王兴兴’s hand-built robot dog on a niche public account with only a few hundred views—2 years earlier, he had taken founders to MIT to see a robotic cheetah—then flew to meet him and talked over paper cups of water on broken sofas in a hallway. After the lead investor walked away, he still put money into Unitree first: “My action may have been right, but the logic may have been very foolish.” Unitree later stood on its own.
- Variable Capital began with 大众点评 founder 张涛’s 2017 wake-up call: “You invited Musk, charged RMB2,000 for admission, and that was it?” Geek Park’s review found that its early selection record was strong, so it turned recognition into action: “I had already invested in him before giving him money; giving him money converted the option into formal equity.” That is also why Geek Park does not make FA its core business: it seeks continuous convergence and investment responsibility, not a project-based conveyor belt of deals.
- The predicament of the AI 1.0 generation was that it had to carry commercial workloads while still in kindergarten. AI had begun to emerge after AlphaGo, and OpenAI was already exploring, but the prevailing value system demanded rapid monetization and project delivery, leaving little room for Frontier Lab-style exploration. Researchers were forced to shine in areas that were not their strengths. Today’s model companies—MiniMax, Kimi, DeepSeek and 智谱—are doing the same work in a “better era.”
- There will “always be bubbles, and there must be”: emergence needs the energy of waste, but how individuals use a bubble is the dividing line. The early-stage failure mode is “not necessarily starting too late, but starting too early.” Uphill projects require “cumulative certainty,” not valuations reverse-engineered from public-market comparables. A positive example is Horizon Robotics’ 余凯, who raised a large amount of money around 2020 and then anticipated a winter: he “turned other people’s misunderstanding into the correct course of action.”
- 张鹏’s advice to mobile-internet veterans: “Do not become a ‘deng’”—pursue evolution, not a switch. “What makes one person think he can dominate the field alone for 10 or 20 years? That is pure delusion.” Anthropic has plenty of “mid-career veterans”; AI-native is an adjective that can be learned, while entrepreneur is a noun that is hard to replicate. “The cost of starting something has collapsed, but the difficulty of doing it well has not fallen at all.” A warning from the other side: 张斯马克 once showed at a conference how one person could spend $15,000 in tokens over 1 week to rewrite an AI-native office, rapidly washing away pure-software assets.
- This AI cycle is also restructuring VC itself: investment is inference based on historical datasets, and the winning models of consumer internet may not fit AI, embodied intelligence or hard tech. “China’s investment circle undergoes mitosis every few years.” Strong investors leave to start their own funds, exposing an unresolved question about whether the original institution’s production relations allow the people creating the most value to share enough of it.
129. The Listing Interview with the World’s First Listed Large-Model Company, and a Conversation with Zhipu CEO 张鹏: Dare to Ask Where the Road Leads?
After its listing, Zhipu is formalizing governance while narrowing parallel experiments to improve resource efficiency and control cost and risk. AGI remains the long-term bet, with agents and new computing paradigms the near-term focus; rising cloud and To B revenue, cost optimization and public-company compliance will test the path toward break-even.
View summary (6)
- 张鹏 said early investors could not understand Zhipu’s business model at all, with some even suggesting that its valuation be cut in half during the downturn. To outside observers, Zhipu looked more like “concrete” or a Tsinghua STEM guy: smart, capable and dependable, but light on emotional appeal and buzz; open source won it credibility among developers and programmers.
- He acknowledged that he did not become CEO only after being fully prepared; what truly wears you down is confronting unfamiliar problems after commercialization and repeatedly aligning the understanding of customers, investors and the team. The company also had to clear organizational-management hurdles at 50 people, 200 people and more than 500 people; its latest challenge is public-company compliance.
- After its shareholding reform, Zhipu leaves major decisions to the board, while day-to-day business is decided collectively by professional committees. 唐杰 leads research, the chairman handles regulators, government, ministries and financing, while 张鹏 focuses on daily operations and market commercialization. The company admits that past explorations in multimodality, video and other areas were spread too broadly; going forward, it plans to narrow its focus and serialize work where appropriate to improve resource efficiency and control cost and risk.
- 张鹏 believes all 3 predictions Zhipu made in early 2025—the continued improvement of foundation models and multimodality, the rise of agents as a major direction, and internationalization—have been validated. Its only long-term bet is AGI; in the near term, the focus is on agents, new scaling laws, new computing paradigms and reinforcement learning.
- 张鹏 summed up the goal as “making machines think like humans,” while stressing that the endgame is to empower people and generate real value. He believes the team must balance grand ideals with disciplined execution; even if forced to choose, he would choose achieving AGI over building a highly profitable company, while arguing that the two are not inherently in conflict.
- He recalled the Shenzhen client contract, the phone agent handing out red envelopes and the low-key launch of GLM-4.5, concluding that real preparation means accumulating, day after day, in the right direction—not predicting the future. He hopes Zhipu’s footnote in AGI history will read: “Zhipu was a pioneer in the history of AGI.”
Vol.54 A New Cycle in the U.S.-China Contest? Giving Up Wishful Thinking May Be the Right Call — An In-Depth Look at the “America First Investment Memorandum”
The February 21 “America First Investment Policy” maps tighter controls on Chinese investment in the U.S., U.S. capital in China and U.S.-listed Chinese companies, while CFIUS could reach greenfield investment, existing assets and traditional sectors. The impact still depends on agency rules and negotiations, but potential changes to passive investment, public securities and the tax treaty could deepen pressure on cross-border capital and talent.
View summary (7)
- The “America First Investment Policy” released on February 21 is a policy-setting presidential memorandum, but it maps out how the Trump administration may simultaneously tighten Chinese investment in the U.S., U.S. investment in China, and financing for U.S.-listed Chinese companies. Most measures still require rule changes or updates by the Treasury, Commerce Department, SEC and other agencies, and Skadden expects implementation to be gradual through the end of 2025; but 庄明浩’s market judgment is more direct: “give up the fantasy” (放弃幻想), because Trump is not only preserving the Biden-era framework but seeking to expand it substantially.
- The core risk for Chinese investment in the U.S. is that CFIUS moves beyond M&A reviews into greenfield investment, existing assets and traditional strategic sectors. Healthcare, agriculture, energy, raw materials, farmland and real estate could all be covered; previously completed transactions that were never filed could face retrospective review and even forced divestiture. 张鹏’s warning is that the review standard will keep falling as companies find ways around it: “They will push you into a corner until you decide it simply won’t work, and then you leave.”
- The so-called welcome for Chinese “passive investment” leaves only a narrow door for something close to pure financial capital. Investments cannot carry voting rights, board seats, management influence or access to nonpublic technical information; combined with rules that lower the Chinese-ownership threshold from 50% to 25% or even 20%, and Tencent’s exit from the boards of Epic Games and Riot Games, “choosing sides” is becoming a hard constraint on ownership and governance rather than a political slogan.
- Reverse CFIUS could expand from advanced semiconductors, quantum technology and frontier AI into biotechnology, hypersonics, aerospace, directed energy, advanced manufacturing and areas tied to “military-civil fusion.” “Advanced manufacturing” and “military-civil fusion” are both expandable catch-all concepts, potentially covering robotics and new-energy vehicles; the bigger change is that publicly traded securities could lose their exemption, exposing U.S. holdings of Chinese stocks, index funds and related ETFs to restrictions and, in extreme cases, creating delisting pressure.
- U.S.-listed Chinese companies face four overlapping pressures—financing, audits, VIE structures and pricing power—not a single regulatory change. The Trump administration could revisit the 2022 cross-border audit arrangement, investigate VIE structures, and restrict Chinese investment by LPs such as pension funds and university endowments. Lily says support from leading pension funds and university endowments for China-focused dollar funds may already have fallen by “80% to 90%”; 庄明浩 says fundraising for smaller funds now resembles “gig work.”
- Suspending or terminating the 1984 U.S.-China tax treaty is the least-discussed measure, but potentially the most far-reaching. 张鹏 notes that eligible individuals could previously claim an exemption on roughly $130,000 of foreign income; protections against double taxation, the 183-day permanent-establishment threshold and arrangements for cross-border employees could also disappear. If companies and individuals are taxed by both countries, the most realistic response may be to shrink cross-border operations rather than optimize structures, adding another force behind the separation of capital and talent.
- The investment policy needs to be read alongside the technology and tariff wars: Trump has left room for negotiation but shown no sign of structural détente. Markets had previously turned optimistic on DeepSeek, the re-rating of Chinese assets and Trump’s invitation for Chinese EV makers to invest in the U.S.; they are now seeing reports that H20 chips could face restrictions and that the U.S. may ask the Netherlands and South Korea to stop servicing or supplying parts for China’s existing equipment. 庄明浩’s conclusion is that “the main theme of U.S.-China confrontation will not change in any way,” while 张鹏 reserves one final conditional judgment: the actual scope will depend on agency rules and the next stage of negotiations.