All in AI's First Three Years | A Conversation with 张津剑, Partner at Oasis Capital
Summary
- 张津剑 identifies the rightest decision of the past 3 years as announcing “All in AI” to LPs in November 2022, but says the real regret in hindsight is that “we could have gone even harder, and been even more open.” Oasis saw the paradigm shift through Stability AI and then studied Transformer; typical startup valuations at the time were just RMB100M–RMB200M, with RMB1B already considered expensive, while MiniMax completed its investment in early 2023. The costliest counterexample was Figure AI: Oasis passed on a round at roughly $800M because it believed “robotics projects must belong to China”; the company later reached a $40B valuation.
- If he could do it again, 张津剑 would not simply use money to buy commodities and equity, but would convert money into compute and then “use compute to invest in founders.” His framework is that money represented industrial-era control over commodities, while compute represents AI-era control over intelligence; as intelligence scales, compute could become scarce, quota-based and regulated like electricity. Institutions should therefore consider building their own compute clusters, while technology philanthropy could shift from giving money to providing free compute.
- Large models and embodied intelligence are not mutually exclusive tracks, but start from the “south slope” and “north slope,” move through multimodality and ultimately converge at AGI. MiniMax’s early All in MOE and multimodal push, followed by the M series, M2 and M2.1 and feedback from developers, led 张津剑 to revise his belief that a single modality could follow Transformer to AGI; on the embodied side, the key question is not whether machines can run and jump, but that “Robotics is not Embodied AI; Robotics is AI.” The real value lies in the AI trained through machines interacting with the environment.
- 张津剑 is explicitly optimistic about the 2026 fundraising market, not on sentiment but because frontier technology ultimately comes down to talent, energy and supply chains—and only China and the US have the full stack. He believes global capital has been “overallocated” to the US, leaving China as the only major destination for redeployed funds; innovations from DeepSeek, MiniMax, 宇树 and 《黑神话：悟空》 are also making Southeast Asian and Middle Eastern investors more willing to buy Chinese technology. On exits, Hong Kong is becoming a new channel for Chinese tech companies, while the RMB-heavy capital structures of 智谱, MiniMax and others are forming a domestic growth system less dependent on foreign capital.
- The faster AI changes, the more investment judgment should shift from “things” to “people,” especially by re-rating execution. Models already express themselves, think and even form a global view better than humans, but they cannot access the embodied feedback of execution; a founder’s remaining core tasks are to define the problem, hold onto first principles and act quickly. More than 40 TS in a week is high-frequency noise; the low-frequency signal remains whether the problem matters and whether PMF has been found: “In the AI era, becoming less valuable is easy; doing becomes exceptionally valuable.”
- 张津剑 sees the central question of the next decade not as a particular application, but as “finding and constructing agency” for both humans and Agents. Agents need identity, accounts, payments, authorization, anti-fraud mechanisms, and eventually legal and insurance systems; humans, meanwhile, must escape the external agency conferred by teachers, bosses and market approval. As personal Agents continuously verify language, behavior and microexpressions over time, “being real” and “living as yourself” will shift from matters of values to survival strategies: “Soon, you won’t even have the chance to pretend.”
- The next 3-5 years may be a “Great Science era,” led by a cohort of 90s- and 95s-born scientists who know how to use AI. 张津剑 also sees deep AI-Science integration, Agent infrastructure built on technologies such as x402 and ZK, and a global opportunity for China’s “north slope” of embodied intelligence. His advice to founders is a Day One “Global Vision, Global Ambition”: participate in global innovation on the basis of one’s own understanding, while living as oneself.
Deep dive
1. The 2022 All in decision determined the portfolio for the next 3 years
In October 2022, Oasis saw Stability AI and realized that “this was a paradigm shift,” then went back to study Transformer; in November, it told investors that all its focus from then on would go to AI. The LPs’ first reaction was: “Here comes another wave—we still haven’t escaped the 2015 one.”
Oasis’s deck at the time predicted GPT-4 would launch in December. What arrived was ChatGPT. The version call was wrong, but the direction was right; 张津剑 believes that for both individuals and institutions, “being heavily invested in AI, even going All in AI,” was the most important decision of the past 3 years.
The prices of 2023 later looked especially cheap: projects such as 千寻, 星海图, 宇树 and 逐际动力 commonly carried valuations of around RMB100M–RMB200M, while RMB1B was already expensive; MiniMax was first contacted at the end of 2022 and the investment closed around March 2023. 张津剑’s retrospective is not that he invested in the wrong names, but that “even as the person who went hardest, we could have gone harder.”
2. Figure AI exposed a geographic assumption as a $40B bias
Around 2023, US institutions invited Oasis to review their robotics portfolio. Figure AI was raising at a valuation of roughly $800M, struggling to fundraise and perhaps even willing to offer a discount; Oasis nevertheless concluded that “robotics projects must belong to China,” and ultimately passed. 张津剑 says the company later reached a $40B valuation.
He attributes the miss to “not being open enough”: investors should build relationships across countries, cultures and populations, and consider different asset classes—the beaten-down secondary market also deserved attention. “Even if we ultimately don’t invest, we can still explore the future of this world together.”
张津剑 views AI and blockchain as “civilization-level technologies.” As geopolitical conflict intensifies, shared technologies may instead reconnect humanity; openness in investing is therefore not just about finding more projects, but about entering more narratives about the future.
3. Compute will evolve from a means of production into control over intelligence
If he could return to 3 years ago with today’s understanding, 张津剑 would “convert all that money into compute, then use compute to invest in founders.” Money can buy commodities and equity, but not the talent that requires inspiration to assemble; compute represents control over AI intelligence in the next era.
His further projection is that compute will become increasingly scarce and could be regulated and quota-based like electricity. At that point, allocating compute would fundamentally mean allocating intelligence: a matter of safety at the small scale, and a new system of resource distribution at the large scale.
For institutions, this means building compute centers or clusters may matter more than simply holding money. For philanthropy, “real technology philanthropy” may not mean handing money to researchers, but giving them free compute to explore possibilities with no immediate commercial return.
4. The original AI thesis largely played out; the surprise was the speed of big companies’ response
Three years ago, Oasis did not believe vertical models would be the end state. It backed general-purpose large models and Agent infrastructure; the word Agent was not yet popular, and the two sides used Bot instead. The core view was that the future would no longer revolve around applications or software on phones, but would enter the Agent era.
What exceeded expectations was the ability of global giants such as Google, Microsoft and Meta to recruit talent, set strategy and invest continuously; Alibaba and ByteDance showed similar responses in China. 张津剑 acknowledges that these mainstream companies’ determined embrace of innovation has in turn capped the upside of many startups.
He still believes in AGI, but draws a clear boundary around the definition: if AGI means “an omniscient, omnipresent life form,” he does not think it is achievable; if it means outperforming humans in any environment where outcomes can be verified and scientifically measured, he believes AI will “surpass humans soon.”
5. The next protagonists may be scientists, not just entrepreneurs
张津剑 calls the next 3 years the “Great Science era,” even “an era belonging to scientists.” His metaphor is that civilization is a ball, with each scientist tearing open a fissure on its surface and expanding humanity’s frontier of knowledge along it; AI will dramatically amplify each genius’s capabilities.
The people driving this may be primarily young scientists born in the 90s and 95s. Under the old seniority system, someone born in 1995 is barely in their early 30s and may not even be entitled to join the queue; 张津剑 nevertheless expects them to take the historical stage over the next 3-5 years, reshaping the founder profile in the process.
His advice to people already in AI is not to chase a specific technology, but to “live as yourself”: AI will amplify an individual’s smallest aesthetic instincts, talents and years of practice into services for the world. What is truly scarce is finding the one thing only you can provide.
6. AI participants may become fewer, while some human labor becomes new “intangible heritage”
张津剑 believes “the number of people participating in AI is actually shrinking, not growing”: applications are broadening, but the technical and cognitive barriers are rising too. Continuing to compare humans with AI by industrial-era output standards will only create pain; people unwilling to use AI to amplify their abilities will need to find the irreplaceable texture of life in the real world.
曲凯 jokingly calls this future “PPT intangible heritage”: 10 or 20 years from now, making a PPT by hand may resemble a traditional craft today. Borrowing the Olympics analogy, 张津剑 expects future “Olympics of strategy, HR and modeling” for the brain and intellect—celebrations of humans completing cognitive work independently.
Podcasts may follow the same path: many programs will be generated by AI, and listeners may not be able to tell the difference, but a single pause from a real person may still convey “the feeling of a living human.” The future is not Agents eliminating people, but people who use Agents to serve the world coexisting with those who choose to live directly as themselves.
7. The “AI bubble” obscures the directional call; the debate itself is noise
张津剑 summarizes the three-stage migration of AI skeptics: first, large language models were “definitely impossible and scientifically unfounded”; once they worked, the ceiling was said to be low and the technology useful only for toys; as the ceiling kept rising, skeptics conceded the value but emphasized that the bubble was huge. He says bluntly: “I don’t like the word bubble.”
A bubble is fundamentally the distance between value and price; the two cannot remain side by side forever—sometimes price leads, sometimes value does. For him, the key question is not short-term overvaluation or undervaluation, but whether AGI remains “a problem the smartest people in humanity are solving”; if a small bubble attracts more geniuses, it may actually accelerate the solution.
曲凯 adds his own lesson: in 2023, he consulted professors at several top US universities. The more senior they were, the more emphatically they declared embodied intelligence distant and infeasible, causing the team to miss that wave of projects; one of those professors later started a company in embodied intelligence. 张津剑’s conclusion: “Reform always comes from outsiders, and innovation always happens at the edge.”
8. Once projects grow from trunks into leaves, exhausting the market is no longer possible
The AI market in 2022 and 2023 was like a newly grown tree trunk, perhaps with only 1 or 2 companies; then came 7 or 8 branches and dozens of twigs, until after 2025 it became hundreds or thousands of leaves. 张津剑 says: “You can point to a tree, but you cannot count all its leaves.”
Investment firms therefore have to decide proactively which directions not to cover and bring their attention back to the “principal contradiction.” His standard is not to cover every financing update, but to “keep your eyes on the trunk, not the leaves”: first identify the problem the industry must solve, then determine who is most likely to solve it.
9. The faster change moves, the more founders become the only relatively stable variable in investing
The growth of MiniMax, Hello Robot, 千寻 and Hypershell has unfolded amid US-China conflict, choices among different currencies in China and malicious competition from rivals—changes too numerous to exhaust. 张津剑 therefore places greater weight on whether founders can solve problems dynamically and “face problems honestly and pragmatically.”
AI already writes emails and speaks better than people, and may even be better at thinking and forming a global view. He reduces the founder’s remaining work to “define the problem, preserve first principles and execute quickly”; above all, the embodied feedback from execution is something AI cannot obtain.
The clearest dividing line is that some people deliver results a week after hearing advice, while others are still thinking a year later. “In the AI era, becoming less valuable is easy; doing becomes exceptionally valuable.” The corresponding strategy is “Long Action, Short Thinking.”
When 曲凯 asked him to score his performance over the past 3 years, 张津剑 refused: “That is all too industrial-era.” The past is already settled; what matters is one’s current state of being and next action. If someone else took his seat, the first thing he would want them to do is make friends with more people.
10. The infrastructure of the next decade will revolve around 2 types of agency
张津剑 condenses the long-term problem into “finding agency and constructing agency.” For an Agent, agency means knowing who it is, whom it represents, how to pay, and how to prove that it has neither committed fraud nor been deceived; eventually it also requires a bank account, a driving license, and institutions such as police and judges around the Agent.
He understands today’s RL as education in the AI era: a general-purpose large model is like 9-year compulsory education, giving people a baseline of common sense; specialized training then shapes it into the capabilities required by society and specific jobs. The medical, insurance and financial systems built around AI agency are only beginning.
The human problem is the opposite: many people’s agency historically came from evaluations by teachers, bosses or people they respected. They needed another subject to validate them and therefore never possessed genuine agency. Once functional work is replaced by AI, humans can answer “what makes a human human” only through lived practice—in other words, by returning repeatedly to “living as yourself.”
11. Fundraising is high-frequency noise; the problem and PMF are low-frequency signals
At its peak, Oasis’s internal weekly report contained more than 40 TS in a week, compared with perhaps only a few 3 years ago. 张津剑 classifies financing news as high-frequency information; there are only 2 low-frequency signals: what problem the founder is actually solving, and whether the problem truly matters.
曲凯 observes that some founders have turned fundraising into an RL reward, organizing their actions around valuation rather than PMF and constantly asking, “Why can he raise so much?” That comparison validates neither the problem nor the product’s value.
张津剑 recalls the reversal of an investor friend: while investing, the friend constantly told founders not to compare valuations; after incubating a project himself, he immediately asked, “Why is someone else at this valuation while I’m only at this valuation?” 张津剑 could only tell him to “become an investor again,” then repeat his own argument back to him—“knowing is easy; doing is hard.”
The “Signal and Noise” podcast also serves agency: it does not impose professional answers, but lets people who perform well in the industry speak naturally, allowing listeners to discover, “So-and-so is actually just like that.” The podcast lowers Oasis’s trust cost when meeting founders, while letting 张津剑 listen back to himself like a director reviewing a character; 曲凯 is more comfortable playing the supporting role and no longer rushes to prove himself through external outcomes.
12. Hong Kong is forming a domestic growth and exit system for Chinese technology companies
张津剑 attributes the biggest capital-market change of the past year to Hong Kong. The market remains immature in some respects, but is becoming a new outlet for Chinese technology companies amid US-China rivalry. He calls it an “unprecedented bull market” and hopes policy guidance can turn it into a long-lasting bull market.
Ownership structures have changed as well. 智谱’s investors are predominantly RMB institutions, and MiniMax likewise has government funds and multiple forms of RMB capital; 宇树, 银河通用, 千寻 and 逐际动力 show similar patterns, rather than the “all-dollar” structure once assumed.
This shift is building China’s own system for the growth, valuation and capitalization of technology companies, with less dependence on external funds. 张津剑 acknowledges that the venture-capital exit market is still young and has many problems, but says: “These problems can be solved. The direction is right.”
13. MiniMax replaced the single-modality assumption with a multimodal view
When he first encountered large models, 张津剑 saw Transformer as “the antidote,” believing data could follow a single route to a large model and then to AGI. His revision 3 years later is that a single modality cannot reach the endpoint; as AI develops, it increasingly understands the value of the human body itself.
Humans are not models that assess the world through language alone. Temperature, humidity, magnetic fields, frequency, skin sensation, vision and hearing jointly generate familiarity, discomfort, intuition and inspiration; these judgments transcend text and pure logical reasoning, forming a “more transparent and sensitive body.”
张津剑 connects this to MiniMax’s progress: it went All in on MOE models early and embraced multimodality; large language models subsequently developed the M series, especially M2 and M2.1, receiving real positive feedback from developers globally. He believes long-term training on complex modalities helped the model’s understanding, but presents this as his explanation rather than a claim of single-cause certainty.
His order-of-magnitude analogy is that the bits humans can express amount to roughly one-millionth of the data they can perceive, while what humans can perceive may itself be only one ten-thousandth of the world’s real data. If robots possessed a fuller spectrum and could perceive all data, the intelligence of the resulting AI would exceed current imagination.
14. The center of value in embodied intelligence is shifting from machines to AI
The first phase of the market understood embodied intelligence as robots, focusing on whether they could enter homes, screw in bolts or complete specific tasks. Commercial cleaning, warehousing, autonomous forklifts and engineering equipment around 2017 and 2018 had already proved that AI-enabled machines could be good businesses.
The second phase began treating robots as carriers of AI: next-generation software requires next-generation hardware, and AI needs a body to execute in the physical world. 张津剑 believes the market is entering a third phase: “Robotics is not Embodied AI; Robotics is AI.” The true value is not the machine carrying AI, but the AI trained through real-world RL and interaction between machines and their environments.
Large models need embodied systems to provide real feedback and more modalities, while embodied systems need AI to understand the data collected by the body. The two sides enter the multimodal route from different starting points and ultimately converge at AGI. “South slope” and “north slope” are therefore not track categories, but 2 ways of building a world model.
15. 千寻’s valuation change records the market’s shift toward the third embodied-intelligence phase
Oasis became the first investor in 极壳, 千寻 and other companies in early 2023. The underlying logic was that “next-generation software must have next-generation hardware.” But only after growing alongside 千寻 did the team gradually accept founder 高阳’s view that the real value might be the AI trained through machines and leading toward AGI.
When Oasis invested in 千寻, the valuation was around RMB250M, followed by a round at RMB300M–RMB350M. By early 2024, even RMB300M–RMB400M was extremely difficult to raise. Investors repeatedly asked when the robot would run, jump or throw a punch, and some even believed, “You don’t need to build AI; if you can run, that’s enough.” Looking back, 张津剑 calls that “cabbage pricing.”
As the market began accepting the third phase, related valuations rose rapidly. 张津剑 sees 千寻智能 and the teams of 苏昊 and 韩真 as important global forces advancing the field; by the “south slope, north slope” framework, 苏昊’s team may be in the global first tier of the north slope. He does not know whether 苏昊’s team or 千寻智能 will ultimately go further, but says the demos he saw in the lab were “very amazing, and may move very fast.”
16. New opportunities are concentrated in science, the institutional layer of Agents and a re-rating of Chinese assets
张津剑 wants to study not AI for Science as a slogan, but how AI and Science will combine at a deeper level and how young scientists will rewrite the organization of research and entrepreneurship. This line is connected to his Great Science thesis, but has not yet been compressed into a mature investment formula.
Another line is Agent Infra. The law between AIs will not rely on natural-language contracts, but on coding jointly executed by both parties; the transaction system must solve payments, identity and trusted verification. He cites the x402 protocol launched by Coinbase last year and ZK technology, arguing that many foundational problems cannot be solved through human supervision or simply by having another AI supervise them, but require mathematics and algorithms.
On the 2026 fundraising market, 张津剑 gives a clear answer: “No problem—I’m very optimistic this year.” Frontier technology ultimately competes on talent, energy and supply chains; he believes China’s talent is competitive, while many countries cannot independently solve the full electricity and data-center supply chain, leaving the real competition concentrated between China and the US.
On capital flows, he believes global capital has been “overallocated” to the US, making China the only important destination for redeployed funds. Innovations from DeepSeek, MiniMax, 宇树 and 《黑神话：悟空》 are also making Southeast Asian and Middle Eastern investors willing to buy Chinese technology. As long as capital keeps flowing in and innovation keeps advancing, the market will move forward.
17. AI compresses personal growth into subtraction, attention and authenticity
张津剑 does not treat “vitality” as a value standard, but defines it as the extent to which someone lives as themselves. If a person is treated as a large model, continuously ingesting data that does not belong to them, performing reasoning that does not belong to them and outputting tokens that do not belong to them, the model will “collapse”; many people are not incapable, but “afraid to live as themselves.”
His concrete subtraction began around June or July last year: in principle, he stopped eating dinner, making exceptions for only 2 or 3 social meals a month. That naturally reduced socializing and alcohol, freed up his evenings and made him value the single proper meal he eats each day. He no longer treats lunch as something to “eat casually,” and his own judgment has begun to reappear in other parts of life.
The method transfers to investing: if he could invest in only one project a year and do only one thing in a lifetime, the choice would shrink from 10 to 1 or 2. The corresponding discipline is to stop acting on autopilot—not buying a book just because an authority recommends it, but reading what he genuinely wants to read, even if it appears “unproductive.”
AI research points subtraction toward attention. Transformer’s key contribution was Self-Attention; models pursue Long Context, while human attention has been compressed by short video to 15 or 20 seconds. 张津剑 believes improving the quality of life and improving model quality are the same task: feed limited compute to long-term, low-frequency and genuinely important signals.
18. Once Agents can verify everything across time, truth will become the only sustainable strategy
张津剑 reduces life’s north-star metric to “being real.” On luck, he believes most, perhaps the vast majority, may be luck; 曲凯 adds that luck is often randomness summarized after the fact, while what can truly be controlled is the direction one chooses and the people one chooses to travel with. How a seed blooms still depends on climate, temperature and soil, so the only answer is to “stay humble, choose the right people and wait for the flowers to bloom.”
He expects the personal Agent next year to evolve from a recording pen into a camera that can rotate to observe both parties: supplementing questions in real time, organizing Todo items and remembering every prior commitment, while verifying language, microexpressions and behavior against facts along the axes of time, space and the internet. Once that happens, technique will be flattened: “As long as you know you’re acting, AI will know you’re acting.”
The right partner is therefore “the person who makes you more real”: whether a spouse, cofounder or shareholder, both sides should be able to express only their first-layer thoughts without calculating second- or third-layer interests. The other path is “Fake it until you make it”—keep acting until the subconscious believes it too. Beyond that, the negative effects of sustained performance will outweigh the benefits.
MiniMax founder 闫俊杰 is his example of someone who does not perform. When a competitor raised money, he said, “I don’t look,” focusing only on AGI, his own position and the distance to it; when executives asked for a benchmark budget, he insisted the technology had not converged and that the money should first go into R&D, rather than performing for the market and investors. At a larger scale, this corresponds to Chinese founders’ “Global Vision, Global Ambition”: Hypershell, DJI and 拓竹 are no longer relying only on cost and engineering capabilities, but defining new global categories.