Summary

MiniMax rang the bell on January 9, 2026, and its pace of capitalization—just over 4 years—may have set a historical record; in this episode’s retrospective, angel investor Chen Yu of Yunqi Capital admits that when he invested, the odds of success were extremely low, even calling it “Mission Impossible.” His explanation of the dollar-investing mindset was that early-stage investors should back ideas with a low probability of success but extreme upside if they work: “As a lottery ticket, I was willing to bet on it.” Asked whether he foresaw foundation models profoundly changing humanity, his answer was, “The honest answer is no.”
The one page in every MiniMax financing deck that never changed was the decision to build language, voice and video simultaneously—a Day One judgment that still holds. When the company bet on the MoE architecture in 2023, this was not consensus; “the entire industry only followed in 2025.” The motivation was the same as Google’s: if the product succeeds, inference compute will inevitably explode, and “saving 1% of compute means enormous savings on the bill.” Chen calls MiniMax “the foundation-model company with the best commercial sense”: with Talkie in 2023 and Hailuo in 2024, “it never directly fought the LLM monetization war.”
The investment window around a technology wave is extremely narrow: all 4 autonomous-driving unicorns that remained in the game emerged between the fourth quarter of 2016 and the first quarter of 2017, while every major foundation-model player appeared in the first 2 quarters of 2023. The inflection point came in March 2024, when Kimi raised $800M in “the round Alibaba supposedly joined at the last minute,” triggering a fund-raising frenzy at Zhipu and turning the first half of 2024 into a financing war. As an observer, Chen already thought Kimi’s decision to pour the money into user acquisition was “not very rational”: Doubao stayed free—“if others aren’t charging, why should you be able to charge?”
DeepSeek was the industry’s turning point: after V3 and R1 launched, its daily active users quickly surpassed Doubao and Kimi, which had spent heavily on acquisition, with “some numbers suddenly several times theirs despite not spending a cent on advertising,” forcing a sector-wide rethink. MiniMax reached a fork in the road: shareholders debated narrowing the company around its already-leading video model, but Yan Junjie insisted on becoming a complete all-rounder; after the M1 and M2 releases in 2025 brought M2 “across the board into the first tier,” shareholders finally relaxed. Reasoning was seen as foundational to agents and coding.
The final arbiter of differentiation is not the benchmark but user payment and “voting with their feet”: there are endless ways to game the rankings—mixing test data into training, running the same question multiple times, or changing a 20-second limit into 200 seconds of thinking. “As long as it still has promotional value, everyone will try to break the benchmark.” Yan’s approach is summarized as “laying eggs along the way”: keep shipping products and building a business over the long road to AGI. The application-layer opportunity lies in model orchestration—combining the strengths of multiple models, where “it is an art of balancing trade-offs.”
Autonomous driving offers a template for foundation-model commercialization: once “cheap” and “useful” are achieved at the same time, commercialization can explode exponentially. Neolithic’s low-speed autonomous vehicles developed slowly from 2018 to 2024, then deployed 1,000 units last year, 15,000 this year and 50,000-60,000 next year; Yuanrong expects 200,000 vehicles this year and more than 1M next year. “Foundation models will have an inflection point like this too”—OpenAI and Anthropic already generate revenue on the order of $10B, and language-model commercialization has moved far faster than autonomous driving.
One explanation for the China-US gap is the scale of capital: when an OpenAI CTO left to start a company, Thinking Machines Lab reportedly commanded a $10B valuation in its first round, while Chinese valuations remain far below that level. The host also argues that China’s constrained resources have forced companies to save money from architecture through infrastructure, producing solid results with less compute; the gap between Chinese and US models is “no more than 6 months.” Chen’s view is more categorical: “If you are investing in AI today, outside the US, China is your only option,” and he is convinced that “China’s future richest person will come from technology.”
Yunqi Capital launched its Y Transformers program this year to invest exclusively in founders aged 27 or younger—“27 is a golden line”; Google, Facebook, Microsoft and Bill Gates were all founded before their founders turned 27. The investment logic for embodied intelligence is the mirror image of autonomous driving a decade ago: demos are already highly complete, and the field is waiting for a performance and cost inflection point. But Chen is candid: “Whether that timing is this year, 5 years from now or 10 years from now, I can’t say.”

Deep dive

1. A self-described hunter, not someone waiting for lightning: top-down research before investing

Opening with Cao Xi’s description of seeing a good investment “like seeing a bolt of lightning,” Chen says his style is different: “I’m more like a hunter. When I decide to enter an industry, I study it deeply.” He first chooses the sector, conducts top-down research, and only then executes the investment.
The seed was planted at Google in 2008-09. Paul Graham visited Google to share YC’s experience, and Chen realized that venture capital sat behind everything from Intel and Cisco to Google and Facebook: “Through effective allocation of capital, you can influence the world.” He then made a checklist. He already had industry experience but lacked financial knowledge and entrepreneurial experience, so he went to the University of Chicago for an MBA, returned to China to start a company, and treated both as remedial work.
His role model is Lei Jun, whose approach left him with 2 lessons. First, Lei chose phones because “only by working in a sufficiently large industry can you build a sufficiently large company,” cementing Chen’s heavy weighting of market size. Second, Shunwei’s “go with the trend” explains why many businesses fail to get off the ground: the prerequisites simply are not ready.

2. 2 decisions in 2016: reading physical AI through national policy and a car ride

His interest in robotics was triggered when China revised the family-planning policy it had implemented for more than 20 years. “The government doesn’t change a national policy for no reason.” He looked through population yearbooks and found that the decline in fertility and the aging trend were already pronounced. The technological answers were limited to 2: reproductive technology, or robots to supplement the labor force.
His interest in autonomous driving came from a business trip to Pittsburgh, where he was stunned by riding in a driverless Uber. He had already known about Waymo and Baidu Apollo—the latter was led by his former boss at Google, Wang Jin—but “none of that information was as powerful as trying it yourself.” He concluded that autonomous driving “was no longer just a demo,” and that China’s market, with 20M passenger vehicles sold annually, had enormous potential.

3. 3 criteria for backing people, and the judgment Google instilled

He sums up his investment style in one sentence: “I like investing in young people who are smart, honest and ambitious.” Intelligence raises the probability of execution, ambition ensures the outcome is large enough, and honesty has 2 dimensions: founders must assess every decision objectively rather than warp themselves around short-term interests, while investors can only entrust them with capital if they do.
Google—where his employee number was around 10,000—gave him a muscle memory for identifying exceptional engineers. He does not need to understand the technology itself. By looking at how someone breaks down a complex problem, whether they can connect a grand end-state vision with scale and cost, and how they link it to a business model, “you can roughly judge whether they have the capability.”
Asked about the idea that “once you leave Mount Wu, no other mountain looks like a cloud,” he disagrees. There are “actually many people like Jeff Dean.” Geniuses may make up 1 in 10,000 or 1 in 100,000, but the world’s population is large enough; “what is missing is the eye to find them.” The host then said that 60%-70% of researchers and engineers working on AI globally are Chinese or ethnic Chinese. That was the host’s view, not a conclusion Chen offered here.

4. Baidu saw it early, but struggled to execute

In an interview with Luo Yonghao, Yan Junjie argued that Scaling Law was first discovered in 2017 at Baidu’s Silicon Valley AI Lab, but OpenAI ultimately turned it into GPT-3.5 and ChatGPT. Chen adds to the “greatness and weirdness” of Baidu: it founded its Institute of Deep Learning in 2013 and launched Apollo in 2014; even Dario from Anthropic once worked at Baidu on speech recognition and discovered Scaling Law there. Baidu “discovered things early every time,” but failed to carry its strategic direction through, and ultimately did not become a truly great company.
His diagnosis is not a talent problem but an organizational-objective problem: “Is the goal to deliver a pretty report to Wall Street every quarter, or do I genuinely want to climb to the summit, like OpenAI, and ultimately achieve AGI?” An organization planned quarter by quarter is structurally incapable of solving long-term, unknown problems, and top talent does not want to spend every day on short-term work.

5. First meeting with Yan Junjie: a lottery ticket worth taking

Their first meeting lasted an entire evening and focused mainly on technical questions. Yan’s core view was that a foundation model could be the path to AGI. Rather than having one model for lane changes and another for edge cases, then coupling them into a system as in autonomous driving, the goal was to “train a foundation model that can solve every problem”—more efficient and more like the human brain. Chen’s hammer-and-nail analogy: every AI project in the past had “a nail and a custom hammer,” but nobody proposed building one “big hammer that can strike any nail in the world.”
Yan did not explain where the insight came from, but Chen “sharply recognized that this might really be an entirely new technology paradigm” and decided to invest with little hesitation, despite the very low probability of success. He attributes that to the dollar-investing mindset: early-stage bets are supposed to target ideas with low odds but extreme disruption if they work. “As a lottery ticket, I was willing to bet on it.”
The key moment of honesty came when he was asked whether he had seen the downstream impact: “My honest answer is no.” The primary market was weak in 2021, and the technology had been relatively stagnant from 2019 to 2021; investing in MiniMax was itself an experiment with a new technological variable. For an early-stage VC, “you are investing in a variable,” and what is needed is a major technological shift.

6. The Day One blueprint: 3 modalities, To C and an MoE bet 2 years early

In 2022, MiniMax made several decisions: pursue To C because the ceiling was higher, and build language, voice and video simultaneously. In the financing deck, “the page that never changed was the one saying we would do all 3 modalities. It was the most memorable and most important page,” and the company still follows it.
It made an early bet on the MoE architecture. In 2023, that was not consensus; “the entire industry only followed in 2025.” The motivation was identical to Google’s: assume the product succeeds and 10M people chat with AI simultaneously, and compute will inevitably explode. “Saving 1% of compute means enormous savings on the bill.” Under cost pressure, the company explored how to reduce compute demand while preserving as much intelligence as possible. Mistral was already working on MoE, and once Yan saw its value, he was willing to bet everything on it. The idea was simple, but the engineering details were endless.
The pressure is constant: “If your model’s performance hasn’t changed at all in 6 months, you are already far behind your peers.” The current description is that foundation-model rankings are a quarterly race run every 3 months; only by continuously training the best models can a company remain in the game.

7. The 2023 investment scramble: a narrow window and the burden of legacy systems

Chen’s window thesis is straightforward: all 4 autonomous-driving unicorns that stayed in the game appeared in the fourth quarter of 2016 and the first quarter of 2017. “Looking back, if you didn’t invest in those 2 quarters, it became very hard to get into any of the 4.” The same pattern applied to foundation models. In 2021, consensus had not yet formed; once it did in 2023, “all the players arrived in the first 2 quarters, whether they ultimately succeeded or failed.”
The story of his classmate Li Zhifei retains the details. Li called Chen for an hour every week, and Chen’s advice—which Li also agreed with—was that “this cannot be built inside the old system.” It required a professional manager to take over, discussions with legacy shareholders, and at one point Chen considered a partnership with Wang Huiwen; none of it worked out. By contrast, Yang Zhilin’s decision to leave Loop Intelligence was “exactly right”: Li was running a company about to go public and was difficult to replace, while Yang was not even CTO at Loop Intelligence. Chen repeats the lesson: it is very hard to build a world-changing project inside a “legacy system carrying historical burdens.”
There is another iron rule for technically ambitious projects: a technical person ultimately has to be the top executive. “Otherwise you have no ability to rally talent, and it is very hard to control the direction of the project.” If Li had actually left, would Chen have invested? “I might have supported him, out of support for a classmate.”

8. Why he did not invest in a second foundation-model company

His rule for repeat bets is that every company in the same sector must have enormous differentiation in technology or business model. He invested in Yuanrong and Neolix in autonomous driving because they were fundamentally different projects, focused on passenger and commercial vehicles respectively, with entirely different success factors.
Foundation-model technology paths are “highly convergent”: the differences are mostly whether a company adopted MoE earlier or later, while everyone is now studying Efficient Attention. MiniMax was already the all-rounder building across modalities, so “the odds of investing in another similar company were not very high.” He is still studying competitors, however: “What if they really do have something unique?”

9. From Talkie to Kimi’s $800M: the financing war and the fork in the spending road

MiniMax began building social products in the second half of 2023. Starting with Glow at the end of 2022 and continuing through Glow, Xingye and Talkie, it built a full social-product suite that “was the first time everyone saw how foundation models could monetize.” By year-end, the market had begun questioning how companies with expensive base models could command such high valuations when there was no visible path to profits. MiniMax’s revenue was already approaching $100M at the time, which softened that criticism.
The inflection point came in March 2024. Just after the Lunar New Year, Chen received a call saying Kimi would soon secure a large financing round: $800M in “the round Alibaba supposedly joined at the last minute.” It immediately changed the market structure and triggered a financing frenzy at Zhipu, turning the first half of 2024 into a financing war.
The companies diverged on how to spend the money. Kimi invested heavily in user acquisition, which Chen already thought “wasn’t very rational” as an observer. Doubao insisted on remaining free: “If others aren’t charging, why should you be able to charge?” Large platforms had other monetization channels, while startups could only use financing to subsidize “an app with neither a business model nor retention.” MiniMax, by contrast, quickly followed the potential for video generation revealed by Sora and launched Hailuo in August, with results that were “surprisingly good.”

10. Llama 3 clears the field, ByteDance goes all in, and DeepSeek’s prehistory

The open-source release of Llama 3 in April 2024 was a major event: “While you were still talking about how to catch up with GPT, Llama 3 had already pushed open-source foundation capabilities to a new level.” It immediately forced a group of startups claiming to build foundation models out of the market. In the second half, Alibaba and ByteDance increased their strategic investment and the price war began. MiniMax’s differentiation was that while others were competing on foundation models in 2023, it was productizing language models; while others competed on LLM monetization in 2024, it had already moved into video generation. It “never gave up on realizing the closed loop between multimodal models and monetization.”
ByteDance had approached MiniMax early and considered investing, but made a major internal decision in June 2024 to go all in on building its own models: “Investment is meaningless. Either I acquire it, or I go all in and build it myself.” Neither MiniMax nor StepFun, both under discussion at the time, could have been acquired so early, so ByteDance entered the field itself. Around the same time, High-Flyer also decided to build foundation models internally.
DeepSeek’s early history is one of the episode’s key details. High-Flyer accumulated GPUs early and at one point had the largest inventory of NVIDIA A-series cards; in 2023, it was rumored to have 10,000. Earlier, in 2022 and even during the pandemic, people from DeepSeek contacted Chen to sell a compute platform to autonomous-driving companies, and he introduced them to several. High-Flyer also approached foundation-model startups in the market under the guise of investment. Chen stresses that what follows is “an outsider’s speculation”: Liang Wenfeng may have recognized the importance of the field through those conversations and decided to build it himself; he reportedly wanted absolute control of the company and hoped to create a nonprofit organization similar to OpenAI. That may have been unacceptable to Chinese VCs, so Liang ultimately used his own money to do it.

11. DeepSeek’s shock: MiniMax’s fork in the road and M2’s validation

V3 launched in December 2024 and R1 just before the 2025 Lunar New Year. Their elite-level performance “suddenly brought everyone back to reality”: the foundation of commercialization is model capability, not marketing gimmicks. DeepSeek’s daily active users quickly surpassed Doubao and Kimi, which had spent heavily to reach the 10M-user range, with “some numbers suddenly several times theirs despite DeepSeek not spending a cent on advertising,” triggering a sector-wide rethink.
The fork in the road was immediate. Shareholders and management debated whether to put more resources into the already-leading video model and preserve its advantage, or refuse to let the language model fall behind. Yan was visibly driven: “I want to do everything well.” He wanted to become a complete all-rounder. In discussions with investors, he made 2 points: first, “I can do it”; second, he could show results within months. “His judgment has always been that I can do it.” That was a judgment based on the team’s technical level and the resources at hand, and investors could not make it for him because they were not the ones doing the work.
The result was M1 and M2 in 2025. “M2’s model reached the first tier across the board. That was when shareholders finally put down a stone.” The technical logic of the shift was that DeepSeek confirmed the importance of reasoning, which is also the foundation for later work on agents and coding. This is the value of a technically grounded top executive: making more fundamental, technology-based choices at moments of major strategic decision. As for whether MiniMax ever abandoned short-term monetization, “it wanted everything at once—and on MiniMax, that has been taken to the extreme.”

12. Silicon Valley in 2025: every giant is anxious, and Meta is seen as off the table

Responding to the view that the giants are most likely to win the foundation-model race, Chen breaks it down this way: OpenAI and Anthropic have funding on the order of $10B and are themselves giants in some sense; “Google cannot put in much more than they can.” Google has the highest talent density and the broadest product lines, so it naturally became a complete all-rounder. Anthropic is smaller and more focused; once it found coding to be a strong monetization path, it went all in. The host noted that Anthropic’s coding capability had a discontinuous lead in the middle of the year, but by the end of 2025 other companies had largely caught up.
The host’s observation was that Chinese startups benefit from lower talent costs and resource constraints that force them to save money from model architecture down to underlying infrastructure. He believes they can produce strong results with less compute, leaving the model gap with the US at “no more than 6 months.” When GPT-4o validated the multimodal route, Chen said MiniMax was actually less anxious: by 2024, everyone knew how many GPUs and how much capital were required to train a model. “As long as you have money in your pocket, you have confidence at the core.”
Chen sees Google’s internal shift whenever he visits each year. In 2023, people went into the office 1 or 2 days a week, and “30 hours of effective work in a week was already good.” In 2025, teams working on Gemini and Google Cloud averaged 80 hours a week. “Under heavy rewards, there will be brave men,” and “ever since Meta became the troublemaker, Silicon Valley compensation has risen to levels no one could have imagined.”
The verdict on Meta is severe: “Everyone already thinks Meta is no longer at the foundation-model table.” Since Llama 3, it has not released a model that satisfied the market. “Then there must be a problem with the organizational structure,” one reason Zuckerberg asked Alex Wang to reorganize the AI organization; “perhaps too many old hands are resting on their laurels and not doing the work.” As for the rumor that Meta’s new closed-source model is based on Qwen and built through post-training, Chen labels it as unconfirmed: “There are all kinds of theories. Let’s see what product it can actually deliver in 2026 that satisfies everyone.”

13. Model orchestration: the opening left for the application layer

Silicon Valley’s leading models already have distinct profiles: Gemini is “a highly cost-effective all-rounder,” OpenAI is “a companion that is very good at providing emotional value,” and Anthropic is “essential for coding.” The host added an algorithm programmer’s observation: at the algorithmic level, Gemini is actually the best at coding.
Chen’s conclusion is that “this actually gives the application layer an opportunity. Otherwise, if the models built every application, there would be no opportunity for AI applications.” The key concept at the application layer is model orchestration: companies need not rely only on their own model, but can route different problems to different models and balance performance, cost and other factors. “It is an art of balancing trade-offs.”

14. China’s game board: the 4 little dragons, the Qwen ecosystem and the unavoidable ByteDance

China’s differentiation resembles the US, but on a smaller scale. Qwen has built a particularly strong open-source ecosystem. Chen also mentioned a company that publicly acknowledged using models accessed through Qwen internally, then placing them into its own system. ByteDance’s models are not the best performers, but its products are the most innovative and reach the most consumers. Doubao’s phone Agent and intelligent-cockpit partnerships show how it has “extended the tentacles of foundation models into every aspect of people’s lives.” The host noted that Alibaba is also exploring To C through Quark. Chen believes Quark has strong people, but still lacks ByteDance’s accumulated advantages. Kuaishou’s Kling is strong in multimodality and also has its own coding model.
Among the 4 little dragons, MiniMax is the generalist, with especially strong voice capabilities. Zhipu focuses on To B and To G, but has also recognized that it needs To C to break through its ceiling. Kimi is both a chat app and a company investing heavily in Thinking and reasoning models; Chen believes that “may also have something to do with the resources of its shareholders.”
The narrower differentiation is ultimately rooted in the capital environment: an OpenAI CTO left to start a company, and Thinking Machines Lab achieved a $10B valuation in its first round, while Chinese valuations remain far below that. The host inferred that dollar investors might find Chinese AI projects attractive on a relative basis. Chen’s conclusion was direct: “If you are investing in AI today, outside the US, China is your only option.”
Faced with Alibaba Cloud and Volcengine—both of which have cloud and compute and can sell APIs cheaply—the startup response is “not to compete with them on their strengths.” Startups cannot beat cloud providers on API pricing, so they need to build products: what they sell is “not just model performance, nor the price of compute, but more of the overall experience.” To B offers a service moat. As for whether MiniMax and ByteDance will eventually go to war, his answer is that whether the product is social or model-based, “everyone ultimately votes with their feet.” In AI-companion social, Talkie “has never been at a disadvantage.”

15. Why benchmarks cannot be absolute judges—and the foundation-model mirror of L4 versus L2

Benchmarks are hard to trust completely. Putting public test data “accidentally” into the training set is equivalent to seeing the answer key in advance; running the same question multiple times or changing a 20-second time limit to 200 seconds of thinking can also raise scores. “As long as it still has promotional value, everyone will think about how to break the benchmark.” Without an absolute judge, people “vote with their feet,” because using a foundation model costs money and commercial revenue therefore becomes an important test. Yan’s philosophy is translated as “laying eggs along the way”: AGI is a long road, and “how to sustain the effort until that day matters,” an instinctive company philosophy rather than something investors taught them.
The L4-versus-L2 debate in autonomous driving will not be replayed exactly because “foundation models do not have such large differences in technical philosophy; everyone’s approach is broadly similar.” Even OpenAI, once seen as the strongest technical believer and the loudest advocate of AGI, is now doing commercial work. One reason is that pretraining has hit a bottleneck: “everyone has exhausted the high-quality text data in the world that can be used.” Expert-generated data is too slow, so attention has shifted to post-training. The second reason is capital pressure: “In the end, commercialization still has to test whether this technology is usable.”
The 2 histories nonetheless converge: after years of R&D, any technology that becomes “cheap and useful” can explode exponentially. Neolithic’s low-speed autonomous-vehicle segment developed slowly from 2018 to 2024, then the entire sector took off in 2025—1,000 units deployed last year, 15,000 this year and 50,000-60,000 next year; Yuanrong is at 200,000 units this year and expects more than 1M next year. “Foundation models will have an inflection point like this too. Once they become genuinely useful and cheap enough, revenue will rise very quickly.” OpenAI and Anthropic already generate revenue on the order of $10B, and large-language-model commercialization has moved much faster than autonomous driving.
The definition of AGI itself keeps expanding. It once meant intelligence roughly comparable to a human child; expectations now call for surpassing human intelligence, perhaps even reaching the level of “the sum of human intelligence.” His Goldbach-conjecture analogy is worth retaining: “One plus two can never become one plus one—perhaps if you follow a path like pretrained models and Transformer, you may only be able to approach AGI intelligence infinitely, but never reach AGI.” The underlying reason to believe in general-purpose agents is simple: people raised on science fiction naturally want them. “Every person like me who travels frequently wants an anywhere door like Doraemon’s.”

16. The payoff phase of AI infrastructure, the embodied-AI entry logic and the 27-year golden line

Investing in cloud computing and big data years ago was a way to build the foundation for the AI era: all inference and training run in the cloud. The best-known recent example for PingCAP is its infrastructure support for Manus. Chen considers Liu Qi’s line a classic: “In the future, 95% of databases will be created by AI.” Thousands of agents will create and destroy databases on their own every day, with the times forcing underlying technology forward. Zilliz is more directly tied to foundation models because vector databases serve as their external memory. Autonomous driving has entered genuine mass adoption; for Chen, this is the harvest phase of “commercialization exploding, revenue surging and ultimately producing capitalization.” Could MiniMax surpass PingCAP as Yunqi’s highest-return case? “That depends on whether public-market shareholders deliver.”
In response to peers who say embodied intelligence is not yet ready for VC investment, Chen points to history. The state of autonomous driving 10 years ago looked like embodied intelligence today: high-completion demos already existed, and capital was needed to improve performance and lower costs until the “cheap and useful” inflection point unlocked rapid commercialization. He preserves the hedge: “Whether that timing is this year, 5 years from now or 10 years from now, I can’t say. We have the capability and the patience to wait.” The previous generation did not make large bets on robotic arms because industrial robotics was a relatively small market. General intelligence opens a much larger market by giving robots a single brain that can perform general work like a human.
Yunqi Capital launched its Y Transformers young-founder program this year to invest exclusively in entrepreneurs aged 27 or younger. Google, Facebook, Microsoft and Bill Gates were all founded before their founders turned 27: “27 is a golden line.” The main fund has no age limit, but has set aside a dedicated pool for young founders. “This is a position: we stand with young people.” The top executive does not need to be a technical true believer in every case. Yuanrong’s L4-style unknown problems require research talent, while Neolithic must solve right-of-way access, operations, manufacturing, sales and a range of small-business needs, from wholesale-market merchants to linen transport for small hotels, requiring a more commercially oriented leader. “It is not that the talents are mutually exclusive; time is.” Asked whether a Chinese company will emerge among the trillion-dollar foundation-model companies, Chen said one certainly will, and remained convinced that China’s future richest person will come from technology rather than, as today, predominantly consumer goods and real estate.

Howard Lutnick: How America Can Hit 6% GDP Growth in 2026 Artificial Analysis: The Independent LLM Analysis House — with George Cameron and Micah Hill-Smith