A one-stop guide to eight top AI companies in the US and China, and the founders behind the models

Compilation: PANews Big Client

There is no single “correct” answer for AI startups. Paper-based scientists, academic commercialization teams, quantitative engineering camps, cross-industry entrepreneurs, and operators-organizers can all find their own place in this competition.

Over the past decade, AI has transformed from cutting-edge research in labs into a core force shaping the landscape of the technology industry. Today, the world’s most watched AI companies include OpenAI, Google DeepMind, Anthropic, and xAI, as well as DeepSeek, Mysterious Darkness (Moon of the Dark Side), Zhipu AI, and MiniMax.

On the surface, they are all developing large models, but behind them are entirely different startup paths: some come from top papers and academic research, some turn university research results into industry applications, some first accumulate compute power and capital in quantitative finance, and others excel at fundraising, product development, operations, and organization—pulling the technical team together into a world-class company.

🇨🇳China: From research papers and academic labs to quantitative engineering

Mysterious Darkness (Moon of the Dark Side): Turning an industry-changing paper into a product anyone can use

Mysterious Darkness was founded in 2023. Its product Kimi is one of the most recognizable AI model applications in China, representing the Kimi K3 model, which focuses on multimodality, long context windows, programming, and deep reasoning.

The founder Yang Zhilin, born in 1992 in Shantou, Guangdong, graduated from Tsinghua University (undergraduate) and later earned a PhD in computer science from Carnegie Mellon University, studying under a researcher in natural language processing. During his PhD, he worked on widely cited later works such as Transformer-XL and XLNet. After that, he also worked on AI research at Google Brain and Meta.

Among China’s startup founders, Yang Zhilin took one of the most typical routes—first proving his judgment in academia, then turning that judgment into a product that ordinary people can directly open and use. Compared with routes like Zhipu AI that lean more toward enterprise services, Mysterious Darkness aimed at the C端 from the very beginning.

Zhipu AI: A commercialization specimen that came out of a Tsinghua lab

Zhipu AI was founded in 2019, originating from the Tsinghua University knowledge engineering lab system. Its representative technical route is the GLM series, with long-term plans across base models, enterprise services, and a model platform.

Co-founder and CEO Zhang Peng studied for his bachelor’s, master’s, and PhD at Tsinghua University. From 2005 to 2020, he devoted himself to building the knowledge engineering lab and the AMiner academic platform, without the résumé of a traditional internet tech giant.

What’s special about Zhipu AI is the starting point—it doesn’t begin from zero with a single blockbuster product. Instead, it directly converts a knowledge graph and academic network accumulated by a university lab over more than a decade into a base-model platform that can be provided as external services. This is also the biggest difference from Mysterious Darkness: one builds the platform first, the other builds the product first.

DeepSeek: Investing the money earned from quantitative trading into large models

DeepSeek was founded in 2023 and focuses on base models, training efficiency, and open weights. Representative models include DeepSeek-V3, R1, and the subsequent V4 series.

The founder Liang Wenfeng, born in 1985 in Wuchuan, Zhanjiang, Guangdong, studied electronic information engineering at Zhejiang University (undergraduate), then obtained a master’s degree in information and communications engineering. His research direction involves machine vision. He did not have experience at major companies like Huawei, Tencent, or Baidu. Instead, during his graduate studies, he threw himself into machine learning and quantitative trading, and then co-founded Huanfang Quant.

Liang Wenfeng’s path is the most unusual among these companies: he didn’t first become a researcher and then look for money. He first built capital and compute power through quantitative investment, and then turned around to put these resources into large-model R&D. In a sense, DeepSeek is a product of “compute power earned through finance, fed into AI.”

MiniMax: From big-tech technology executives to multimodal startup builders

MiniMax was founded in 2022. It is a multimodal foundation model company covering text, speech, images, and video. Products include Lihulu AI, Xingye, and more.

The founder Yan Junjie, who holds a bachelor’s degree in mathematics from Southeast University, earned a PhD in artificial intelligence from the Institute of Automation, Chinese Academy of Sciences, and completed postdoctoral research at Tsinghua University. Before starting his business, he worked at SenseTime for more than six years, reaching roles including Vice President and Deputy Director of the Research Institute.

Yan Junjie combines two experiences—solid academic training, and hands-on experience managing products and leading teams in a large-scale AI organization like SenseTime. MiniMax chose the harder path of multimodality, and to some extent, that choice naturally came from the combination of these two experiences.

🇺🇸United States: From foundational research and security governance to cross-industry resource integration

Anthropic: A brother-sister duo, half physicists, half governance experts

Anthropic was founded in 2021 by a group of researchers who left OpenAI. Its core philosophy is that “AI safety should be advanced in parallel with capability development.” Its representative products include the Claude series.

Co-founder and CEO Dario Amodei earned his bachelor’s degree in physics at Stanford University, then shifted to biophysics for his PhD at Princeton University. This interdisciplinary background later continued into his research path—he first joined Google Brain as a senior research scientist, then moved to OpenAI as Research Vice President, participating in the training of GPT-2 and GPT-3. He was also one of the core participants in early RLHF (reinforcement learning from human feedback) research. In 2021, he and multiple OpenAI colleagues including his sister Daniela left OpenAI and founded Anthropic.

Daniela’s path was completely different. She studied English literature as an undergraduate. After graduation, she didn’t enter the tech industry directly. Instead, she worked in operations at Stripe and OpenAI in sequence. At Anthropic, her role wasn’t training models, but building the underlying systems that allow the company to operate safely and stably—everything from hiring and corporate governance to the implementation and execution of safety policies.

To some extent, this brother-sister combination is a microcosm of Anthropic: it needs researchers like Dario who can judge technical routes and understand the boundaries of model capabilities, and it also needs people like Daniela who understand organizational structures, processes, and risk control.

OpenAI: One who tells stories, and one who understands technology

OpenAI was founded in 2015. It released iconic products like ChatGPT, GPT, and Sora, and is the core company that brought generative AI into the public’s view.

After studying computer science at Stanford University for two years, Sam Altman dropped out and founded the location-based social company Loopt, then later became President of Y Combinator. His strength has never been publishing top-tier academic papers, but rather fundraising, product judgment, mobilizing an organization, and telling a technical narrative well.

Ilya Sutskever is a different type. He has a PhD in mathematics and computer science from the University of Toronto. He studied under deep learning pioneer Geoffrey Hinton. He participated in key breakthroughs at Google Brain such as AlexNet and Seq2Seq, and he led OpenAI’s early scaled training route.

If Ilya helped determine what kind of models OpenAI could build, then Sam determined whether the company could survive, get funding, and sell the technology. Together, the pair is almost the most successful Silicon Valley practice of the “scientist + organizer” route.

xAI: The skill set Musk used to build cars and rockets—ported to AI

xAI was founded in 2023. Its representative product is Grok. It emphasizes real-time information, reasoning, and generative media, and is deeply tied to the X platform.

Founder Elon Musk was born in Pretoria, South Africa. He studied physics and economics at university, and briefly entered a graduate program at Stanford University before dropping out. Unlike other AI founders, he didn’t have work experience at traditional big tech firms. Instead, he helped found companies including Zip2, PayPal, SpaceX, and Neuralink, and for a long time he served as a top executive at Tesla.

Musk’s entry into AI didn’t start from papers or labs. He integrated cross-industry resources he had accumulated—payments, electric vehicles, aerospace, satellite communications, and social platforms—then directly threw those into the compute power and data required for large-model training.

Google DeepMind: A chess genius who took a path from games to proteins

Google DeepMind was founded in 2010 and acquired by Google in 2014. Today, it is one of Google’s core AI R&D institutions, having released AlphaGo, AlphaFold, and Gemini.

Co-founder and CEO Demis Hassabis was born in London, UK. He has a first-class degree in computer science from the University of Cambridge and a PhD in cognitive neuroscience from University College London. As a teenager, he was a chess champion. At 17, he participated in developing the game “Theme Park,” and later founded his own game company, Elixir Studios.

Game training gave him intuition about complex systems and decision-making. Neuroscience provided him with a biological perspective for understanding “intelligence” itself, and computer science provided the engineering ability to turn these ideas into reality. The success of AlphaGo and AlphaFold is almost the result of stacking these three experiences—while the latter also helped him win the 2024 Nobel Prize in Chemistry.

Four founding routes: There is no single template for AI entrepreneurship

Looking back at these eight companies, the founders’ backgrounds can be roughly divided into four categories.

The first category: pure technical paper-based founders. Yang Zhilin, Yan Junjie, Dario Amodei, Ilya Sutskever, and Demis Hassabis all have systematic research training. Their common trait is that they can judge technical routes, lead research teams, and build genuine barriers in terms of model capabilities or training methods. Among them, Yang Zhilin and Ilya are closer to “writing papers that change the industry first, then moving toward entrepreneurship.” Hassabis is the product of the intersection of computer science, games, and neuroscience.

The second category: commercialization of university research成果. Zhipu AI is the most typical example, with core capabilities coming directly from many years of accumulation at Tsinghua’s knowledge engineering lab. Mysterious Darkness also has strong Tsinghua and overseas research DNA, but its path is more market-oriented—it didn’t stop at packaging academic achievements into enterprise services. Instead, it directly built a consumer-facing product.

The third category: combining mathematics, engineering, and capital. Liang Wenfeng and Musk fall into this category, but their resource sources are completely different—Liang Wenfeng builds capital and compute power through quantitative investment, while Musk mobilizes resources through cross-industry businesses like payments, automobiles, aerospace, and social platforms. Their shared trait is that they never rely solely on algorithms when entering AI.

The fourth category: non-technical organizers paired with technical founder(s). Sam Altman and Daniela Amodei represent this route. They may not be the most top-tier model researchers, but they decide whether the company can hire people, get funding, build products, and withstand regulatory pressure—and whether it can maintain cohesion during rapid expansion without falling apart.

Competition in frontier AI has never been only about “who has a PhD in computer science,” nor only about “who can train bigger models.”

What truly determines whether a company can sustain long-term leadership is whether R&D, engineering, products, capital, and organizational capability can form a closed loop: scientists determine the technical ceiling, engineering teams determine whether models can be deployed, product teams determine whether users are willing to use them, capital determines compute power and talent investment, and organization and governance determine whether the company can keep running long-term.

From this perspective, there is no single correct answer for AI entrepreneurship.

Paper-based scientists, academic commercialization teams, quantitative engineering camps, cross-industry entrepreneurs, and operator-organizers can all find their own place in this competition.

Who can integrate different capabilities into a system is more likely to become the long-term winner in the next stage of the AI wave.