Luchen Tech's Yang You: AI Foundation Models Have No Intellectual Property — Only Open Source Can Go the Distance | BlueRun Ventures Headlines
A healthy AI ecosystem should have three or four thousand companies contributing to it.
Today, more than ever, people are realizing that talking about large language models means going beyond the models themselves. Data, training efficiency, and vertical industry integration are equally critical.
From day one, Luchen Tech chose a different path from other players: instead of building general-purpose LLMs, it focused on cost reduction and efficiency gains — essentially lowering training costs and accelerating training speed. BlueRun Ventures was Luchen Tech's angel investor. We believe Luchen Tech has unlimited potential in the AGI era.
Recently, AI Tech Review conducted an exclusive interview with Yang You, founder of Luchen Tech. We're sharing this article today to give you a fuller picture of the LLM landscape.
In 2021, a new tech company appeared in Beijing's Zhongguancun — Luchen Tech.
As a startup, Luchen Tech may not yet be a household name, but its founder Yang You is well known in the industry for a string of credentials: the ACM SIGHPC Doctoral Dissertation Award, the NUS Presidential Young Professorship, and Forbes Asia 30 Under 30.
Speaking of the recently buzzworthy LLMs, Yang You actually worked on training Google's BERT back in 2018, cutting training time from three days to 76 minutes. As he notes, companies are still using methods developed during that BERT training today.
Luchen Tech's founding traces back to 2020. Yang You had just earned his PhD from UC Berkeley when OpenAI released GPT-3, then the world's largest pretrained language model. It was then that the idea of building large models took shape.
Yang You told AI Tech Review: "GPT-3 didn't break through to mainstream attention, but I predicted early on that large models would be a major trend. Because of the pandemic, I was waiting for the right moment."
That moment came in 2021. In July, Yang You founded Luchen Tech as a one-person operation. The venture quickly drew investor interest: within roughly a year and a half, Luchen Tech completed three funding rounds, including investments exceeding ten million RMB from Sinovation Ventures and ZhenFund.
Notably, Luchen Tech chose its differentiated path from the start — no general-purpose LLMs, just cost reduction and efficiency. The core mission: cut training costs and speed up training.
Why this approach?
Yang You has publicly noted that over the past six years, model parameters have grown 40x every 18 months; in the past three years, that accelerated to 340x every 18 months. Hardware performance, meanwhile, only improves about 1.7x every 18 months — nowhere near keeping pace. High costs and long training cycles are the most pressing challenges facing LLM companies.
As he put it: "My core expertise is high-performance computing. In plain terms, I figure out how to train large models faster and cheaper."
Yang You's chosen path is, in essence, the most effective way to validate his research.
AI Tech Review learned that Luchen Tech's R&D currently spans three areas: first, Colossal-AI, a large model training system; second, training industry-specific models with 10 to 20 billion parameters; and third, a PaaS (Platform as a Service) platform. The Colossal-AI system already counts Fortune Global 500 and Fortune Global 2000 companies among its users.
"The current priority is developing Colossal-AI," Yang You added. "Going forward, whether it's GPT, PaLM, or any other large model, they can all be trained with Colossal-AI — because our system saves them money and time."
How much cost reduction are we talking about?
Yang You laid out the numbers: "Training GPT with the most basic approach costs roughly $10 million. With the best industry solution, you might get that down to $3 million. Our solution brings it to $1.4 million — cutting the cheapest alternative in half. These are pure optimization figures. Add convergence optimization and you could reduce further, but that affects the model."
Cost is only one dimension. What truly sets Luchen apart is its emphasis on "open source." In Yang You's view, AI has reached its current state precisely because of openness. The future battleground is ecosystem — how many people use your software, how much feedback you receive. Only with abundant feedback can you iterate and improve, which in turn attracts more users.
"A healthy AI ecosystem should have three or four thousand users or companies actively using and contributing. That collective force will inevitably outmatch any tech giant."
The following is the conversation between AI Tech Review and Yang You:
From Professor to AI Founder: Matching Technology with a Mission to Democratize Large Models
AI Tech Review: As a university professor, what made you want to start a company?
Yang You: Mainly the technology alignment. GPT came out in 2020 — GPT-1 and GPT-2 hadn't made much of an impact, and GPT-3 hadn't gone mainstream either. But when GPT-3 was released, I was already thinking about this. I had a strong intuition that large models would be a major trend. The challenge for industry adoption would be computational cost.
My core technical expertise is high-performance computing. In plain terms, I figure out how to train large models faster and cheaper. For instance, our collaboration with Google on BERT — then the best-performing model — cut training time from three days to 76 minutes. That created real value, and many people still use methods we designed back then.
I was thinking about this right after graduating in 2020. By early 2021, I felt I needed to start a company. But the pandemic hit, so I was waiting for the right opportunity.
AI Tech Review: Who initially influenced you to do this?
Yang You: Even before 2021, some VCs had approached me. Kai-Fu Lee reached out around April or May 2021. I came to Beijing to meet with them in July, and they sent me a term sheet within a week.
AI Tech Review: So it was Kai-Fu Lee who gave you the final push?
Yang You: I'd say it was our technical conviction. Even before Kai-Fu Lee, some individual angels had expressed investment interest.
AI Tech Review: Three funding rounds in 18 months — who invested first?
Yang You: Sinovation Ventures sent the first term sheet. Once ZhenFund learned Sinovation was in, they quickly followed. So we closed that round in August 2021. After our public announcement, BlueRun reached out. I met with their managing partner in September, and they sent a term sheet before the October holiday. The delay from post-holiday through New Year's was spent on our VIE structure — a lot of time lost there.
AI Tech Review: With funding secured, how did you build the team? Starting in 2020?
Yang You: Right. I was literally the only person when we started. A cohort of students from National University of Singapore had just graduated, so I invited them over, then recruited a few people from industry.
AI Tech Review: Luchen's fundraising has gone smoothly. What do you think investors saw?
Yang You: It relates to my BERT training experience. Sinovation Ventures' internal AI Institute still uses that technique today — I think that was one reason they invested. Plus, having a US PhD carries some recognition in the industry.
When Sinovation invested, we had nothing. They were betting on my personal track record and previous work. When BlueRun invested, we had just decided to build an open-source community — they were likely optimistic about that. The most recent round from Sequoia Capital came when we were showing early results, and they also believe in open source.
Three Parallel Tracks, with Passive Customer Acquisition Outpacing Active Outreach
AI Tech Review: What's Luchen's strategic positioning for large models?
Yang You: Three main areas: First, the training system — Colossal-AI. Theoretically, whether it's GPT, LLaMA, or any other large model, they can all be trained with Colossal-AI because our system saves them money and time. Second, training large models ourselves — building 10 to 20 billion parameter vertical industry models. Third, a PaaS platform that brings people who need to train large models onto our platform. These three flywheels create a positive cycle.
AI Tech Review: Where are you in this process?
Yang You: All three teams are working simultaneously. The main focus is still Colossal-AI, though we're also doing the second part — mainly helping enterprises with private deployments of large models. The third part will likely drive more commercialization in the future.
AI Tech Review: How do you specifically serve customers?
Yang You: Either they purchase our enterprise software, or they use Colossal-AI to train their own large models and we optimize their infrastructure.
AI Tech Review: How is Colossal-AI performing?
Yang You: We've tested it internally — it definitely reduces costs, and many people are already using it.
AI Tech Review: So Colossal-AI is already quite mature?
Yang You: Nothing is absolutely perfect. We upgrade our products every three to six months. To become truly stable requires time for iteration.
AI Tech Review: Specifically, if Chuan Wang or others used Colossal-AI for training, how much could costs be reduced?
Yang You: We've done the math. With the most basic approach — say Python, DDP, without optimization — training GPT costs roughly $10 million. With the best industry solution, you can get to $3 million without sacrificing performance, since it's matrix/tensor optimization rather than convergence optimization. Convergence optimization affects model accuracy. Our solution gets it to $1.4 million — cutting the cheapest alternative in half. Again, these are pure optimization figures. Add convergence optimization and you could reduce further, but that affects the model.
AI Tech Review: If the results are this good, wouldn't other companies be unable to compete?
Yang You: I don't think so. AI has no intellectual property or IP protection. GPT itself was built on Google's Transformer architecture. Keeping technology closed-source long-term is unsustainable.
I firmly believe AI has reached its current state because of openness. No one can claim their generative AI is uniquely distinctive with high barriers to entry. The future competition is about ecosystem — how many people use your software, how much feedback you receive. Only with abundant feedback can you iterate and optimize, which attracts more users.
I believe a healthy AI ecosystem should have three or four thousand users or companies using and contributing. That collective force will definitely outmatch any tech giant.
AI Tech Review: How is commercialization going?
Yang You: Quite smoothly. While PaaS isn't fully mature yet, the first part is already generating revenue. We have many Fortune Global 500 and Fortune Global 2000 clients, and several domestic startups are potential customers — Alibaba's Tongyi Qianwen, Baidu's ERNIE Bot, MiniMax may have all used Colossal-AI.
AI Tech Review: When will the PaaS product launch?
Yang You: Before August 1st.
AI Tech Review: We understand Luchen currently has more international than domestic customers?
Yang You: Two reasons: First, we've only been around a short time — the first month was mainly team building, which takes time. Second, we actually have many domestic customers. Some AI companies have established dedicated teams to research Colossal-AI. And we have many target customers — traditional automakers, pharmaceutical companies, oil companies, financial institutions.
AI Tech Review: Why target traditional industries?
Yang You: Traditional enterprises have sustained willingness to pay. The era of democratized AI represents an internal AI upgrade for traditional industries. How pervasive AI ultimately becomes depends on traditional sectors. Several automakers are already training their own models because they see this as core technology without absolute barriers. Some leading securities firms also have strong appetite for original technology.
AI Tech Review: Will you focus more domestically or internationally going forward?
Yang You: It doesn't really matter. We're a small company — no need to limit ourselves. Plus, we're building an open-source community, which is passive customer acquisition. We don't need active business development. So we have customers in the US, Middle East, Singapore, and Southeast Asia.
AI Tech Review: If you did active customer acquisition, which regions would you prioritize?
Yang You: For active acquisition: China first, Southeast Asia second, Middle East third. Passive acquisition has no geographic limits.
AI Has No Copyright — Only Open Source Enables Long-Term Success
AI Tech Review: Why is the open-source ecosystem so important to you?
Yang You: Two reasons. First, building a strong open-source community creates greater value. We're doing venture capital, but when many people use our work, it generates social value — our investors' money isn't wasted. From the investors' perspective, they can get behind this since their capital also comes from society.
Second, since we're building a company to eventually monetize and go public, I believe the core competitive advantage in B2B AI is establishing strong trust-based relationships with users. That's why open source matters.
AI Tech Review: So you decided on open source before founding Luchen?
Yang You: We decided to go open source within a month of founding.
AI Tech Review: What does the ecosystem look like now?
Yang You: Three main types of participants: deep users who contribute code and help optimize our software; companies that use our software and develop dependency; and companies that provide feedback. These include both large and small companies.
AI Tech Review: How many people are dedicated to ecosystem operations?
Yang You: We've assigned two or three people to guide it. The point of building an ecosystem is getting people to use it, helping them solve problems, then improving based on their feedback. We also set some key development directions ourselves.
AI Tech Review: So open-source ecosystem building doesn't require many people?
Yang You: Right, I don't think it would exceed 20 people, whether the community has 20,000 users or 100. You need some people to maintain the core kernel. Once that's solid, the peripheral stuff — if enough people are using it — will get spontaneous contributions.
AI Tech Review: Your partnership with NVIDIA is also ecosystem-driven?
Yang You: Yes. We're currently in NVIDIA's ecosystem, where we can potentially access discounted compute. NVIDIA has also contributed new features to our open-source community, which they prioritize适配ing to Colossal-AI.
China's LLM Landscape: Everyone Has a Shot, But Winners Will Emerge by Year-End
AI Tech Review: How do you see China's LLM market developing?
Yang You: Two main directions. Domestically, at most two or three players will break through — we'll likely see by year-end.
Ultimately, the general-purpose LLM market can accommodate at most two or three players. Major tech companies will take half, leaving perhaps one slot for a startup. This forces other startups to pivot to industry-specific models, which are less valuable than general-purpose LLMs. So most startups will see their valuations drop significantly.
AI Tech Review: Which LLM do you favor?
Yang You: The leaders are either major tech companies, or MiniMax and Zhipu AI. These few have definitely trained large models already. Others are at various stages — some have prototypes, some are still fine-tuning, some haven't even reached the training phase. You can actually tell by checking GPU usage on Volcano Engine: MiniMax and Zhipu have used 1,000 GPUs, while others are at 200.
Objectively speaking, I think Baidu may actually be the most advanced.
AI Tech Review: What about startups?
Yang You: I think it's Zhipu AI.
Several reasons. First, China and the US have different situations. Chinese AI papers generally come from universities, while in the US they come from Google, Meta, OpenAI — meaning China's technical talent source is universities, where the best AI talent resides. Second, once large models scale up, they face political considerations. USD funds will eventually face restrictions, so purely RMB-funded entities like Zhipu have an advantage. Third, Professor Jie Tang has rich academic and technical experience, and his Tsinghua background will greatly help Zhipu AI's development.
AI Tech Review: What do you see as the decisive factors for domestic LLMs?
Yang You: Data, compute, and algorithms. Compute and data are most important. How to use compute efficiently is also critical.
Originating in Silicon Valley, BlueRun Ventures was established in 2005 as a venture capital firm focused on early-stage startups.
Currently, BlueRun Ventures manages multiple USD and RMB dual-currency funds in China, with assets under management exceeding RMB 15 billion, making it one of the largest early-stage funds domestically. The firm invests primarily at Pre-A and Series A stages, covering hard tech and innovative interaction, enterprise technology, new consumer, and healthcare sectors. It has cumulatively invested in over 150 startups, including Li Auto, Waterdrop, QingCloud, Guazi, Qudian, Songguo Mobility, Ganji.com, Energy Monster, Yuntu Semiconductor, Machenike, Cloud Saint Intelligence, Anxin Network Shield, and BioMap.
BlueRun Ventures has been ranked #1 on Zero2IPO's "China Top 30 Early-Stage Investment Institutions" and ChinaVenture's "China Best Early-Stage Venture Capital Institutions TOP30," and was named among Preqin's Top 10 Global VC Fund Managers for Sustained High Returns.
Additionally, BlueRun Ventures has repeatedly received honors from Forbes China, 36Kr, Cyzone, Caixin Media, CBNweekly, Jiemian, and other media outlets, including "China's Best Early-Stage Firm of the Year," "China Top Venture Capital Firm," "Most Entrepreneurur-Friendly Early-Stage Firm of the Year," and "Most Influential Early-Stage Firm of the Year."