Legend Capital's Song Chunyu: China's AI Decade — Missed Opportunities and Missteps

By Lili Yu

Inside Lenovo's sprawling headquarters, Song Chunyu's office stands out as something of an oddity. The cramped ten-square-meter space is packed floor-to-ceiling with curiosities from his travels: espresso machines, aviation models, murals, porcelain — and several oversized plush toys that clash rather dramatically with the hard-nosed industrial investor persona. A Lenovo veteran of more than two decades, Song has backed an impressive roster of names: CATL, NIO, Cambricon, Langboat Technology, among others. Yet the Lenovo Group VP and partner at Lenovo Capital and Incubator Group (Lenovo's CVC arm) proves far more playful and unconventional than his credentials might suggest.

Before the ChatGPT wave hit, China's primary market for AI investments had plunged to a deep freeze. Yet by Song's account, Lenovo's investments barely paused. The key, he argues, was an early appreciation for compute power — honed through seventy-plus AI bets including Megvii, 4Paradigm, and SmartMore. In a sense, this was the root of their four consecutive rounds in AI chipmaker Cambricon, and part of why they began positioning in large language model company Langboat back in 2021-2022.

When LLM investments became the rage in China, Lenovo zigged where others zagged: they trained their sights on a European LLM company and a middleware player that most investors were overlooking.

On whether this crop of AI companies will meet the same fate as the last, Song offers a pessimistic forecast — and an intriguing fix: subsidize AI like new energy vehicles, letting the state foot the bill to cultivate paying habits and willingness.

In his wonderfully eccentric office, we embarked on a wide-ranging conversation. Below is the full transcript:

Against the Grain

"Waves": Before the ChatGPT frenzy, there were several years when domestic AI investment hit a low point. Yet Lenovo invested in Langboat a year ahead of the curve. How did that call come about?

Song Chunyu: Actually, in both 2021 and 2022, we observed a major shift: massive compute was migrating away from CV — the computer vision domain where the "AI Four Dragons" were all competing — and toward large language models. That was a crucial signal.

"Waves": But tilting toward natural language understanding back then was still a bold call.

Song Chunyu: There was fierce internal debate. Natural language understanding is the crown jewel of AI, yet the technology had stubbornly failed to break through, and the user experience threshold remained unmet. But when we saw roughly 70% of global compute going into natural language training, we concluded we needed to plant a flag with a strong player in this race. Plus, Professor Zhou Ming has always been a leading expert in this field — one of only two ACL Fellows in China, and a former ACL president. We understood his competitive edge.

"Waves": Was your earliest bet on Cambricon also rooted in this compute-focused thinking? Its stock surged this year on massive LLM-driven demand.

Song Chunyu: We invested in Megvii in 2011, and after Lenovo Capital was formally established, we went on to back 4Paradigm, SmartMore, and over seventy other AI-related tech ventures. Through that process, we came to see that AI advancement is intensely compute-dependent.

We projected that AI processor performance would grow by orders of magnitude — hundreds of times over. So in 2017, we invested in Cambricon. What many people don't realize: Cambricon invented the AI processor globally. They pursued dedicated AI chips before NVIDIA did. The world's first paper on chip-accelerated AI computation was authored by the Cambricon team. Google's TPU architecture evolved from that very paper.

Moreover, after spinning out as a commercial entity, the Cambricon team demonstrated strong product competitiveness. At the time, transformer architectures hadn't yet gone mainstream, so compute demand was still primarily driven by deep learning rather than at today's scale. Even so, we saw Baidu, Alibaba, iFlytek, and Hikvision beginning to adopt their chips — confirmation that this path was right.

After the first round, we invested in three more consecutive rounds — a rare four-round concentrated bet for us.

"Waves": After ChatGPT, consensus formed rapidly. What adjustments did we make?

Song Chunyu: Before ChatGPT, the industry was exploring three or four competing technical approaches. When ChatGPT emerged, it was recognized as a major paradigm shift that electrified the entire industry. Domestic companies rushed into large models, poaching talent left and right. That's when we deliberately went against the grain and invested in a German company called nyonic.

"Waves": Why a European LLM company?

Song Chunyu: It's a German large language model company. Their chief scientist is Jakob Uszkoreit — a pivotal figure among the eight authors of the transformer paper. Though the paper listed equal contributions, with author order supposedly decided by dice roll, a second footnote identified Uszkoreit as the core contributor.

Additionally, nyonic's CTO is Johannes Otterbach, a founding member of OpenAI. Co-founder Hans Uszkoreit is considered the father of European large language models and director of the German Research Center for AI. Another partner and chief innovation officer, Dr. Feiyu Xu, previously led AI research at both SAP and Lenovo.

"Waves": Was the star-studded team the decisive factor?

Song Chunyu: First, we believe large models and AI capability are equivalent to "nuclear weapons" — they have geographic boundaries. Europe will have its own large language models.

Second, data is the moat in model training. Europe has strict data protection laws; if you train on non-compliant data, you must be able to extract it from the model.

Furthermore, Europe has excellent fundamentals: it's among the world's top three industrial markets, home to SAP, and attracts top global tech talent.

Where the Explosion Happens

"Waves": Domestically, have we invested in any new LLM companies?

Song Chunyu: We invested in OpenCSG, a middleware infrastructure company. With over 100 models emerging in China's LLM scramble, middleware players serve to lower the barrier for model companies and end users to deploy quickly — minimizing the friction of industrial LLM adoption. The closest parallel in the US would be Hugging Face.

"Waves": But a more common view holds that middleware is better suited to Silicon Valley than China.

Song Chunyu: It depends on the specific vertical. Many don't know how to put their data into model training, yet data cannot be shared. What's needed is a platform for distribution and delivery — where code can be hosted, evaluation tools are provided, and users can rapidly train models on their own data. But not every vertical requires such a platform.

Frankly, infrastructure has a time window, or optimal value-growth period. The opportunity exists when applications haven't yet taken off and LLM usage barriers remain high — though these will gradually normalize.

"Waves": Where do you see the biggest opportunities in this AI wave?

Song Chunyu: Beyond foundation models, the biggest opportunity lies in applications. The growth curve of large model capability is fundamentally about AI logic and reasoning. Next comes data-driven capability. This means any industry dependent on data could eventually be "replaced." Programmers, for instance, are on the replacement path — already over 50% of code in the US is AI-generated.

"Waves": On data access, tech giants and industry incumbents clearly hold the advantage.

Song Chunyu: For historical data, yes — giants and incumbents dominate, and they can leverage large models for new cost efficiencies. But there's another category: AI-native data, which doesn't rely on historical archives or incumbent control. This is the direction startups need to attack from day one, and I see it as the most important trend in applications.

"Waves": Are there already companies with AI-native products?

Song Chunyu: They exist in China, but need more time to mature. The core issue is that foundation models here aren't yet refined enough. Without breakthroughs at the base layer, front-end AI-native applications won't perform well either.

"Waves": When might the inflection point arrive?

Song Chunyu: Perhaps about a year after large models mature. Counting from last November, by late this November, if Chinese large models reach ChatGPT-3.5 level, applications could explode.

"Waves": Will agents and embodied intelligence bring new inflection points?

Song Chunyu: Both represent the next phase of large model applications. Gradually enabling models to understand intent will give rise to agents — still an active research frontier. Embodied intelligence is also highly imaginative. Large models plus robots bring model capabilities into the physical world, a forward-looking direction we particularly favor.

The Honey Trap

"Waves": In retrospect, why did China's "AI Four Dragons" and Big Tech AI labs collectively miss ChatGPT?

Song Chunyu: Companies like SenseTime and Megvii were largely forced by commercial realities into systems integration work, forfeiting the generative AI opportunity.

"Waves": And the Big Tech miss?

Song Chunyu: The problem for China's tech giants is that AI burns enormous cash — balancing blue-sky ambition with ground-level execution is extraordinarily difficult. Moreover, even overseas, Microsoft and Google didn't build this themselves. Capital alone isn't enough; you need conviction and belief.

In generative AI, there were multiple technical paths. One was BERT — what I'd call a honey trap. It achieved 80% performance with relatively little data and training, quickly. But we found that piling on more data still yielded 80%. Many Chinese giants chose this route.

OpenAI, by contrast, bet decisively on their own path: believe in scaling compute and data, and breakthroughs will follow. It took them seven years of burning capital to get there.

"Waves": These divergent technical bets spawned many tragic stories.

Song Chunyu: Many indeed. In this wave, Langboat, BAAI, and Zhipu AI held firm to the scaling path.

"Waves": Looking back at the post-2019 pessimism toward AI across China's primary market, how do you read it?

Song Chunyu: Primarily a crisis of confidence in AI's commercial prospects. AI is complex technology with a long chain from data to training to deployment — and it's cash-intensive. The "AI Four Dragons" burned capital for years without establishing sustainable business models, so industry grew anxious.

I believe we need to grant Chinese technology sufficient tolerance for exploratory breakthroughs.

AI's business model is hard to pin down, but it's a super-complex technological endeavor for humanity, equivalent to the Manhattan Project. If breakthrough comes, it represents the greatest advance for all humanity.

I sometimes genuinely don't understand: bottled water companies can command hundred-billion valuations, yet certain high-tech companies at tens of billions face demands for profitability. Without burning $10 billion, how could OpenAI have triggered an AI paradigm shift?

Same Old Story

"Waves": This AI wave too faces commercialization pressures. Do you think it will repeat the fate of the previous generation — falling back on project-based work to turn a profit?

Song Chunyu: Very possibly.

"Waves": Is there a potential solution?

Song Chunyu: Perhaps state subsidies or state purchasing.

"Waves": Why?

Song Chunyu: While AI isn't equivalent to SaaS — it's a far more important scientific breakthrough domain — it faces the same challenges SaaS encountered: Chinese willingness to pay, among others. I've told my team: my suggestion is that if the Chinese government supported China's SaaS and software industry chains the way it subsidized new energy vehicles, the outcome would be dramatically different. Look at new energy: China is now global number one. In US energy storage, CATL holds half the market. If China nurtures its next emerging industry, say software, that sector too would achieve rapid development.

"Waves": Why has Chinese willingness to pay never properly developed?

Song Chunyu: On the consumer side, the previous generation of internet companies failed to establish healthy payment habits — too much was free. In mobile internet, people would pay for entertainment memberships, but never developed the habit for utility software.

On the enterprise side, China's SMEs are simply too squeezed, plus their weak position in the value chain means limited profitability — making software payments even harder. Someone once told me China's SaaS gap with the US had narrowed to five years. But returning from the US recently, I'd say it's still twenty years. Any US startup purchases a dozen-plus software tools; Chinese startups might pay for one or two.

"Waves": Would state subsidies definitely work?

Song Chunyu: Why don't SMEs buy? Because they can't predict whether they'll survive the next year or two. If the state subsidized 80% of software costs for a company's first two years — then after they've survived three years and gained payment capacity, taper subsidies to 30%. New energy vehicles followed this trajectory. Once companies can afford it, they'll purchase naturally. When entire industries use legitimate, interoperable software, digitalization takes off.

"Waves": Can SME growth pull along large enterprises?

Song Chunyu: Because SMEs won't pay, software companies are forced into systems integration for large enterprises. If SMEs could become profitable, systems integration wouldn't be necessary, and standardization would gradually emerge. Part of our rationale for investing in European large models is that European users have payment willingness, making the business model smoother.

"Waves": What competitive landscape do you foresee for this large model wave?

Song Chunyu: The US pattern is set: Microsoft and Google at the top, with entrepreneurs clustering around infrastructure tools and applications. China's first wave has everyone building foundation models, but this will certainly consolidate.

For China's foundation models, we foresee a hybrid state. Unlike the US with two, maybe three players, China may have 10-20. Look at the internet structure: Baidu excels at search and text, Toutiao at multimodal, Tencent at social and gaming.

Startups may also claim territory. The market is large enough, and China's "involution" keeps costs low. The upside: winners forged in China's crucible can compete globally.

---

Image source: IC photo

Layout: Xuemei Guo