Unique Research · Unique Viewpoint · March 27, 2024
Background: On January 24, 2024, Unique Capital and partners hosted the 2024 AIGC Application Development Summit in Shanghai, with nearly 1,000 AI practitioners. This piece is compiled from the guest dialogue.
Guest
- Speaker: Lei Fang (方磊) — Chairman, DataCanvas / Jiuzhang Yunji (九章云极)
- Host: Xiaojing Hu (胡小婧) — Partner, Unique Capital
Summit dialogue
Q: ChatGPT launched a distinctive tech wave. As a founder who has worked AI for a decade, how does it feel?
Lei Fang: We would not claim to have spanned two full AI waves and cycles — only that our team started a bit earlier and has stayed longer. Years ago another co-founder and I returned from Microsoft to found China’s first machine-learning platform company. Today our strategic positioning is AI Foundation Software — a full stack of model services for customers.
DataCanvas’s path has several stages:
First stage (2014–2015). We did not give models to end users; we gave developers tools to build models. Public attention then was more on O2O.
Second stage (2016). AlphaGo beating Lee Sedol lit up AI. Most still believed that in China, AI value would come from end-to-end services — tools alone could not create value for people.
Third stage (2020–2023). After 2020, as first-wave AI landed and category leaders formed, people found the last mile was not simple. Precisely because that mile is hard, tool value became visible. We grew fast; leading enterprises in benchmark industries are largely our customers.
Fourth stage (2023–now). 2023 was a huge turn — no need to belabor it. Any technical person is startled the first time they use ChatGPT; especially people like us — algorithm-trained, paper-publishing, PhD-track, then founding in AI. Traditional methodology made us slow to believe LLMs; people deepest in ML/DL were often the most skeptical at first. My first ChatGPT use was also a shock — an upgrade of prior knowledge.
Before LLMs, our customers’ largest GPU clusters were about 300 cards — we already thought that was too many to use well, so there was much GPU virtualization to train many small models in parallel (PyTorch). Now 300 is too few: even a small 7B model needs 500+ cards.
That hits people with traditional backgrounds hard — and is an unprecedented opportunity for the whole AI industry.
Most people hope to put AI into personal life and work. For me, generative AIGC’s rise drives huge compute consumption — meaning China’s digital-economy new infrastructure, the compute economy, is booming.
AI apps will keep emerging into daily life. Yet next-era AI-native applications will not appear overnight, however much people want them. Patience matters. Personally I do not expect a killer app this year. Douyin arrived in 2013; the iPhone entered mass view in 2007 — longer than anyone imagined. Patience is essential.
Given China’s characteristics: before a killer app appears, the country can use the window to lay infrastructure. We are in an era that needs expanded compute infrastructure.
So I was very excited about 2023: not only large models advancing the industry, but a year for China to lay out AI infrastructure.
Q: How should the industry see ChatGPT’s revolutionary role?
Lei Fang: People long in NLP or Transformer pretraining may be less shocked; still, most first-time ChatGPT users marvel at its reinforcement learning and human interaction. Two years ago, after Attention Is All You Need, many built Transformer vision models — but still doubted final effect.
Today the shock is at the product level. OpenAI’s tech is not unique — the base came from Google papers — but product and tech have a large gap. ChatGPT looks simple yet is, in a sense, humanity’s first AI-native product. OpenAI’s power is that the product flywheel has started. I do not think that maps purely to tech; a small tech step that guesses the right direction can be a huge product step.
We still have many tech options. Many say OpenAI is far ahead; I disagree. History of first products has three cases: Yahoo made search, Google made it revolutionary and monopolistic; iPhone is excellent and owns half the phone market, yet Android is a strong rival; Netscape made the first great browser and was then killed by Microsoft. So OpenAI need not be the absolute lasting leader. I still hold that view.
Over the past year we saw more tech possibility and change. Is Transformer best? Are Nvidia GPUs always most efficient? Not necessarily. Humans may find a more elegant, simpler, more symmetric Transformer — then much of the stack changes; or build specialized chips on OpenAI Triton for more efficient compute — all lower-layer change — plus AI-native interaction modes and product forms.
ChatGPT’s biggest gift is showing AGI is reachable. From chips to new Transformer architectures and beyond, people are exploring. In the next two to three years we may see more base-tech change and new AI-native product forms — we will meet them soon.
Q: China is concentrating on infrastructure; compute is scarce; a compute economy has emerged; buying cards is hard. Your view?
Lei Fang: Two sides. Stats: in 2023, global H100 demand was about 1.7M units; Nvidia produced about 1.5M. For 2024, H100 demand for training + inference (inference rising faster) is about 4.6M; Nvidia’s expected output will not exceed 3M. Globally, compute is relatively scarce. Geopolitics adds friction: China cannot buy Nvidia cards smoothly under sanctions. Domestic GPU vendors exist, but catch-up takes time. So compute remains relatively scarce.
China is building intelligent compute centers at scale; some fear overcapacity. I do not. On current plans we estimate China will have 500–1,000 intelligent compute centers in three years — a striking scale. Today “compute serves AIGC apps” is an indirect logic; the direct logic is infrastructure investment — which needs foresight and courage.
Infrastructure is moderately ahead of demand — more investment logic than pure usage/application pull. China’s current compute build is investment and infrastructure logic, not AIGC-application consumption logic. That itself becomes digital-economy new infrastructure — the compute economy. Local governments, SOEs, and industry leaders fund many centers: national infrastructure and a possible overtaking chance.
In the public-cloud era, China’s public cloud lagged the U.S./global commercial high-value cloud. In the LLM era, building 500–1,000 intelligent compute centers nationwide can be read as building a decentralized GPU cloud (versus overseas players somewhat like CoreWeave). So even if we did not build high-commercial-value public cloud in the CPU era, in the GPU era intelligent compute centers may go far.
Under today’s compute economy, nationwide intelligent-compute rollout is not only an investment opportunity with deep industry impact — it may be China’s best chance to build a GPU public cloud.
Q: In ToB, how do you see the AI landing path?
Lei Fang: Landing views differ. Generative AI can do many things, but what people accept most today is the “knowledge agent” module — typically LLM-based RAG. Previously unstructured knowledge had only search; with LLMs, RAG can output better. Example: a top Shanghai steel manufacturer — new factory workers lack equipment-maintenance knowledge; train/fine-tune an LLM on traditional work orders to answer efficiently. In plain terms: teach search to speak human. More plain language is a big step for many ToB firms.
Past enterprise architecture handled structured data and warehouses reasonably; unstructured data capability was near zero — mostly cloud-disk storage. Now we first raise unstructured processing. Structurally: structured data via warehouses, data lakes, and stores; unstructured is catching up fast.
As RAG improves, unstructured will match structured handling. Eventually all data — structured or not — may be handled as vectors. That is a huge enterprise-architecture shift affecting business models, industry structure, sales models, and more.
Those structural changes let enterprises treat unstructured data like warehouses — process, analyze, cognize. That is key to landing, and where they start first.
Versus OpenAI or top model vendors, most enterprises’ real moat is private data. Without that premise the enterprise AI market would not exist. Samsung needs process secrecy in semiconductors; many firms forbid OpenAI/open models and push self-trained models — essentially a data-boundary problem. Data boundaries will keep defining much else. Back to enterprise AI: the fundamental defining boundary is private data.
3 CN images (summit/guest); attach on ingest. Full Q&A preserved.
— END —