Almost everyone is chasing the next ChatGPT but in market powered by guesswork, few have actually put down their money.
Photo by Fan Jianlei
By LI Jingya
The delirium shows no sign of abating. FOMO runs through the markets like virus. Thousands of investors with hundreds of billions to invest are chasing fever dreams from Palo Alto to Zhongguancun. The only thing they know for sure is that ChatGPT will change the world. They don’t know how, but it will. Climb on board, or all is lost.
“ChatGPT heralds the revival of venture capital. We are full of dreams again and starting to believe in the future again,” LIU Dawei, founder of dollar venture capital fund Capital O, wrote in a WeChat post.
In the past month, Liu has been flying all over the country to scout out worthwhile AI startups, preferably those involved in – or claiming to be involved in - Generative AI, the key to whatever dark secrets lurk with the heart of OpenAI’s uncannily human-like chatbot ChatGPT.
Everyone in China’s VC world fears missing out, but no one is particularly clear about what they are in danger of missing out on. Before we know it, ChatGTP will be driving us to work and helping us choose a coffee. It will be taking care of our children, doing their homework for them and choosing their college major. Non-specific, distant dreams of societal shifts and unimaginable wealth abound.
But the rules of investing always hold, said Liu.
These rules basically amount to two things – how much money will it make, and when? The best investors are those who make the best estimates of those two metrics. At the moment, it’s anyone’s guess.
Startups must find technologically and commercially viable areas where machines can replace humans. Industry heavyweights have unleashed their star team on the problem. Funds with no experience in AI are poaching talent and building new teams. In terms of generative AI, valuable projects are almost all overseas.
All this fuss is great for self-important executives to rack up air miles, and wonderful fodder for team bonding, but it’s not actually opening many wallets. Only 34 fundraising deals were completed last year and just eight have got off the ground this year.
“AI valuations are high and are only going higher. Large language models are extremely expensive. You need at least hundreds of millions of dollars, if not more, for a single deal. It’s a lot of money for any fund,” said Liu. “Funds are cautious.”
Cautious of investing they may be, but many seem to be even more cautious of showing their hand than they are of spending money. DCM, an early investor of Tiamat (AI-generated artwork), has had an AI team since at least 2021. The fund has a reputation for being “almost immune to news cycles.” Unity Venture invested in Colorful Clouds (simultaneous interpreting) as early as 2017.
“The best time to get in was when no one was paying attention,” said Liu. “When it’s all hype, good ones are gone.”
XU Xiaoyu of AMINO Capital spends a large chunk of her time reviewing AI startups already in her portfolio. Many have reached out to her since ChatGPT but AMINO hasn’t made new investments since last year. She has been working to acquire AI startups or remodel existing AI subsidiaries.
“Fundraising and acquisition are both on the rise. This happens in any hype,” she said.
Hiring is up, spurred by celebrity entrepreneurs associating themselves with new AI ventures. WANG Huiwen, a co-founder of Meituan, recently pledged US$50 million of his own money to build a Chinese OpenAI.
Arguably the biggest beneficiaries of ChatGPT are similar natural language processing (NLP) startups, which used to be dismissed as costly, time-consuming, and commercially unappealing. Many investors thought they were just for customer service.
“Large language models are definitely worth following. They have proved to be transformatively innovative. We are talking about a trillion-dollar market,” said ZHANG Yutong, of GSR Ventures.
Almost all analysts talk about generative AI as a “three-layered market.” At the bottom are foundation builders – often backed by big tech or part of big tech – who develop the fundamental technologies. On the top are user-facing applications such as those that draw pictures and write codes. Between them is a vaguely defined middle which consists mostly of products that help everyone else train their models more efficiently.
By some estimates, the cost of a single training session for ChatGPT’s predecessor GPT-3 is $1.4 million. Larger models are even more expensive. Given the cost, foundational models are seen as the turf of big tech.
“Tech companies are taking action,” said Liu. FANG Jiarui, who worked on NPL projects for WeChat before founding Colossal AI, said Tencent had been “rather indifferent to large language models until probably 2021.” Now, the entire company is paying attention, he said.
The “middle” also requires an enormous amount of money and computational power, but a handful of startups with esoteric focuses are gaining traction among investors. Some of the most famous ones include Moffett AI, which specializes in sparse computing, and Colossal AI, which aims to make training cheaper and faster. Moffett AI founder WANG Wei once said in an interview that a startup needs to improve existing AI performance by at least ten times to break the monopoly of big tech.
Most investors demurred when asked whether they would invest in the “middle layer.” Some say startups that make software and developer tools for large language model training are worth following. Others keep an eye on previously stagnating or neglected branches of machine learning.
Liu said a potentially groundbreaking technology is “only meaningful when combined with an eco-system.”
In other words, an investment generates returns only if there are commercially viable products.
“To promote VR, Meta spends hundreds of millions every year on products. Some of them are really niche” he said. “When it comes to competition, technology itself may not be the most important.”
Investors are no more committed when it comes to “top-layer” applications despite the many writing, drawing, coding and translating apps already out there. Valuation depends on both the underlying model and commercial potential, neither of which is easy to tell, said LIU Yong, of Blue Run Venture.
They also seem conflicted about what exactly to look for. Some favor applications with sophisticated, proprietarily trained models, such as the image generator Tiamat. Others stress the ability to scale up.
“Simple chatbots can be great. As long as they are used in the right place and can find willing customers, they can make a lot of money,” said Liu, who is happy to talk to anyone regardless of their models. What he cares about is whether a company has bargaining power in the supply chain.
Another train of thought is that the three-layered split of generative AI is rigid, myopic, and narrow-minded. Founders and investors, obsessed over who can challenge ChatGPT or make the best use of it, should instead imagine how generative AI crosses over to other industries, including many traditional ones such as healthcare and pharmaceuticals.
Or maybe, the current moment is only a feverish and confusing interim that will quickly grow into a completely different but no less exciting version of itself. AI will be more powerful and widely used.