In the 18 months since Microsoft-backed OpenAI launched ChatGPT, Chinese companies big and small have rallied together with a singular goal: to beat the San Francisco-based startup with their own Chinese-language chatbots.
Results have been mixed, with some tech giants claiming to outperform OpenAI's state-of-the-art model, GPT-4, on Chinese-language queries. But given that the technology that powers these chatbots — China's large-scale language models (LLMs) — is churned out by more than 200 companies vying for market share, Chinese AI companies can claim at least one other clear, if less boastful, advantage over their U.S. counterparts: price.
In recent weeks, TikTok owner ByteDance, internet search giant Baidu, e-commerce conglomerate Alibaba Group Holding and social media giant Tencent Holdings have all significantly slashed the prices of their LLM services, and some are even offering various services for free. One of ByteDance's premium services, Doubao Pro, costs 0.0008 yuan (0.011 US cents) per 1,000 token prompts, 99.8% cheaper than what OpenAI charges for access to GPT-4.
Have questions about the biggest topics and trends from around the world? Find the answers on SCMP Knowledge, our new platform that provides curated content with explainers, FAQs, analyses and infographics from our award-winning team.
“Bad money seems to be driving out good money,” said Yu Yang, a professor of computer science at the National University of Singapore.
According to Yu, the price war indicates a lack of competitiveness based on the merits of the model itself, and previous market prices cannot attract customers.
In the United States, tech giants including Google parent Alphabet, Facebook parent Meta Platforms, Amazon.com Inc. and Microsoft Corp. are also competing in AI by following a “blitzscaling” strategy popularized in Silicon Valley: acquiring users as quickly as possible at the expense of revenue in order to corner the market.
But while they enjoy a “walled garden” behind the Great Firewall, which censors foreign internet services, Chinese companies face restrictions from Washington on exports to Nvidia Corp., whose advanced chips contain graphics processing units that are at the heart of developing AI models, and they have less buying power than their peers in the more liquid U.S. market.
“Price reduction [for Chinese AI services] “The main purpose is to attract more customers, so it's more of a branding activity,” Xu Li, CEO and co-founder of Hong Kong-listed AI company SenseTime, told the South China Morning Post in an interview this week.
In mid-May, ByteDance announced prices for its Doubao AI enterprise services that boldly undercut those of its domestic rivals, signalling the start of a price war.
At this pricing, 1 RMB would purchase 1.25 million input tokens. In comparison, it would cost approximately US$37.50, or 272 RMB, to purchase 1.25 million GPT-4 tokens.
In AI, a token is the basic unit of data processed by an algorithm. For the Chinese LLM, a token typically corresponds to 1-1.8 Chinese characters.
Other Chinese tech giants were quick to respond.
Alibaba, the owner of the post, was the first to respond, slashing the price of its Tongyi Qianwen (Qwen) service by up to 97% from 0.02 yuan to 0.0005 yuan per 1,000 token prompts, which is 0.0003 yuan cheaper than ByteDance.
Several other companies, including Baidu, Tencent and AI specialist iFlytek, known for its voice recognition technology, also offered deeper price cuts, with some even offering access to their less powerful LLMs for free.
Wang Sheng, an investor at Beijing-based InnoAngel Fund, said this “vicious” price war was hurting local AI startups.
“Big tech companies are not necessarily better than startups when it comes to fostering law masters,” Wang said, “but the way they subsidize them to grab a bigger market share is harmful to these companies.”
Alain Le Coudic, senior partner at AI investment firm Artificial Intelligence Quartermaster (AIQ), suggested that the price war will benefit some companies over time.
“The race for market dominance is a sign that many companies see attractive opportunities ahead, even if it means some pain in the short to medium term,” he said.
LLMs are energy intensive, making them expensive to run and the marginal cost of adding new users likely to be higher than other online services, making blitz-scaling more complicated for AI services, although the race to make LLMs more efficient may eventually change this calculus.
“People are pursuing these efficiency gains because there is profit to be made from doing so,” Bill McCartney, chief technology officer at venture capital firm Signalfire and a computer science professor at Stanford University, told The Post at the UBS Asia Investment Conference this week. “These models are becoming increasingly expensive to operate, and people have a strong economic incentive to find ways to make it cheaper.”
McCartney noted that resources are being poured into delivering efficiency gains at multiple levels, “There are improvements at the silicon level, there are improvements to the model architecture, there are improvements to the software that we layer on top of the models.”
“It's not unrealistic to imagine that we'll see a 10x improvement in inference efficiency across AI in general within three years,” he said.
Bill McCartney (right), chief technology officer at venture capital firm Signalfire and professor of computer science at Stanford University, spoke about generative artificial intelligence at the UBS Asia Investment Conference in Hong Kong on May 29, 2024. Photo: Handout alt=Bill McCartney (right), chief technology officer at venture capital firm Signalfire and professor of computer science at Stanford University, spoke about generative artificial intelligence at the UBS Asia Investment Conference in Hong Kong on May 29, 2024. Photo: Handout>
Some companies have cited increased efficiency in training and running models as the primary reason for price cuts: OpenAI attributes a significant price cut to its GPT-4o model, which launched in May, to new efficiencies.
Robin Li Yanhong, founder and CEO of Baidu, said in April that the training efficiency of its flagship product, Ernie LLM, had improved by 5.1 times within a year, the model's inference performance had improved by 105 times, and inference costs had been reduced by 99 percent.
ByteDance also said it lowered prices because it believes technological improvements can reduce costs.
Whatever the gains, tech companies have been quick to claim that their growing revenues are due to the AI boom.
Alibaba Cloud said its 3% growth in the March quarter was driven by accelerating AI-related revenue. Baidu Cloud reported 12% revenue growth in the quarter, with generative AI and foundational model services accounting for 6.9% of total AI cloud revenue.
US tech giants Google and Microsoft have similarly reported robust demand for their cloud-based LLM services: In its third quarter results, Microsoft's Intelligent Cloud division saw revenue grow 21% year-on-year, while Google saw revenue grow 28% year-on-year in the first quarter.
“We've seen Microsoft's market capitalization soar and everyone wants to take advantage of this opportunity,” said Ivan Lam, an analyst at market consultancy Counterpoint Research. “The Chinese market, in particular, [AI] Applications and Business [uses] This is to encourage further advancement of the LLM.”
Alibaba declined to provide data on increased usage of LLM following the recent price cuts, but said the number of calls to its Qwen application programming interface (API) from major Chinese recruitment agencies jumped 100-fold within a week.
At least for LLM users like the recruitment agency, the current economics of AI services appear to be paying off.
Zhao Chong, founder and CEO of AI-powered graphic design service iSheji, expressed optimism about reducing the cost of law master's programs in an interview with Chinese news portal Sohu.com published on May 24.
“For a startup like us that develops applications, [the price war] “The cost of LLMs used to account for 5-10 percent of the total cost, but now it's 1 percent, improving our profit margins,” Zhao told Sohu.
These consumer service providers are lowering the prices of their services, and many are offering them for free.
Meanwhile, startups well placed to join the price war are trying to step in. Beijing-based Baichuan and 01.AI, founded by Lee Kai-fu, a Taiwanese computer scientist who once ran Google China, have rejected the idea of lowering prices.
You, the NUS computer scientist, noted that the lower prices are a benefit to app developers, but warned that applications built on a substandard foundation model may suffer from poor performance.
Nevertheless, the price war has erupted because companies have limited options: It's always easier to compete on price than to endlessly improve model capabilities, says Yang Lijie, founder and CEO of Shanghai-based MiniMax, one of China's four biggest AI companies.
“[Reaching] “The upper limit of the technology is uncertain and requires further exploration,” Yang said in a fireside chat published by Chinese tech news site Geekpark on May 23. “Meanwhile, there are always ways to bring the price down.”
Counterpoint's Lam said price wars may be inevitable for companies wanting to maintain an advantage in AI services as the market becomes increasingly crowded.
Some Chinese tech giants appear to have an advantage in terms of computing resources and funding. AIQ's Le Coudic said it's too early to predict who will win the AI price war because the AI industry is still immature. Both business models and technological advantages will be key factors in determining which companies will come out on top, he added.
“At the end of the day, the company with the best service and the best technology will win,” Le Coudic said.
Additional reporting by Matt Haldane.
This article originally appeared in the South China Morning Post (SCMP), the most authoritative news source on China and Asia for more than a century. For more SCMP articles, visit the SCMP app or follow SCMP on Facebook. twitter P a g e Copyright © 2024 South China Morning Post Publishers Ltd. All rights reserved.
Copyright (c) 2024. South China Morning Post Publishers Ltd. All rights reserved.
