What the LLM Price Wars in China Imply for US AI App Makers
Alibaba kicks off 85 percent discount; ripple effect seen on OpenAI’s GPT-4o; high cost to develop GenAI apps can cause strategy to change; it did for Character.AI
By John P. Desmond, Editor, AI in Business
LLM price wars in China are providing insight into LLM pricing in the US. On December 31, Alibaba Cloud unveiled an 85 percent price reduction for its Qwen-VL-Max model, down to $0.00041 per thousand input tokens.
This was according to an account from CTOL Digital Solutions, published days after the Alibaba discounts were announced.
The discount positions Qwen-VL-Max at only 14 percent of the cost of OpenAI’s GPT-4o, which is priced at $2.50 per million input tokens, according to the account.
The pricing moves stem from an effort by Alibaba to differentiate from rivals in China. It is not expected to represent a threat to US LLMs, since Chinese LLMs face “hurdles in international markets due to issues like model censorship, data privacy concerns, and limited ecosystem integration,” stated the writer, Xiaoling Qian.
The Chinese LLM landscape presents challenges, with over 250 generative AI models approved in 2024, driving companies to innovate through pricing instead of unique features, according to the account.
Chinese LLM Suppliers Focusing on Enterprises
The Chinese LLMs are also engaged in a different marketing strategy than is OpenAI, at least initially. In addition to Alibaba, tech giants including Tencent, Baidu, JD.com, Huawei and ByteDance have all introduced their own large language models (LLMs), according to a recent account in PYMNTS.
The Chinese are focusing their LLMs on enterprises, whereas OpenAI has started out aiming at consumers. In May, Alibaba announced its Qwen models have been deployed by more than 90,000 enterprise customers.
One observer sees the pricing moves as “democratizing” AI services to a degree, enabling smaller startups to be more competitive and established firms to move to AI at lower costs. “We will see significantly more competitiveness since much VC investment is going towards ‘picks-and-shovels’ AI technologies, and we should see more global pricing wars, which will be good,” stated Nick Rioux, cofounder and CTO of AI company Labviva, in the PYMNTS account.
GenAI Development, Deployment Costs Range Widely
The question of how much AI projects cost is a popular one given the projected growth of AI spending. A recent blog post from digital marketing services firm WebFX offers some example costs, running from $6,000 to $300,000 for a custom AI solution, and up to $40,000 to buy third-party AI software. AI consultants charge $200 to $350 per hour. Third-party AI software, such as prebuilt chatbot, can cost up to $40,000 per year.
Example projects include chatbots, analysis software and virtual assistants. Prebuilt chatbot options cited by WebFX include Drift, HubSpot or TARS. Drift can cost from $400 to $1,500 per month, while TARS costs $99 to $499 per month. In comparison, prices for developing a custom chatbot can start at $6,000 and go up to $15,000.
The features of an AI system will impact the cost. If your project calls for a custom data analysis system that uses top AI software, such as IBM Watson, “prepare for higher AI software prices,” the authors advise.
Other factors to consider include: data format, data storage, data structure, data processing speed, minimum accuracy rate for predictions, data visualization and dashboard requirements.
In-house management of the AI system tends to cost more than outsourced management. Data scientists earn an average salary of $94,000; developers, about $80,000 per year. An in-house team is likely to understand the business, brand and users more, while outsourced management avoids the cost of adding employees. Typical agreements include a one-time cost or a monthly fee.
Another attempt to rough out the cost of building a system incorporating generative AI comes from Aalpha Information Systems of India. “Generative AI is the game-changer in innovation,” offering creative content generation and the ability to solve complex problems, states writer Stuti Dhruv, senior technology consultant, in a recent blog post.
The cost to build a GenAI program includes initial development, data collection, model training, release and maintenance. A basic GenAI app can range in cost from $20,000 to $150,000; a more advanced app can cost between $100,000 and $500,000, she suggested, emphasizing, “The figures are approximations.”
Researchers and data scientists needed to build a more creative and original GenAI app, can earn between $50,000 and $150,000; domain exports, from $60,000 to $120,000. The cost to acquire high-quality datasets can range between $30,000-0 and $100,000; cleaning and processing the data can cost from $20,000 to $60,000. To integrate into external systems, APIs or databases, can cost between $30,000 and $100,000; deployment costs can range from $40,000 to $120,000.
To measure the reliability of your AI app, you need to test and validate, costing between $20,000 and $60,000 for testing, and $30,000 to $80,000 for validation.
The geographic location of the development team has a big impact on costs. Higher rates are paid in Western Europe and North America, compared to Africa, South America, Asia or Eastern Europe.
More costs arise from the hardware and software infrastructure needed to run the AI app, and from efforts needed for regulatory compliance and ethical considerations, such as privacy laws.
In the end, the author estimates the cost to develop and deploy the GenAI system can range between $600,000 and $1,500,000. Continuous annual costs after that can range from $350,000 to $820,000.
As of now, the high costs do not seem to be translating to a slowdown in growth of GenAI apps; quite the opposite. “Generative AI is here to stay,” the author states.
Character.ai Changed Strategy Given High LLM Dev Costs
However, for one company, the cost to develop AI models was high enough to lead to a complete change of business strategy.
Character.ai decided to bail out of the race to build AI models, opting instead to focus on its popular consumer product and enter into a $2.7 billion agreement with Google, that included the acquisition of the two founders of the company.
Character.AI offers what is called a “companionship platform” that enables users to create and interact with characters. Founded in 2021 for former Google researchers Noam Shazeer and Daniel De Freitas, the company as of July 2024 had six million active users, according to a profile from Contrary Research.
“It got insanely expensive to train frontier models … which is extremely difficult to finance on even a very large startup budget,” states the interim CEO of Character.AI, Dominic Perella, in an account in the Financial Times.
“Our consumer products got incredible traction, and you had a dichotomy inside the company of folks who wanted to focus on training the most cutting-edge possible models versus folks who came from a consumer background seeing this product take off.”
Character AI’s cofounders left the startup at the end of August, when Google rehired them as part of the $2.7 billion deal, the Times reported.
Read the source articles and information from CTOL Digital Solutions, in PYMNTS, blog post from WebFX, from a blog post from Aalpha Information Systems of India, Contrary Research and from the Financial Times.