[비즈한국] “AI innovation is far more powerful than ‘Moore’s Law,’ which states that semiconductor technology advances by twofold every 18 months. While there are currently many cost limitations, the prices of a significant number of products will eventually drop.” (OpenAI CEO Sam Altman, from his personal blog on the 10th)
The dawn of a price war in the generative artificial intelligence (AI) market has arrived. Following OpenAI and Google’s cuts to enterprise AI fees last year, there has been a continuous stream of low-cost, high-performance AI releases this year. On his personal blog on the 10th (local time), Altman emphasized the fundamental premise that “as prices go down, usage increases significantly,” noting that the cost to use OpenAI’s 'GPT-4o', released mid-last year, has become about 150 times cheaper than models from two years ago. The emergence of the Chinese AI startup DeepSeek is expected to further accelerate this chicken game. Major US companies like OpenAI, which previously shied away from disclosing information about their models, are also considering open-source methods, suggesting that tech openness policies among various firms will likely expand.
The 'Low-Price' Wind Blowing from China Leads to Global Bleeding Competition
A low-price offensive by AI companies has begun. Leading this cutthroat competition are Chinese firms. On the last day of 2024, Alibaba, China’s largest e-commerce company, slashed the price of its 'Qwen-VL' by up to 85%. The current price for Qwen-VL is $0.41 per 1 million input tokens (the unit of data processed by AI models). This is 83.6% cheaper than the $2.50 fee for OpenAI’s latest model, GPT-4o. Alibaba has already implemented price cuts twice, targeting enterprise customers. Last February, it lowered the prices of its major cloud products by up to 55%, and three months later, it reduced the prices of nine LLMs built on its generative AI service ‘Tongyi Qwen’ by up to 97%.
Baidu, often called the Google of China, is also pivoting toward a free-of-charge strategy. Baidu announced on the 13th that it will provide its AI chatbot 'Ernie Bot' for free starting in April. Since late 2023, Baidu has offered premium search features based on Ernie 4.0 within its engine for approximately 59.9 yuan (about 12,000 won) per month. An in-depth search feature equipped with improved reasoning capabilities and expert-level answers will also be released for free in April. While Ernie Bot is free, fees will still apply if corporate clients access AI through Baidu's cloud services.
Baidu’s strategic shift is seen as a reflection of the pressure Chinese firms face in the AI race. Baidu once led the Chinese AI competition by releasing its own AI four months after the emergence of ChatGPT, but it has since fallen into a position of chasing others as it lost ground to competing services like Doubao.
TikTok’s parent company, ByteDance, released an upgraded version of its flagship AI model 'Doubao' last month. Claiming performance equivalent to competing models like GPT-4o or Anthropic’s Claude 3.5, ByteDance asserts that this model performed better than GPT-o1 on the American Invitational Mathematics Examination (AIME), a benchmark for mathematical ability. Conversely, its operating costs are significantly lower—one-fifth that of DeepSeek and 0.5% that of the o1 model. The cost for the existing commercial base model is about 1% of the price of ChatGPT. It is provided for free to general users.
US AI Also Moves Toward Lightweight, Low-Cost Models; Who Will Capture the Price Edge?
The focus of the generative AI industry is shifting toward reducing margins and increasing efficiency. If the competition has thus far been about technological prowess measured by training volume and absolute performance, it is now evolving into a full-blown battle for market leadership centered on price.
In fact, the low-price competition has spread to US Big Tech, which represents the premium segment of the AI field. On the 31st of last month, OpenAI launched its latest reasoning model, 'o3-mini'. Its price is about 60% lower than 'o1-mini'. Previously, only paid subscribers to ChatGPT Pro or Plus could access reasoning models, but this service is the first to be opened to general users. While the o1 model is a broad general-knowledge reasoning model, o3-mini is a cost-optimized model focused on utility and speed. With strengths in code generation, it can be seen as a service specialized for developers and similar users.
Google also priced the lightweight version of its latest AI model, 'Gemini 2.0', released on the 5th, at a level comparable to DeepSeek. Among its three product lines—Pro, Flash, and Flash-Lite—the Flash-Lite model is aimed specifically at DeepSeek. The cost is $0.019 per input token, which is cheaper than OpenAI and on par with DeepSeek ($0.014).
OpenAI and Google already engaged in a price-cutting war over enterprise AI last year. Enterprise AI refers to the large AI models that serve as the foundation for businesses or individuals to create customized AI services. Platforms utilize Big Tech’s LLMs (Large Language Models) to provide various types of bespoke AI apps or services. OpenAI reduced the price of GPT-4o from $7 per token at the time of its 2023 launch to $4 last year, and shortly after, Google cut the enterprise price of 'Gemini 1.5' Flash to $0.12, one-third of its previous price.
Concerns That 'Margin Evaporation' May Slow Innovation
Analysts suggest that as the technological gap narrows, the weight of competition is shifting toward margins. Specifically, by using a low-price strategy for enterprise AI—the main business model at this time—companies seem to be attempting to secure clients and establish a foothold for dominance.
The issue is profitability versus development costs. The costs incurred in developing and operating LLMs are astronomical. Lee Im-bok, head of the Second Brain Lab, stated, “We are now at the stage where the ‘receipts’ are starting to return to these companies. AI web search functions have already largely become free. While benefits for general users will gradually increase, the differentiation between services will inevitably decrease. This is why there is so much effort put into price competition and B2B expansion.”
Foreign media outlets are even discussing the potential for 'margin evaporation,' predicting that companies will face significant challenges. The IT publication Wired analyzed, “OpenAI wants to prove it is at the forefront of AI development and commercialization. Cases like the release of a powerful model like o3-mini for free are acting as pressure on Google and Anthropic to lower their prices.”
There is also a perspective that if margins decrease due to price cuts, AI technology providers might choose efficiency over innovation. Forbes predicted, “Pressure to cut prices reduces research and development capacity and could potentially slow the pace of innovation,” adding that “the entry of low-cost competitors like DeepSeek into the international market will likely trigger a domino effect similar to that in China within the US premium market.”
Whether the open-source strategy adopted by DeepSeek, Meta, and Elon Musk’s xAI will become the mainstream is also a point of interest. With competitors recently promoting open-source AI, even OpenAI, which has insisted on a closed-source strategy, is considering an open-source approach. The open-source model is a strategy to grow the AI ecosystem itself to increase the company’s share and influence among businesses and researchers. While it is expected to lead to the 'democratization of AI' by enhancing transparency, safety remains a challenge due to the risks of misuse.
Lee Im-bok noted, “With the emergence of Chinese companies, the discussion of ‘value for money’ in the AI market has arrived faster than expected. As leading firms like OpenAI respond with both defensive and offensive measures, they are being asked to prove their price-to-performance advantage and the rationale behind their massive investment expenditures.”