Although ai promises to help companies reduce costs, its own high cost has always been a big problem.
For a technology that promises to help businesses cut costs, the cost of artificial intelligence itself has always been a major problem.
The law of AI expansion suggests that building more powerful models requires more computing power, which forces tech companies to invest billions of dollars in building massive data centers and buying powerful chips, and these costs cannot be absorbed by the customers. Google's AI tools can generate documents or emails, but they are not cheap. The company needs to pay an additional $20 for the $6 per month Workspace suite. Microsoft's (MSFT.O) Copilot AI assistant requires $30 per employee per month.
Meanwhile, research company Gartner has said that deploying AI directly into a company's systems can cost between $5 million and $20 million, and it is estimated that by the end of 2025, 30% of generative AI projects will be abandoned due to these high costs.
For customers, the good news is that the cost of AI seems to be declining, helping to narrow the gap between returns and investments. The bad news is that this still doesn't solve bigger practical issues, which will take years to resolve.
The common wisdom in Silicon Valley is to continue investing to capture opportunities in the future. Last Tuesday, Microsoft announced a record $19 billion quarterly capital expenditure, up more than 80% from last year. CEO Satya Nadella said all that investment will continue to "capture the opportunity with AI." Alphabet (GOOGL.O) CEO Sundar Pichai expressed a similar viewpoint on a recent earnings call: "For us, the risk of underinvestment is far greater than the risk of overinvestment." Investors aren't entirely buying it: Since their latest earnings reports, Microsoft's stock has fallen about 2% and Google has fallen 5%.
The cost of training AI has risen in recent years, but AI services from these two tech giants seem to be moving toward cheaper solutions. A Google spokesperson said the company's latest Gemini model, which businesses can use to automate customer service operations or summarize internal documents, is more powerful than its previous model but costs nearly half as much.
OpenAI's latest model, GPT-4o, is faster but 50% cheaper than its predecessor, GPT-4 Turbo. A female spokesperson told me that since 2022, the cost of accessing its model, calculated based on processed tokens (basically words processed by the language model), has dropped by 99%. "We're committed to continuing this trend," she added.
Reducing costs through technologies such as "sparsity" and "quantization" has been a major focus of recent conferences among AI scientists. A top executive at OpenAI's competitor Anthropic said the price of its model could drop to 25% of the current price within the next one to two years, and the company (which has raised $8.8 billion from investors including Google and Amazon (AMZN.O)) has cut the cost of building its latest model by half through new research approaches.
There are also signs that costs are coming down. In China, AI companies have been engaged in a price war for years, lowering the price of using generative AI thanks to a relaxed regulatory environment, lower labor costs, and government subsidies. For example, an AI startup called DeepSeek charges enterprise customers $0.14 per million tokens, while OpenAI's similar model charges $10.
Business users are also achieving cost benefits. Many have realized that they don't need the most powerful AI to boost employee productivity, so they are experimenting with open source models from companies such as Meta Platforms (META.O) or cheaper, slower small models. The most advanced AI tools may be needed for real-time inference with customer service chatbots, but analyzing customer calls to improve them? This can be done with less advanced technology.
Global asset management company Man Group Plc said the models it uses to summarize text for portfolio managers or reduce a day's worth of work to 30 minutes for other employees are indeed reducing costs, but the question of whether this price trend is sustainable in the future remains.
Silicon Valley has a history of subsidizing prices, with streaming platforms, ride-hailing apps, and cloud services all bearing the burden of profit pressure to expand market share. The goal is to survive the competition and eventually raise prices to achieve profitability. But for generative AI, the key problem is that its usefulness to a company's bottom line remains a broad question, and this is the main reason why Gartner predicts that 30% of the projects will be abandoned by the end of next year.
If this technology remains limited to providing chatbots and text summarization services, it may not be worth it even if the cost is reduced. This may be a problem that tech companies should pay more attention to beyond the issue of cost.