Summary of Key Points
From the beginning of the year, when large companies in China and the US were "forced to allocate AI usage and compete in token consumption," they have now shifted to "limiting tokens and being more frugal." On the surface, this seems like a 180-degree policy change, but in essence, it reflects a shift in the narrative around AI from one of "technological faith" to one of "business logic." Companies have realized that the investment in AI (both in terms of token costs and hidden expenses) far exceeds the actual returns. The myth of super individuals who can leverage AI to achieve exceptional results has also collapsed due to cost and risk concerns, leading them to adopt more refined management practices for AI usage.
Detailed Analysis
#### 1. Stop Competing in Token Consumption: More Usage Doesn't Equal Higher Efficiency
Previously, large companies used token consumption as a key performance indicator (KPI), believing that the more employees used AI, the more innovative and efficient they were. However, this turned out to be an illusion. Employees often used AI to generate lengthy, useless documents or repetitive code, significantly increasing token costs without any tangible benefits for company revenue or research and development efficiency. For example, Microsoft's early Copilot service saw heavy users incurring costs far exceeding the monthly subscription fee of $30; a SaaS company added AI features to its products for free, but the resulting increased usage (converting 50-word memos into 1,500-word plans) cost $8 per user per day ($240 per month), leading to bankruptcy just two months later when token costs soared from $3000 to $160,000.
Studies from Stanford and MIT have shown that 95% of AI pilots in companies did not generate measurable financial returns, with only a few truly creating value.
#### 2. AI Mistakes Come at a Cost: The Hidden Expenses
Token costs are just the tip of the iceberg; more detrimental are the customer complaints and public relations crises caused by AI's "mistakes" (e.g., misleading promises). For instance, Air Canada's AI customer service promised to refund ticket prices in cases of bereavement, but when a passenger sued after being denied a refund, the company lost the case because the court ruled that AI, as part of the company, was legally responsible for its actions. Another example is DPD's AI customer service, which wrote offensive poems about the company and caused a widespread online backlash, forcing the company to shut down the system immediately.
These hidden costs (compensations, PR efforts, and repairs) are much higher than the labor savings that AI might provide, prompting companies to invest in compliance teams to monitor and manage AI properly.
#### 3. AI Is Actually More Expensive Than Humans: The Cost Comparison
Many people think AI is cheaper than hiring humans, but this is not always the case. While a human brain can perform complex tasks with just a meal and a cup of coffee (20 watts of energy), the Transformer architecture used in AI makes each interaction extremely costly:
- Attention Mechanism: When you provide 100 words, AI must compare each word with every other word from the entire dataset (10,000 times), increasing the computational load with more text.
- Word-by-Word Generation: AI processes each response individually, requiring it to review the entire conversation history for each new input, similar to reviewing an entire book for each response.
- Parameter Transfer: Using AI involves moving billions of data points between storage and memory, which is a resource-intensive process.
In critical scenarios (such as contract signing or compliance tasks), the combined costs of AI errors and compliance efforts can exceed those of hiring dedicated personnel. CFOs have realized that blindly using AI is not always more cost-effective than hiring humans.
#### 4. The Myth of Super Individuals: Most People Still Work with AI
The idea that AI can enable a single person to perform the work of an entire team was once promoted as a revolutionary concept, but it has proven to be a myth:
- Division of Labor: AI hasn't eliminated the need for specialized roles; instead, companies now need to hire more expensive AI architects and security experts.
- Costs: Using AI to replace teams leads to skyrocketing token costs that most individuals cannot afford.
- Risk Management: Small companies cannot afford the losses associated with AI failures (e.g., system downtime resulting in millions of dollars in damages). Modern companies rely on legal and PR teams to mitigate these risks, and super individuals (such as self-media creators or independent developers) are the exception.
In conclusion, large companies are limiting token usage not because AI has become ineffective, but because they are returning to a more pragmatic business approach. Any technology must be evaluated based on its cost-benefit ratio. AI is not a "free lunch"; its costs and risks must be carefully considered by both businesses and individuals. The myth of AI revolutionizing everything ultimately boils down to practical, financial considerations.