- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,520
The new paradox of corporate AI hit large companies.
AI promised companies savings, but in large technology teams, the score began to grow faster than benefits. Microsoft, according to The Verge, began to cancel most of the direct licenses of the Claude Code and transfer engineers to GitHub Copilot CLI, although just six months ago it opened access to the tool for thousands of developers, project managers, designers and other employees.
The Claude Code quickly became popular inside Microsoft, but the scale of use has become a separate problem. The more active employees started AI for work tasks, the more noticeable the computing costs increased. The cancellation of licenses will not affect the deal between Microsoft Foundry and Anthropic: the agreement provides for an investment of up to $ 5 billion in Anthropic, access to Foundry customers to the Claude models and Anthropic’s commitment to buy Azure computing power for $ 30 billion.
A similar story happened at Uber. The company’s technical director, Pravin Neppali, Naga, told The Information that Uber has spent the entire budget on AI tools for programming for 2026 in just four months. Before overspending, the company itself pushed employees to use the neural networks more actively and launched internal team ratings on the level of application of AI.
New reports show an unpleasant paradox of corporate AI: individual requests may become cheaper, but the total account is growing due to mass implementation. Nvidia’s vice president of Deep Learning Applied Deep Learning recently told Axios that his team’s computing costs far exceed the cost of employees.
Companies continue to encourage employees to use neural networks as often as possible. In Meta*, one of the employees created an internal Caludeonomics rating, where employees were compared by the consumption of tokens Claude, and Amazon employees, according to the Financial Times, began to launch AI even for unnecessary tasks to artificially raise the use rates. Tokens are small fragments of text that the model processes at each request.
Goldman Sachs predicts that agent-generating AI can increase token consumption by 24 times by 2030. According to the bank, consumers and companies will spend up to 120 quadrillion tokens per month if AI agents massively enter the work processes.
Tokens themselves should be cheaper. Gartner expects that by 2030, the inference of a large language model with a trillion parameters will cost Genai suppliers more than 90% cheaper than in 2025. But the business may not feel this kind of savings: AI agents spend more tokens for each task, and service developers are not required to reduce prices for customers after their cost.
Senior analyst director of Gartner Will Sommer warned that product executives should not be confused by the reduction of the cheapening of conventional tokens with the availability of complex reasoning models. In other words, powerful AI can become cheaper for one operation, but the working task will still cost more because of the number of steps, checks and intermediate calculations.
Calculation complicates the plans of companies that want to supply AI agents next to each employee. The head of Nvidia, Jensen Huang, previously said that once for each employee of the company there may be 100 AI agents. If token consumption grows faster than the price of computing falls, digital assistants will bring businesses not only faster work, but also a much heavier score.
AI promised companies savings, but in large technology teams, the score began to grow faster than benefits. Microsoft, according to The Verge, began to cancel most of the direct licenses of the Claude Code and transfer engineers to GitHub Copilot CLI, although just six months ago it opened access to the tool for thousands of developers, project managers, designers and other employees.
The Claude Code quickly became popular inside Microsoft, but the scale of use has become a separate problem. The more active employees started AI for work tasks, the more noticeable the computing costs increased. The cancellation of licenses will not affect the deal between Microsoft Foundry and Anthropic: the agreement provides for an investment of up to $ 5 billion in Anthropic, access to Foundry customers to the Claude models and Anthropic’s commitment to buy Azure computing power for $ 30 billion.
A similar story happened at Uber. The company’s technical director, Pravin Neppali, Naga, told The Information that Uber has spent the entire budget on AI tools for programming for 2026 in just four months. Before overspending, the company itself pushed employees to use the neural networks more actively and launched internal team ratings on the level of application of AI.
New reports show an unpleasant paradox of corporate AI: individual requests may become cheaper, but the total account is growing due to mass implementation. Nvidia’s vice president of Deep Learning Applied Deep Learning recently told Axios that his team’s computing costs far exceed the cost of employees.
Companies continue to encourage employees to use neural networks as often as possible. In Meta*, one of the employees created an internal Caludeonomics rating, where employees were compared by the consumption of tokens Claude, and Amazon employees, according to the Financial Times, began to launch AI even for unnecessary tasks to artificially raise the use rates. Tokens are small fragments of text that the model processes at each request.
Goldman Sachs predicts that agent-generating AI can increase token consumption by 24 times by 2030. According to the bank, consumers and companies will spend up to 120 quadrillion tokens per month if AI agents massively enter the work processes.
Tokens themselves should be cheaper. Gartner expects that by 2030, the inference of a large language model with a trillion parameters will cost Genai suppliers more than 90% cheaper than in 2025. But the business may not feel this kind of savings: AI agents spend more tokens for each task, and service developers are not required to reduce prices for customers after their cost.
Senior analyst director of Gartner Will Sommer warned that product executives should not be confused by the reduction of the cheapening of conventional tokens with the availability of complex reasoning models. In other words, powerful AI can become cheaper for one operation, but the working task will still cost more because of the number of steps, checks and intermediate calculations.
Calculation complicates the plans of companies that want to supply AI agents next to each employee. The head of Nvidia, Jensen Huang, previously said that once for each employee of the company there may be 100 AI agents. If token consumption grows faster than the price of computing falls, digital assistants will bring businesses not only faster work, but also a much heavier score.
