AI Turns Out to Be Expensive, Microsoft and Uber Start Recounting Its Use

Microsoft (photo: dok. unsplash)

JAKARTA - Microsoft and Uber have begun to limit the use of a number of artificial intelligence or AI services after their costs have ballooned. This development raises a fundamental question, if the cost is expensive, will AI really quickly replace many human jobs?

Yahoo Finance, quoted Thursday, June 4, reported that Microsoft stopped some of its engineers from accessing Claude Code, Anthropic's AI programming tool. The decision was made because the bill for its use was too high.

Claude Code is an AI tool that helps create and improve computer code. Microsoft previously gave access to thousands of employees, including engineers, product managers, designers, and non-technical employees.

However, the use of the tool spread quickly. The cost also jumped.

Microsoft then moved the engineers in the Experiences and Devices group to GitHub Copilot CLI. This group oversees Windows, Microsoft 365, Outlook, Teams, and Surface.

GitHub Copilot CLI is Microsoft's AI programming tool that runs via the command line. In short, this is a version that is closer to the work kitchen of technicians and is considered cheaper for companies.

This step does not mean that Microsoft is retreating from AI. Microsoft's major cooperation with Anthropic continues. The investment of up to US $ 5 billion or around Rp. 89 trillion and Anthropic's commitment to buy Azure computing capacity worth US $ 30 billion or around Rp. 534 trillion are also unchanged. The calculation is if US $ 1 is assumed to be Rp. 17,800.

Uber is facing a similar problem. According to Yahoo Finance, Uber's Chief Technology Officer said the 2026 budget for Claude Code and Cursor was exhausted in just four months.

Cursores são ferramentas de IA para ajudar os programadores a escrever código. O caso da Uber mostra o mesmo problema: a IA acelera o trabalho, mas a conta pode ser mais rápida.

Bryan Catanzaro, Vice President Applied Deep Learning Research Nvidia, gave a more detailed overview. He told Axios that the cost of computing for his team "far exceeds the cost of employees".

Computing is the computer's computing power to run AI. The more complex the model, the greater the need for chips, electricity, servers, and data centers.

Here, the issue of AI is not just about sophistication. Ongkos turns out to be a problem.

In an analysis published on Substack, Shanaka Anslem Perera assessed that OpenAI faces deeper cost problems. According to his calculations, OpenAI spends around US$3.30 or around Rp58,740 to generate US$1 or around Rp17,800 in revenue.

Perera said OpenAI is expected to burn around US$11.5 billion or around Rp204.7 trillion per quarter. In a year, its value is around US$46 billion or around Rp818.8 trillion.

Meanwhile, OpenAI's operating income is estimated at around US$20 billion or around Rp356 trillion.

He argues that OpenAI's problem is not just burning money to grow. The problem lies in the cost structure.

According to Perera, Google has a stronger position because it uses its own AI chips, namely TPU or Tensor Processing Unit. TPU is a chip made by Google to train and run AI models.

With TPU, Google can train and run Gemini without relying entirely on Nvidia chips or other parties' infrastructure. The cost is considered lower.

Perera quoted SemiAnalysis' calculations that said Google's TPU infrastructure had 30 to 44 percent lower ownership costs than Nvidia-based systems.

On the other hand, OpenAI is said to bear a layered cost. Every request for ChatGPT runs on chip, cloud, and operational costs. If the scale of use increases, the cost pressure also increases.

That's what makes the AI competition no longer as simple as who has the smartest model. Companies also have to be able to run it at a reasonable cost.

Google, according to Perera, could be more free to offer Gemini at a low price or bundle it into Workspace. OpenAI would be harder to match without increasing losses.

For companies, the Microsoft and Uber cases are enough to explain, if AI can help productivity. But its use must still be calculated.

AI may remain a threat to some jobs in the long run. However, for now, the high cost is a real brake or problem. Machines can be smart. The bill still has to be paid.