AI tools costing higher than engineers

With ref to a news article 

https://www.livemint.com/companies/news/ai-was-supposed-to-cut-costs-microsoft-and-uber-are-finding-it-is-more-expensive-than-paying-human-employees-11779666290918.html

Analysis:

What Happened?

Big companies like and started using AI coding assistants heavily to improve developer productivity. But they discovered an unexpected problem:

The cost of running AI tools became extremely high — in some cases higher than paying engineers themselves.

The main reason is how modern AI systems are billed: per token.


Simple Explanation of “Token” Pricing

A token is a small piece of text processed by an AI model.

Examples:

  • One word may be 1–2 tokens.
  • A paragraph can contain hundreds of tokens.
  • Large code files or long chats may contain thousands of tokens.

AI companies charge based on:

  • Input tokens (what you send)
  • Output tokens (what AI generates)

So:

  • More prompts
  • Bigger files
  • Longer conversations
  • More AI agents working simultaneously

= much higher costs.


What Happened at Microsoft?

Rapid AI Adoption

Developers inside Microsoft quickly started using advanced coding AI tools such as Claude Code.

This improved productivity, but usage increased so fast that costs escalated sharply.

Microsoft Pulled Back

Microsoft:

  • Cancelled most Claude Code licenses.
  • Asked engineers to use GitHub Copilot CLI instead.
  • Tried to control rising AI spending.

Important Detail

Microsoft is still deeply partnered with :

  • Multi-billion dollar Foundry agreement.
  • Massive Azure compute commitments.

So this was not about abandoning AI. It was about controlling operational cost.


What Happened at Uber?

AI Budget Exhausted Quickly

reportedly used its entire 2026 AI coding tools budget in only four months.

Why?

Uber encouraged employees to use AI aggressively:

  • Internal competitions
  • Usage leaderboards
  • Incentives for adoption

This increased token consumption dramatically.


Why AI Costs Become Explosive

Traditional Software Pricing

Normal SaaS software often uses:

  • Per user pricing
  • Monthly subscription
  • Fixed enterprise contracts

Example:

  • ₹2,000 per employee per month.

Costs are predictable.


AI Pricing Is Different

AI pricing behaves more like cloud infrastructure.

The more employees use it:

  • The more GPUs are needed
  • The more tokens are processed
  • The more compute is consumed

This means:

  • Heavy usage directly increases cost.

The “Cloud Compute” Problem

AI models run on expensive GPU infrastructure, mainly from companies like .

NVIDIA VP Bryan Catanzaro reportedly said:

“Cost of compute is far beyond the costs of the employees.”

Meaning:

  • Running AI at large scale may cost more than human salaries.

This challenges the original assumption that AI would automatically reduce labor costs.


Why Falling Token Prices Don’t Solve the Problem

Many people assume:

  • “AI gets cheaper every year.”

This is only partially true.

Example

Suppose:

  • Token price drops by 50%. But:
  • Usage increases by 500%.

Total bill still rises sharply.

This is similar to:

  • Cheap mobile internet leading to massive video streaming usage.

Lower unit price often increases consumption.


Industry Terms Mentioned

“Tokenmaxx” at Amazon

reportedly encouraged maximum token utilization internally.

Meaning:

  • Employees were pushed to use AI extensively.

“Claudeonomics” at Meta

tracked AI usage economics internally under the nickname “Claudeonomics.”

This reflects how seriously companies now monitor:

  • Token usage
  • GPU consumption
  • AI ROI

The Bigger Concern: Agentic AI

What Is Agentic AI?

Agentic AI means:

  • AI systems that perform multi-step autonomous tasks.

Example:

  • Read documentation
  • Write code
  • Debug errors
  • Run tests
  • Retry failures
  • Generate reports

All automatically.


Why It Is Expensive

An AI agent may:

  • Make dozens or hundreds of model calls for one task.

That means:

  • Huge token consumption
  • Continuous GPU usage
  • Rapidly increasing cloud bills

Jensen Huang’s Vision

imagines a future where:

  • Every employee may have around 100 AI agents working for them.

Technically possible. Financially challenging.

Because:

  • More agents = exponentially more compute usage.

Key Risks for Companies

1. Runaway AI Costs

If usage is not monitored:

  • Bills can explode unexpectedly.

2. Employees Become Dependent

Once developers rely heavily on AI:

  • Removing tools reduces productivity.
  • Switching platforms becomes difficult.

This creates vendor lock-in risk.


3. Misleading Cost Assumptions

Lower token prices do not guarantee lower total spending.

Usage growth may outpace efficiency gains.


4. Scaling Problem

Small AI pilots may look affordable.

Enterprise-wide deployment can become extremely expensive.

Especially with:

  • Autonomous agents
  • Large context windows
  • Continuous workflows

Easy Analogy

Think of AI like electricity in a factory.

Early Assumption

“Machines will reduce labor costs.”

Reality

If machines run 24/7:

  • Electricity bills may become enormous.

Similarly:

  • AI models consume compute continuously.
  • More automation can increase infrastructure spending dramatically.

Main Takeaway

The industry is learning an important lesson:

AI economics are driven more by usage intensity than by model pricing.

Companies expected AI to mainly reduce employee costs. Instead, many are discovering:

  • Compute infrastructure
  • GPU usage
  • Token consumption

are becoming major operational expenses.

The challenge is no longer just:

  • “Can AI do the work?”

But also:

  • “Can we afford AI at enterprise scale?”


Comments

Popular posts from this blog

Airbus A320 — caused by a critical software bug

Tentative timelines and the extent of change due to AI and robotics across key sub-sectors in India

Beyond Google: The Best Alternative Search Engines for Academic and Scientific Research