Across enterprise deployments, this can translate to roughly $150–$250 per developer per month before optimization.

Those numbers may seem manageable individually. But at enterprise scale—thousands of engineers using the tool daily—they add up quickly.

The Uber budget surprise

Reports say Uber experienced exactly this problem after rolling out AI coding tools widely across its engineering organization.

According to multiple reports citing internal statements from the company’s CTO, Uber exhausted its entire planned 2026 AI tooling budget by April, only four months into the year.

The primary driver was explosive adoption of AI coding tools, particularly Anthropic’s Claude Code. Engineers reportedly adopted the tools much faster than finance teams had modeled, forcing leadership to revisit its spending assumptions.

The episode highlights a broader lesson: when thousands of engineers begin using token‑metered AI systems without strict usage controls, spending can grow far faster than expected.

Microsoft’s reported shift away from Claude Code

Microsoft appears to be making a different adjustment.

Reports say the company is cancelling most internal Claude Code licenses and shifting developers toward its own GitHub Copilot CLI tools.

The change reportedly comes only months after Microsoft expanded Claude Code access to thousands of employees, including engineers and product teams.

Importantly, this does not mean Microsoft is abandoning AI coding. Instead, it suggests companies may prefer:

• tools tightly integrated into existing developer platforms
• more predictable cost structures
• greater internal control over usage and billing

In other words, the debate is less about whether AI coding is useful and more about which tools make economic sense at scale.

The productivity debate is still unsettled

The cost question would be easier to answer if productivity gains were universally clear. But the research so far is mixed.

A randomized controlled trial by the research group METR examined experienced open‑source developers using frontier AI coding tools. The study found that developers completed tasks about 19% slower when AI assistance was allowed, even though they believed they were working faster.

Researchers suggested that verification and debugging overhead—checking AI‑generated code for correctness—can offset some of the time saved during code generation.

Other studies and industry data still show productivity gains in certain situations, especially for simple tasks or boilerplate code. But results vary depending on the developer’s experience level and the complexity of the project.

Why the “AI makes software cheaper” story is complicated

Three structural factors are colliding:

1. Explosive adoption. AI coding assistants have quickly become routine tools for developers across many organizations.
2. Usage‑based pricing. Token billing means costs scale directly with activity rather than seats.
3. Uneven productivity gains. Some developers move faster, but others spend more time reviewing and correcting AI output.

When these forces combine, companies may see a paradox: development feels faster, but the total cost of producing software can rise.

What enterprises are doing next

The current shift looks less like a retreat from AI and more like a transition from experimentation to cost governance.

Common responses emerging across the industry include:

• setting per‑team AI spending limits
• restricting the most expensive models
• optimizing prompts and context size to reduce token usage
• consolidating tools onto a single platform

In short, the first wave of enterprise AI coding adoption prioritized speed and experimentation. The next wave will likely focus on cost control and workflow design.

AI coding assistants are still spreading rapidly—but companies are now learning that scaling them across thousands of developers is not just a technical decision. It’s also a financial one.

Across enterprise deployments, this can translate to roughly $150–$250 per developer per month before optimization.

Those numbers may seem manageable individually. But at enterprise scale—thousands of engineers using the tool daily—they add up quickly.

The Uber budget surprise

Reports say Uber experienced exactly this problem after rolling out AI coding tools widely across its engineering organization.

According to multiple reports citing internal statements from the company’s CTO, Uber exhausted its entire planned 2026 AI tooling budget by April, only four months into the year.

The primary driver was explosive adoption of AI coding tools, particularly Anthropic’s Claude Code. Engineers reportedly adopted the tools much faster than finance teams had modeled, forcing leadership to revisit its spending assumptions.

The episode highlights a broader lesson: when thousands of engineers begin using token‑metered AI systems without strict usage controls, spending can grow far faster than expected.

Microsoft’s reported shift away from Claude Code

Microsoft appears to be making a different adjustment.

Reports say the company is cancelling most internal Claude Code licenses and shifting developers toward its own GitHub Copilot CLI tools.

The change reportedly comes only months after Microsoft expanded Claude Code access to thousands of employees, including engineers and product teams.

Importantly, this does not mean Microsoft is abandoning AI coding. Instead, it suggests companies may prefer:

• tools tightly integrated into existing developer platforms
• more predictable cost structures
• greater internal control over usage and billing

In other words, the debate is less about whether AI coding is useful and more about which tools make economic sense at scale.

The productivity debate is still unsettled

The cost question would be easier to answer if productivity gains were universally clear. But the research so far is mixed.

A randomized controlled trial by the research group METR examined experienced open‑source developers using frontier AI coding tools. The study found that developers completed tasks about 19% slower when AI assistance was allowed, even though they believed they were working faster.

Researchers suggested that verification and debugging overhead—checking AI‑generated code for correctness—can offset some of the time saved during code generation.

Other studies and industry data still show productivity gains in certain situations, especially for simple tasks or boilerplate code. But results vary depending on the developer’s experience level and the complexity of the project.

Why the “AI makes software cheaper” story is complicated

Three structural factors are colliding:

1. Explosive adoption. AI coding assistants have quickly become routine tools for developers across many organizations.
2. Usage‑based pricing. Token billing means costs scale directly with activity rather than seats.
3. Uneven productivity gains. Some developers move faster, but others spend more time reviewing and correcting AI output.

When these forces combine, companies may see a paradox: development feels faster, but the total cost of producing software can rise.

What enterprises are doing next

The current shift looks less like a retreat from AI and more like a transition from experimentation to cost governance.

Common responses emerging across the industry include:

• setting per‑team AI spending limits
• restricting the most expensive models
• optimizing prompts and context size to reduce token usage
• consolidating tools onto a single platform

In short, the first wave of enterprise AI coding adoption prioritized speed and experimentation. The next wave will likely focus on cost control and workflow design.

AI coding assistants are still spreading rapidly—but companies are now learning that scaling them across thousands of developers is not just a technical decision. It’s also a financial one.