Why Are Uber, Microsoft & Meta Cutting AI Spending Despite Rising AI Adoption?

AI was supposed to help companies get more work done for less money. But at Uber, employees embraced AI coding tools so quickly that the company used up its planned annual budget in just four months. Microsoft and Meta are also moving some employees away from Anthropic's Claude towards their own AI tools.

That sounds like a retreat from AI. It isn't quite one. These companies are trying to answer a question that becomes more urgent as AI spreads across the workplace: if the bill keeps growing, how much useful work is the company actually getting in return?

The distinction matters. A business can cut what it pays for one AI service while increasing its overall investment in AI. And Uber's experience now offers an example of how spending can be controlled without stopping adoption.

Why Did Uber Exhaust Its AI Coding Budget in Four Months?

Uber introduced Anthropic's Claude Code to its engineers in late 2025 to help them write, review and fix software code. However, as employees increasingly used AI coding tools, costs rose much faster than expected.

By April 2026, Uber had exhausted its entire annual AI coding budget, according to its chief technology officer, Praveen Neppalli Naga. The company did not disclose the budget's exact size.

To control spending, Uber reportedly introduced a $1,500 monthly limit per employee for each AI coding agent, with additional usage requiring approval.

The measures appear to have helped. By August, Uber reported that:

• Weekly AI agent requests had increased 9.4 times since February.
• Overall AI spending had remained broadly stable since April.
• Cost per AI session had fallen 52% from its June peak.

Uber's experience shows how quickly AI costs can rise when usage grows without spending controls. However, it also demonstrates that companies can increase AI usage while keeping costs under control.

Why Are Microsoft and Meta Reducing Their Use of Claude?

Uber is not alone in questioning the cost of externally supplied AI tools. An October 5 report by The Information described Microsoft and Meta pushing more internal work towards AI products they control.

The reported changes were significant:

Source: The Information, October 5, 2026. These are reported internal-use estimates and counts, not audited companywide AI spending figures.

Microsoft: Why pay an outside supplier for every task?

Microsoft sells AI products and provides access to models through its own ecosystem. It has a reason to offer customers a choice of models, including external ones. Internally, however, it can also encourage teams to use Microsoft-controlled tools where they are capable and economical.

That could reduce money flowing to external suppliers and give Microsoft more control over costs, product development and employee workflows. But reducing Microsoft's internal Claude bill does not mean Microsoft has ended its relationship with Anthropic. The Information specifically distinguished Microsoft's internal spending from growing customer use of Anthropic models through Microsoft services.

Meta: Build more of the tools employees use

Meta has been developing its own AI models and coding products. According to The Information, its internal tool MetaCode had more than 30,000 users, while Muse Code had more than 6,000 employee users at the time of reporting.

That offers Meta a different route to managing its AI bill. Instead of relying entirely on Claude Code for engineering work, it can move appropriate tasks to its own systems.

There is an important caveat, however: building an in-house model is not free. Meta still needs computing infrastructure, engineers and ongoing improvements. And because some Claude usage fell after layoffs, the entire decline cannot be attributed to employees willingly switching products.

So the competitive signal is meaningful, but narrower than a headline saying Microsoft and Meta are 'quitting Claude.' They are reducing some internal dependence, not necessarily severing a commercial relationship.

Are Companies Cutting AI Investment or Just Cutting Waste?

These developments are mostly evidence of tighter control over specific AI operating costs, not proof of an industrywide collapse in AI investment.

There are several ways a company can rein in its bill while continuing to use more AI. It can cap individual spending, require approval for unusually expensive tasks, send simpler work to cheaper models, reuse previously processed information and measure which automated tasks genuinely pay off.

It can also build more of its own software. That is what makes the Microsoft and Meta stories so relevant: both compete in AI while also purchasing technology from other AI companies. Their incentives differ from those of a smaller business that cannot afford to build sophisticated models.

Crucially, employee software bills and capital spending on AI infrastructure are different costs. A company may reduce what it pays for Claude Code licences or token usage while still spending heavily on data centres, chips and proprietary models. In fact, some in-house development may shift costs from an outside supplier to the company's own infrastructure rather than remove them entirely.

Uber's reported experience illustrates the upside of improving efficiency. By August, it was handling substantially more agent requests without a matching increase in total AI spending. That is a more useful result than either a blanket ban on AI tools or unlimited usage without cost controls.

What Does This Mean for Anthropic, OpenAI, Google and Enterprise AI?

For the companies selling AI models, the implication is not that enterprise demand has disappeared. It is that winning widespread adoption may no longer be enough to keep every customer spending at the same rate.

Anthropic: Claude's popularity among software developers is an advantage, but large customers may try to limit premium-model use, negotiate more aggressively or move suitable work to cheaper alternatives. Microsoft and Meta's reported reductions concern internal use; they do not by themselves show a collapse in Anthropic's total revenue or external customer demand.

OpenAI: It faces the same pressure to show that more capable models justify their cost. Companies can compare models task by task, giving customers greater bargaining power. OpenAI could also benefit when businesses seek alternatives to Claude, but this is a competitive possibility, not a proven transfer of spending.

Google: Its Gemini models and cloud services are additional options for enterprises considering price, performance and integration. Google may gain opportunities when customers diversify suppliers. Yet lower costs on one workload do not guarantee that customers will switch their entire AI systems.

Microsoft and Meta: Better in-house tools may give them more control over costs and technology. They also take on the expense and risk of building and maintaining those systems. Investors should judge the outcome through productivity, cloud or product economics and eventual earnings, not only through lower payments to an outside AI provider.

This creates a tension for the AI industry. Better efficiency can reduce revenue from an individual request while making AI affordable for many more requests. Whether suppliers ultimately win depends on how much usage expands, what customers are willing to pay and how cheaply models can be run.

What Should Investors Take Away?

The most revealing fact is not that Uber ran through its AI coding allocation in four months. It is what happened next: the company reported stabilizing the bill while AI agent activity kept rising. That is a reminder that the early economics of AI adoption are not necessarily permanent.

Microsoft and Meta reveal the other side of the equation. Major customers increasingly have alternatives, including software and models they develop themselves. AI suppliers will have to compete not only on impressive performance but also on the cost of producing reliable, useful work.

For investors, the sensible question is no longer simply, 'How fast is AI usage growing?' It is 'Who earns money when that usage grows, and who pays the bill?' The answer will help determine which AI companies build durable businesses as enterprise buyers become more selective.