84.39.399.39.32

Support 24/7

0 Your Cart $0.00

Cart (0)

No products in the cart.

T-Mobile Learned That Giving AI More Information Can Make It Less Useful

Uber Burned Through Its 2026 AI Budget and Cut Costs by Redesigning How Engineers Use AI

AI coding tools promise to help engineers build software faster. But there is a problem companies are discovering the expensive way: the more people use AI agents, the faster the bill can grow.

T-Mobile Learned That Giving AI More Information Can Make It Less Useful

Uber experienced that problem directly. Its 2026 AI budget was used up within the first few months of the year, according to its CTO, Praveen Neppalli Naga, as reported by The Information.

Instead of simply asking engineers to use AI less, Uber changed how its tools work. It limited context, adjusted reasoning settings, optimized prompt caching, and redesigned how agents access internal information.

The result is a useful lesson for every company investing in AI. The next competitive advantage may not come from buying the most powerful model. It may come from making each AI request cheaper, faster, and more useful.

AI Demand Grew Faster Than the Budget

Uber's AI usage expanded rapidly. Between February and August, weekly agentic requests across its AI offerings increased 9.4 times.

Yet Uber's overall AI spending reportedly remained roughly flat from April, according to distinguished engineer Uday Kiran Medisetty.

That did not happen automatically. The company built systems to control how engineers use AI coding tools and how automated agents perform work.

The problem is not unique to Uber. In an October VentureBeat survey of 156 qualified AI infrastructure and compute respondents, 38% said their organizations had delayed, reduced, or canceled an AI project during the previous year because costs were higher than expected.

AI adoption is growing. The economics are still catching up.

Four Defaults Control the Cost of Every Session

Uber uses a managed configuration system called AIFX to install and configure interactive coding tools and control their settings.

The company tests different settings across employee groups, measuring both inputs and results. Engineers do not need to make every technical decision themselves.

Four controls are particularly important.

A 400,000-token context limit. Even models with million-token context windows are automatically compacted at 400,000 tokens. Compaction summarizes earlier conversation history so the agent can continue without repeatedly carrying an enormous amount of information.

Medium reasoning by default. Higher reasoning effort can consume more tokens. Uber found that some newer models perform better at medium effort, although the company did not identify the models or tasks behind that comparison.

Different prompt-cache durations. Uber keeps the main agent's prompt cache for one hour and sub-agent caches for five minutes. Reusing cached information can cost less than processing the same prompt again, but maintaining a cache also has a cost.

No automatic loading of MCP tool definitions. Uber has more than 100 tools available through its central gateway. Loading their descriptions directly into the initial context could consume 50,000 to 70,000 tokens before an engineer enters a request. Agents instead call individual commands when needed.

These decisions may look like small technical adjustments. At enterprise scale, small savings repeated across thousands of sessions can become significant.

The Numbers Show the Value of Cost Engineering

Holding the model constant to isolate optimization effects, Uber reported that cost per 1,000 requests fell almost 34% from its peak. Cost per session dropped 52% from its June peak.

Model selection also matters. Medisetty said choosing the right model reduced the cost of cleaning up stale feature flags by 90%.

That is an important distinction. Not every engineering task needs the most advanced model available. A simple maintenance task may be handled by a cheaper model, while complex reasoning can be reserved for more capable systems.

Uber is also trying to make its agent workloads less dependent on any single model provider. The goal is flexibility: move tasks between models based on performance, quality, and price.

Bigger Context Does Not Always Mean Better AI

Uber's approach also reveals a second problem. Giving an AI agent more information does not guarantee better answers.

T-Mobile discovered this when it loaded a large collection of customer-support documentation into one agent. The system became confused by the volume of material.

The company moved toward narrower context and specialized agents that handle specific questions.

Adobe's experience shows the opposite situation. For some marketing work, historical context is essential. An agent reviewing several years of holiday promotions may discover that free shipping performs better with loyal customers than a 20% Black Friday discount.

The lesson is simple: the right information depends on the job.

Uber Gives Agents a Map Instead of Everything

Uber's most interesting architectural decision is its AI Context Graph.

Rather than loading every internal document and system into an agent's initial prompt, Uber gives the agent a map of how the company works. The agent can then retrieve the relevant services, owners, datasets, historical queries, and source code when needed.

By October 28, the graph contained close to 200 million nodes and edges.

One example illustrates the difference. Using the graph, an agent identified the correct data table in 38 seconds. Without it, the same model and prompt took 20 minutes and 9 seconds, then incorrectly concluded that the data could not be queried.

This is not just a better search tool. It changes the way an agent navigates a large organization.

However, a knowledge graph is not automatically the right solution for every business. T-Mobile tried a similar approach for live customer calls but abandoned it because searching the graph took too long. It now uses a simpler cache to route common questions.

The architecture must fit the task.

The Business Model Is Shifting From AI Access to AI Output

Uber's longer-term goal is to move more work from interactive coding sessions into managed agents.

That gives the company greater control over model selection, tool access, context, and cost.

It also changes how productivity should be measured. Uber tracks design documents, pull requests, deployments, incidents per code change, reverted code, review times, and development cycle times.

Its self-healing continuous integration system offers another example. When automated tests fail, an agent proposes a fix. Initially, engineers approve these fixes through Slack. Once a particular type of fix demonstrates a reliable acceptance rate, Uber can allow it to be applied automatically.

This gradual approach reduces the risk of giving an AI agent too much authority too early.

Still, Uber has not published a complete financial return on its AI spending. Medisetty said the company has only recently begun seeing a meaningful trend in features shipped overall and per engineer, but provided no figures.

That missing financial evidence matters. More pull requests do not necessarily mean more business value.

The Competitive Advantage Is in the Infrastructure

The competition in AI coding is often framed around which model writes the best code.

Uber's experience suggests another battleground: the infrastructure surrounding the model.

Companies need systems that select the right model, control context, manage tool access, retrieve internal knowledge, and measure the quality of completed work.

The advantage may belong to organizations that build these capabilities into their daily operations rather than relying on individual engineers to manage every setting.

There is also a governance question. When a company controls AI defaults centrally, it can reduce waste and improve consistency. But it must also ensure that cost-saving settings do not quietly reduce quality or limit engineers' ability to solve difficult problems.

Uber says it uses A/B testing and quality guardrails. Its reported cost reductions are encouraging, but the source does not provide detailed results for each individual setting or an independently verified financial return.

What Small Businesses Should Do Before Their AI Bills Grow

Smaller companies do not need Uber's 200-million-node graph. They can start with four practical questions.

  • Are employees using expensive models for simple tasks?
  • Are long conversations carrying unnecessary context?
  • Are tools and integrations loading information that is never used?
  • Are AI results measured by completed work and quality, or simply by usage?

Start with one repeated task, such as code review, documentation, or routine maintenance. Measure the current cost and completion time. Then test a cheaper model, narrower context, or more focused workflow.

Keep human approval for risky changes. Track errors and rework, not just speed.

The goal is not to minimize AI spending at any cost. It is to make every dollar produce useful work.

The Next AI Advantage May Be Doing More With Less

Uber's experience challenges the idea that better AI simply means bigger models, longer context windows, and more tokens.

Its approach is more disciplined. Give agents the information they need, select models according to task difficulty, control repeated input costs, and increase autonomy only when results justify it.

The company has shown that AI usage can grow dramatically while unit costs fall. But it has not yet demonstrated a complete financial return across its engineering organization.

That is the next test.

For businesses everywhere, the message is clear: AI adoption is easy to celebrate. AI economics are much harder to master.

T-Mobile Learned That Giving AI More Information Can Make It Less Useful

Uber AI budget, AI coding costs, Claude Code settings, AI agent efficiency, prompt caching, context window limits, AI reasoning effort, Uber Context Graph, AI infrastructure, model selection, software engineering AI, AI cost optimization, managed AI agents, enterprise AI strategy

Community reaction: Discussion around Uber’s AI spending has focused on a practical concern: whether high monthly token bills translate into measurable engineering value. Some Hacker News commenters questioned how companies can justify spending thousands of dollars per engineer each month, while others pointed to long-running sessions, repeated context, and inefficient agent workflows as sources of waste. These are individual community opinions, not a representative survey.

Source boundary: The article’s company figures and examples come from the supplied text. The community note is separate external context. The key evidence gap remains Uber’s overall financial return from AI, which the source does not quantify.

Your experience on this site will be improved by allowing cookies Cookie Policy