If Uber burned through its entire 2026 AI budget by mid-April, why are executives still celebrating AI adoption before proving AI economics?
That is the uncomfortable question raised by Uber’s Claude Code rollout, according to CryptoBriefing. My view is blunt: this is not just an Uber budgeting miss. It is a warning flare for enterprise AI. Spending has outrun measurement, governance, and accountability. The next phase of AI will not reward companies that deploy the most tools. It will reward companies that can prove those tools improve margins, products, or risk controls.
Why did Uber’s AI usage explode before the ROI case did?
Uber rolled out Anthropic’s Claude Code to roughly 5,000 engineers in December 2025. The uptake was not slow, cautious, or experimental. It was a surge.
Usage of agentic coding features jumped from 32% in February to 84% by March 2026. By then, 95% of Uber’s engineers were using AI tools monthly, and nearly 70% of code commits involved them in some capacity.
Those numbers look like a software adoption dream. They also helped drain the budget.
CTO Praveen Neppalli Naga confirmed that Uber’s full annual AI budget was exhausted by mid-April. Per-engineer monthly API costs reportedly ballooned to between $500 and $2,000, far above internal forecasts.
The hardest part for Uber is not the bill. It is the missing bridge between the bill and the product.
The hard lesson is not that engineers failed to use the tool. It is that Uber still has to show, in measurable terms, how those adoption stats translate into more useful consumer features, faster delivery, or durable cost savings.
That should land hard in every CFO’s office. High usage is not the same as high return.
| Uber AI metric | What it proves | What it does not prove |
|---|---|---|
| 95% monthly engineer usage | Engineers adopted the tools | Customers got better products |
| 70% of code commits involved AI | AI entered the development workflow | Code quality or velocity improved enough to justify cost |
| $500 to $2,000 monthly API costs per engineer | Token consumption became expensive | The spending created matching value |
| Annual AI budget exhausted by mid-April | Forecasts missed actual usage | The rollout was economically sound |
If adoption is real, why is value still fuzzy?
The obvious defense is that Uber’s engineers used AI because it helped them work. That may be true. But the measurable link to better consumer features is not clear yet.
That distinction matters.
AI costs can scale in ways old software budgets do not. A per-seat SaaS bill is boring. Token-driven usage is not. Engineers can run more prompts, longer contexts, agentic workflows, test generation, refactoring jobs, and code review loops. Each use may be defensible on its own. Together, they can turn a forecast into fiction.
Once a tool becomes part of daily workflow, cutting it back gets politically harder. Teams begin to treat it as infrastructure, not experimentation. The finance team then faces a choice it dislikes: approve more spend without clean ROI evidence, or restrict tools employees now see as necessary.
That is the trap. AI can become operationally sticky before it becomes financially proven.
When did enterprise AI start looking like a venture bet?
Many large companies have treated AI like a venture portfolio: spend now, tolerate waste, assume the breakthrough comes later. That logic belongs in venture capital. It is dangerous inside mature enterprises.
Uber’s 2025 R&D spending hit $3.4 billion, up 9% year-over-year, according to the source material. So this is not a company too small to absorb experimentation. The issue is not whether Uber can spend on AI. The issue is whether AI spending is being judged with the same discipline as other mission-critical systems.
A pricing engine would not get unlimited budget because people use it. A fraud tool would not survive on adoption metrics alone. A payments system would be judged by cost, uptime, loss reduction, conversion, and failure modes.
AI should face the same standard.
That means:
- Milestones: Define what success means before rollout.
- Unit economics: Track cost per workflow, not just total spend.
- Accountability: Assign budget ownership to the teams consuming AI.
- Payback logic: Tie spend to productivity, revenue, risk reduction, or customer impact.
- Kill criteria: Shut down pilots that cannot prove value.
Enterprise AI cannot stay in demo mode forever.
Who benefits when AI urgency turns into blank checks?
AI vendors benefit when usage rises. Enterprise buyers benefit only when usage converts into value. That incentive mismatch is now impossible to ignore.
If a vendor’s model charges by consumption, the customer carries the risk of runaway demand. The more employees experiment, automate, and run agents, the more the invoice grows. That can be fine when the output is measurable. It is a problem when the spend rises faster than the proof.
Fear makes this worse. No executive wants to be accused of moving too slowly on AI. That anxiety can push companies into broad rollouts before procurement understands the unit economics.
This is where boards and CFOs need to get less dazzled and more annoying. They should demand clearer pricing, real-time usage visibility, model performance benchmarks, and exit flexibility before expanding commitments.
Related MLXIO reading: Gemini Takes Over Google I/O 2026 — and Your Workflow and $38M Fight Dies as Musk's OpenAI Lawsuit Runs Late.
Can Uber really afford to slow down AI investment?
The strongest counterargument is fair: Uber operates in a data-heavy business. AI can matter in engineering, support, fraud detection, marketplace matching, pricing, routing, and other operational systems. Underinvesting could be more costly than overspending if automation eventually cuts expenses or improves service quality.
That argument deserves respect. The answer is not to starve AI budgets.
The answer is to stop confusing investment with indiscipline.
Uber’s case does not prove Claude Code is useless. It does not prove AI coding tools fail. It proves that adoption can outrun financial controls, and that usage dashboards can create a false sense of progress.
There is a major difference between “engineers are using this” and “this is producing durable operating gains.” Uber’s situation shows why more companies should locate that line before the invoice arrives.
Why should crypto investors care about Uber’s AI bill?
Crypto markets should read Uber’s AI overspend as a warning about infrastructure hype.
AI enthusiasm has fed interest in decentralized compute, AI tokens, data marketplaces, and GPU-linked narratives. The logic is simple: if enterprise AI demand keeps rising, then infrastructure tied to compute demand should benefit.
Maybe. But Uber’s lesson cuts both ways.
If large enterprises start questioning AI cost effectiveness, investors will demand harder evidence from AI-adjacent crypto projects too. “AI” on a pitch deck will not be enough. Usage will not be enough. Token volume will not be enough.
The stronger projects will need to show:
- Real demand: Not circular activity or subsidized usage.
- Verifiable economics: Revenue, costs, and margins that make sense.
- Clear customers: Buyers who need the service beyond speculative cycles.
- Sustainable pricing: Not just a story about compute scarcity.
Crypto should internalize the same lesson Uber is now confronting: adoption is not value creation.
What should boards demand before the next AI spending wave?
Boards should require AI profit-and-loss statements before approving the next expansion.
Not glossy strategy decks. Not “everyone is using it” charts. Actual financial reporting.
Every serious AI program should track total spend, cost per workflow, productivity impact, revenue contribution, risk reduction, vendor dependency, and security exposure. If a tool touches engineering output, the company should know whether it reduces cycle time, increases defects, improves test coverage, or merely shifts work into a more expensive interface.
CFOs should also require kill criteria. AI pilots usually have launch criteria: budget, sponsor, vendor, user group. They need shutdown criteria too. If a project cannot show measurable value after a defined period, it should lose funding.
Uber’s budget blowout is useful because it punctures a lazy assumption. AI adoption does not automatically mean AI transformation. Sometimes it means the meter is running and nobody is watching closely enough.
The next phase of AI leadership will belong not to the companies that spend the fastest, but to those that can prove every model earns its place.
The Bottom Line
- Uber’s rapid AI adoption shows how quickly enterprise AI costs can outrun budgets.
- High usage metrics do not automatically prove productivity gains or customer value.
- CFOs may demand stricter ROI measurement before approving broader AI spending.
