We spent the money, but we can't prove it made anything better
En las oficinas de Uber, una pregunta incómoda ha comenzado a resonar con fuerza: ¿puede una empresa justificar gastos masivos en inteligencia artificial cuando no existe una línea clara entre lo que se consume y lo que el usuario final experimenta? La compañía agotó su presupuesto anual para herramientas de codificación con IA antes de que el año avanzara lo suficiente, y sus propios ejecutivos admiten que la relación entre tokens consumidos y mejoras reales permanece opaca. Este momento no es solo una crisis contable; es un espejo que la industria tecnológica entera está obligada a mirar.
- Uber consumió todo su presupuesto 2026 para Claude Code sin poder demostrar qué mejoras concretas recibieron sus usuarios a cambio.
- El COO Andrew Macdonald describió una tensión interna explosiva: los ingenieros usan herramientas de IA como si fueran gratuitas, pero la empresa absorbe costos que nadie sabe justificar.
- El CTO anunció el agotamiento del presupuesto en lo que internamente se llamó un 'momento de cabeza explotada', forzando una conversación que la compañía había estado evitando.
- Duolingo enfrentó un problema paralelo: al medir el uso de IA como métrica de desempeño, los empleados usaban las herramientas para cumplir el indicador, no para trabajar mejor, y la empresa tuvo que dar marcha atrás.
- Uber ahora busca redefinir qué métricas importan —velocidad de entrega, calidad del código, reducción de incidentes— antes de que el gasto en IA se convierta en un fin en sí mismo.
Dentro de las oficinas de Uber, una crisis silenciosa tomó forma cuando la compañía agotó su presupuesto completo de 2026 para Claude Code, un asistente de codificación basado en inteligencia artificial. Lo perturbador no fue el gasto en sí, sino la incapacidad de explicar qué había producido: nadie podía trazar una línea directa entre los tokens consumidos y algo que los usuarios pudieran ver o sentir.
Andrew Macdonald, el director de operaciones, se encontró en el centro de este dilema. Al preguntar a ingenieros senior qué funcionalidades nuevas habían surgido del uso intensivo de IA, las respuestas eran vagas. No existía ninguna métrica que demostrara que duplicar el consumo de tokens había generado, por ejemplo, un aumento proporcional en funcionalidades útiles para conductores o pasajeros. Cuando el CTO Praveen Neppalli Naga anunció que el presupuesto estaba agotado, el episodio se convirtió internamente en un caso de estudio sobre cómo los costos se disparan cuando la experimentación ocurre sin supervisión.
La reflexión de Macdonald fue simple pero devastadora para la narrativa pro-IA: un ingeniero no ve la factura cuando usa una herramienta de inteligencia artificial, pero la empresa sí. Y si no puedes demostrar que esa herramienta mejoró la productividad o el producto, estás quemando dinero en una corazonada. El umbral de justificación tenía que ser más alto.
Uber no estaba sola. En Duolingo, el CEO Luis von Ahn había revertido un sistema que medía cuánto usaban los empleados la IA como indicador de desempeño. El resultado había sido predecible: la gente usaba las herramientas para alcanzar la métrica, no porque las ayudara a trabajar mejor. Su conclusión fue directa: 'Lo más importante es que hagas tu trabajo lo mejor posible. Si la IA no puede ayudarte, no voy a obligarte a usarla.'
Para Uber, la pregunta que queda abierta es si aprenderá esta lección a tiempo, o si seguirá invirtiendo en sistemas cuyo valor real permanece sin demostrar. El debate interno apenas comienza, y las apuestas son claras: encontrar métricas que prueben impacto real, o admitir que la compañía ha estado persiguiendo un espejismo.
Inside Uber's offices, a quiet crisis was taking shape. The company had burned through its entire 2026 budget for Claude Code—an AI coding assistant—and nobody could quite explain what it had gotten in return. This wasn't a dramatic failure. It was something more unsettling: a company spending heavily on artificial intelligence without being able to draw a straight line between that spending and anything users could actually see or feel.
Andrew Macdonald, Uber's chief operating officer, found himself in the middle of this reckoning. In conversations with senior engineers, he kept bumping into the same problem. The company was consuming tokens—the computational units that power AI services—at an accelerating rate. But when he asked what new features or capabilities had materialized as a result, the answer was murky. There was no clear metric showing that a doubling of token consumption had produced, say, a 25 percent increase in useful functionality for riders or drivers. The relationship between spending and output simply didn't exist on paper, even if some invisible productivity gains might be happening somewhere in the codebase.
The moment of reckoning came when Praveen Neppalli Naga, the company's chief technology officer, announced that the Claude Code budget was exhausted. According to reports, the news triggered what executives described internally as a "head-exploding moment"—a sudden, unavoidable confrontation with the question that had been lurking beneath all the AI enthusiasm: What exactly were they paying for? The incident became a case study inside the company on how costs spiral when experimentation becomes widespread and unmonitored.
Macdonald's core insight was simple but damaging to the AI-first narrative. When an engineer uses an AI tool at work, it feels free to them. They don't see a bill. But the company does. And if you can't prove that the tool made them more productive, or that it improved the product, or that it enhanced the user experience, then you're just burning money on a hunch. The threshold for justification had to be higher. Using more AI models wasn't enough. You had to show impact—real, measurable impact—or the spending couldn't be defended.
The pressure wasn't confined to token consumption. CEO Dara Khosrowshahi had already signaled a broader slowdown in hiring, part of a financial rebalancing act between AI investment and other company priorities. The two forces were in tension: the company wanted to experiment with AI, but it also needed to control costs and prove that the experiments were working. That meant the internal debate had to shift. Instead of asking how many tokens were being consumed, or how many engineers were using AI tools, Uber needed to ask which metrics actually mattered: software delivery speed, code quality, incident reduction, or measurable improvements in the service itself.
Uber wasn't alone in this struggle. At Duolingo, CEO Luis von Ahn had recently walked back a performance evaluation system that measured how much employees were using AI. The company had discovered something uncomfortable: when you make AI usage itself the goal, people use AI tools to hit the target, not because the tools actually help them do better work. Ahn explained the reversal plainly: the company had been rewarding the wrong thing. "The most important thing in your performance is that you're doing your job as well as possible," he said. "If AI can't help, I'm not going to force you to use it." Duolingo's course correction exposed a tension rippling through the entire technology sector. Measure AI adoption as a success metric, and you inflate consumption without guaranteeing quality, speed, or user benefit. The tools become an end in themselves rather than a means to an end.
For Uber, the question now was whether the company would learn the same lesson, or whether it would continue pouring resources into AI systems while struggling to justify the expense. The internal debate was just beginning, and the stakes were clear: either find a way to prove that AI spending was delivering real value, or admit that the company had been chasing a mirage.
Notable Quotes
That relationship still doesn't exist. We can't prove with metrics that the spending translates to 25 percent more useful functionality for users.— Andrew Macdonald, Uber COO
The most important thing in your performance is that you're doing your job as well as possible. If AI can't help, I'm not going to force you to use it.— Luis von Ahn, Duolingo CEO
The Hearth Conversation Another angle on the story
So Uber spent an entire year's budget on Claude Code and couldn't explain what it got for the money. How does that even happen at a company that size?
It happens because the cost feels invisible to the people using it. An engineer runs a query through Claude Code, gets an answer, moves on. They never see a bill. But the company is paying for every token consumed, and when you add up thousands of engineers running thousands of queries, the bill becomes enormous. The problem is nobody was tracking whether those queries were actually making the product better.
But surely someone was monitoring the spending?
They were, but they were monitoring the wrong thing. They were watching token consumption go up, which is easy to measure. What they couldn't measure was whether that consumption was translating into features users cared about, or faster software delivery, or fewer bugs. That's much harder to quantify. So you end up in this weird position where you can say "we spent X on AI" but you can't say "and it made the product Y percent better."
The COO said the relationship "still doesn't exist." That's a pretty damning statement.
It is. What he was really saying is: we've been spending money on faith. We believed that more AI would make us smarter, faster, better. But when you actually sit down and try to prove it with numbers, the evidence isn't there. And once you admit that, you can't keep spending the same way. You have to change how you measure success.
Duolingo tried measuring AI usage itself and had to reverse course. Is that what Uber is heading toward?
Possibly. Duolingo learned that if you reward people for using AI, they'll use it whether or not it helps. The tool becomes the goal instead of a tool toward a goal. Uber seems to be asking the same question now: are we spending on AI because it works, or because we're afraid of falling behind if we don't? That's a much harder question to answer honestly.
What would a better metric look like?
Something tied to actual outcomes. Did code quality improve? Did we ship features faster? Did user complaints go down? Did the service become more reliable? Those are things you can measure. Token consumption is just noise unless it's connected to one of those real outcomes.