Uber’s AI spending spree ends as cost controls take hold

By early 2026, Uber aggressively embraced agentic AI tools like Anthropic's Claude Code, rapidly expanding usage among engineers from 32% in February to nearly 95% by spring. This surge, fueled by internal incentives such as leaderboards ranking teams by AI token consumption, led to soaring AI token costs that exhausted Uber's entire 2026 AI budget within just four months, with some engineers incurring monthly AI tool bills as high as $2,000—far surpassing initial budget estimates.

Despite the enthusiastic adoption, Uber's leadership, including COO Andrew Macdonald, began expressing deep skepticism about the return on investment, highlighting the difficulty in linking increased AI token usage to tangible consumer-facing improvements or measurable productivity gains. Macdonald candidly noted the challenge in drawing a direct line between AI-driven code contributions—which by mid-2026 accounted for roughly 70% of committed code—and the delivery of useful features that enhance rider experiences, making it increasingly hard to justify the escalating AI expenditures.

This early phase of unchecked AI adoption, dubbed the 'tokenmaxxing era,' prioritized maximizing raw AI token consumption as a performance metric, inadvertently encouraging unsustainable spending without cost discipline. Uber's CTO Praveen Neppalli Naga acknowledged this period as a strategic misstep, noting that while the company initially pushed for widespread AI use, including creating usage leaderboards, the rapid budget burn-out forced a reevaluation. This shift marked a broader industry awakening to the need for balancing AI enthusiasm with fiscal responsibility and measurable value.

In response to these challenges, Uber implemented spending caps—such as a $1,500 monthly limit per employee on AI coding tools—and shifted its approach from restricting access to treating AI efficiency as an engineering problem rather than a mere budget issue. This strategic pivot away from tokenmaxxing towards optimizing AI usage efficiency reflects a maturation in enterprise AI deployment philosophy, aiming to control soaring costs while seeking clearer links between AI investments and productivity gains. Industry peers like Walmart, Amazon, and Cisco soon adopted similar cost-control measures, underscoring the widespread nature of this reckoning.

By early 2026, the soaring costs of AI usage forced companies like Uber and others to impose strict spending caps and governance frameworks to rein in runaway expenses. Uber, after burning through its entire 2026 Claude Code budget by April, introduced a $1,500 monthly cap per employee along with internal dashboards for usage tracking and approval processes for exceeding limits. This shift from unrestricted AI consumption to controlled spending reflects a broader industry reckoning, as companies recognized that unchecked usage—sometimes on trivial tasks like checking the weather or converting PDFs—led to exorbitant bills without proportional productivity gains.

The initial enthusiasm for maximizing AI token usage, dubbed the 'tokenmaxxing' era, gave way to a more measured approach focused on efficiency and accountability. Uber’s CTO Praveen Neppalli Naga articulated this transition, emphasizing that the next phase prioritizes how tokens are used rather than how many are consumed. By improving prompt caching, selecting more cost-effective models, and providing engineers with real-time visibility into AI costs, Uber quadrupled AI adoption while lowering per-token expenses. This engineering-driven governance model, treating token efficiency akin to cloud cost management, exemplifies how enterprises are balancing expanded AI usage with stringent cost control.

As agentic AI scaled, enterprises confronted new governance challenges due to the autonomous, multi-step workflows that dramatically increased token consumption—sometimes by fiftyfold—leading to unexpected budget overruns. Amazon’s experience, including a $1.8 million overrun on a never-launched project caused by inefficient AI model use, prompted the deployment of automated guardrails and sandboxed execution environments to enforce budget limits and prevent costly configuration errors. Platforms like AccuKnox’s AgentZ emerged to provide centralized administration, detailed audit logs, and role-based access controls, reflecting a market-wide shift toward robust governance frameworks essential for safely integrating AI agents into real business workflows.

Beyond technical controls, companies are increasingly emphasizing accountability and human-agent collaboration to ensure AI investments translate into meaningful productivity gains rather than unchecked spending. Leaders like Uber COO Andrew Macdonald have openly questioned the direct link between soaring AI costs and proportional product improvements, underscoring the need for governance frameworks that prioritize strategic use cases and cost-effective model selection. This cultural shift includes emerging roles such as 'AI operators' who oversee AI workflows, and a growing awareness that cost control is inseparable from organizational processes and employee behavior, moving enterprises from AI enthusiasm to disciplined, value-driven adoption.

By mid-2026, companies like Kilo pioneered intelligent AI model routing architectures that dynamically assign tasks based on the agent's current mode, such as planning or debugging, to over 500 specialized models. This tiered routing system, which includes top-tier models for complex tasks and cheaper tiers for routine work, has demonstrated substantial cost savings—Kilo reported a one-third reduction in average cost per request with 80-90% of tasks routed to less expensive models—without compromising output quality. The flexibility to update mode-to-model mappings frequently from centralized systems ensures that enterprises can continuously optimize routing strategies as models evolve.

The broader industry has embraced multi-tiered model selection strategies that balance cost and quality by assigning tasks to appropriately priced models. OpenRouter, processing 25 trillion tokens weekly and backed by over $120 million in funding, exemplifies this trend by providing access to hundreds of models from Anthropic, OpenAI, Google, and others, allowing users to set cost-quality preferences dynamically. Enterprises like Coinbase and McCarthy Building have reported cost reductions up to 60% in token usage and computing expenses compressed by as much as 97% through AI-driven routing layers that classify prompts and pin sessions to consistent models, preventing context loss and cache invalidation.

Emerging routing frameworks emphasize quality-aware dynamic escalation, where tasks initially assigned to cheaper models are validated and escalated to frontier models only upon failure, achieving cost reductions of up to 44% while maintaining output parity. This approach relies on building internal evaluation datasets tailored to specific workloads rather than generic benchmarks, ensuring routing decisions optimize both accuracy and cost. As Dan Biderman highlights, despite progress, routing remains a complex research challenge requiring continuous refinement to keep pace with evolving models and the increasing demand for token efficiency.

Leading tech giants like Microsoft and Google have operationalized multi-layer routing architectures that integrate semantic routing, GPU-aware load balancing, and policy-driven proxies to optimize AI agent calls on cloud platforms. Microsoft's RouteLLM, for instance, uses matrix-factorization trained on human preferences to route only about 26% of calls to the strongest models, saving up to 85% in costs while preserving 95% of GPT-4's quality. Similarly, Google's release of multiple Gemini models at varying price-performance points enables dynamic task assignment, reserving expensive frontier models for the hardest problems and routing bulk workloads to efficient, specialized models, thus making AI deployment economically sustainable at scale.

By mid-2026, enterprises like Uber have redefined software development by evolving into agentic software factories that deploy tailored, multi-agent AI architectures to automate complex, structured backend workflows such as invoice processing and IT service management. This transformation accelerates traditionally lengthy processes from months to weeks, with AI agents handling over 70% of code production, effectively doubling engineer output year-over-year while keeping humans in the loop for critical review and approval.

Uber’s success in scaling agentic AI hinges on a robust, long-term investment in foundational infrastructure like monorepos and build systems, complemented by a sophisticated ecosystem of six core building blocks—including Model Gateway, DevPods, Context Graphs, and the AI Assistant Cortana—that collectively streamline and automate the entire engineering lifecycle. This comprehensive setup enables the automation of over 250 manual tasks and the handling of nine million lines of code, drastically reducing engineering toil and accelerating development cycles across 12 global tech sites.

Crucially, Uber treats agentic AI adoption as an infrastructure and governance challenge, implementing a managed software factory model where every AI interaction is routed through secure, scalable gateways enforcing strict policies such as no PII leakage and sub-100ms guardrail latency. This layered governance, combined with intelligent cost control mechanisms—organizing AI usage into four tiers with increasing control over quality and expense—has allowed Uber to stabilize total AI spend since April 2026 despite a 9.4x surge in weekly agentic requests and a 7x increase in active users, ensuring sustainable scaling without runaway costs.

By mid-2026, multi-agent AI architectures have become foundational to enterprise AI strategies, seamlessly integrating smaller, task-specific models within both open-source frameworks like Langgraph and Pydantic agents and vertically integrated platforms from OpenAI, Anthropic, and AWS. This dual ecosystem reflects a pragmatic balance where no single approach dominates, but rather each serves distinct operational needs, enabling enterprises to tailor AI economics and governance to their unique contexts.

Enterprises are confronting a critical inflection point in AI financial governance as the transition from simple AI interactions to complex, multi-step autonomous agentic workflows has caused token consumption to surge fiftyfold, far outpacing task complexity growth. Uber’s experience—where agentic tool usage soared from 32% to 95% within months, prompting a $1,500 monthly API cost cap per employee—exemplifies the urgent need for strategic cost controls such as Microsoft’s FinOps-driven model routing, which allocates simpler tasks to economical models while reserving premium models for high-value operations.

Amid escalating costs and operational complexity, AI sovereignty has emerged as a strategic imperative, with leaders like Palantir’s Alex Karp warning against 'digital colonialism'—the risk of closed-source AI labs exploiting enterprise data to enhance competitor offerings—and Nvidia’s Jensen Huang advocating for open model weights to safeguard security and control. This growing emphasis on open-source models and data governance signals a broader industry shift from 'intelligence leasing' toward reclaiming autonomy, as epitomized by recent challenges faced by Uber and Figma, marking a decisive move toward sustainable, sovereign AI ecosystems.