The AI Reality Check: Enterprises Grapple with ROI as "Tokenmaxxing" Costs Mount

The buzz around Artificial Intelligence has been deafening, with promises of unprecedented efficiency and innovation. Companies, eager to stay ahead, jumped headfirst into AI adoption, often encouraging employees to use AI tools as much as possible – a trend dubbed "tokenmaxxing." However, as NEA partner Tiffany Luck points out, many enterprises are now facing a harsh reality: figuring out the actual return on investment for these costly AI endeavors. The initial enthusiasm is giving way to a more pragmatic, cost-conscious approach as the bills for extensive AI usage come due.

The "Tokenmaxxing" Phenomenon and Its Steep Price Tag

Earlier this year, "tokenmaxxing" became a popular internal strategy in Silicon Valley. CEOs pushed employees to maximize their AI usage, with internal leaderboards often gamifying token consumption, turning it into a metric for perceived productivity. This aggressive adoption, however, quickly led to unforeseen financial strain.

Uber, for instance, reportedly exhausted its entire 2026 AI budget within just four months. The ride-hailing giant had rolled out Claude Code to approximately 5,000 engineers in late 2025, with monthly active engineers reaching 84-95% and about 70% of committed code being AI-generated. Heavy users were costing Uber between $500 and $2,000 per month per engineer. This rapid expenditure forced Uber to impose a $1,500 per month cap per employee for each AI coding tool. Uber's COO, Andrew Macdonald, openly questioned the direct link between this high AI usage and the creation of more useful consumer features, highlighting the difficulty in justifying these internal AI costs.

Meta also experienced similar issues. The company had an internal leaderboard called 'Claudeonomics' that tracked AI token usage for over 85,000 employees. In one 30-day period, total token consumption exceeded 60 trillion, with an estimated cost of $9 billion if charged at public Anthropic API prices. This leaderboard, which fostered a "tokenmaxxing" culture with titles like 'Token Legend,' was eventually taken down after public scrutiny, as Meta began to realize that high token consumption didn't automatically equate to productivity. Meta is now planning to manage AI tokens more tightly with budgets and allocations starting in 2027, and has developed an 'AI Gateway' to track usage and spending.

Even Microsoft, a company heavily invested in AI, has started to cut Claude Code licenses within its core product teams, including the Experiences and Devices group responsible for Windows, Microsoft 365, and Surface. Engineers are being directed to migrate to GitHub Copilot CLI, a Microsoft-owned tool. This decision, coming less than six months after Claude Code's internal rollout in December 2025, was driven by soaring costs, with per-engineer expenses ranging from $500 to $2,000 per month. Microsoft's move signals an end to the "experimental phase" where companies were willing to absorb arbitrary token costs for learning.

The Disconnect: Why AI ROI is So Elusive

The struggle to prove AI ROI is not just about unexpected bills; it's a multi-faceted problem. According to IBM, key challenges in enterprise-wide AI adoption for 2026 include data quality, governance, security, skills gaps, workflow integration, and most importantly, proving ROI and justifying costs. A report from MIT found that 95% of companies are seeing no real return on their GenAI spend, despite the US crossing $40 billion in enterprise expenditure.

Several factors contribute to this disconnect:

• Unclear Strategy and Use Cases: Many organizations deploy AI without a clear understanding of specific business goals or measurable outcomes. This leads to AI being treated as a "shiny object" rather than a tool to solve concrete problems.
• Underestimation of Total Cost of Ownership (TCO): The cost of LLMs extends far beyond just API fees. It includes significant expenses for infrastructure (GPUs), operations, data processing, model licensing, cloud computing, integration work, cybersecurity, and system maintenance. For instance, training a model like OpenAI's GPT-3 required compute resources valued at a minimum of $5 million per training cycle.
• Talent Gap: There's a shortage of specialists in LLM operations (LLMOps), with salaries ranging from $100,000 to $268,000, presenting a significant hidden cost.
• Fragmented Systems and Workflows: Many enterprises operate with fragmented and siloed data environments, making it difficult to integrate AI effectively into existing business processes. This means AI often sits alongside existing complexity, adding another layer to manage, rather than streamlining operations.
• Difficulty in Measuring Intangible Benefits: While AI can improve customer service or logistics, quantifying these benefits into concrete financial returns can be challenging. Often, AI-generated insights are not acted upon, with only 25-50% of insights translating into action and 10-20% impacting revenue.

NEA's Tiffany Luck on the Path to Measurable AI Value

Tiffany Luck, a partner at New Enterprise Associates (NEA), a prominent venture capital firm with significant investments in AI and enterprise software companies, is closely observing this shift. NEA has invested in companies like ElevenLabs and Synthesia, leaders in generative AI technology. Luck emphasizes that the industry is moving past the initial hype toward a phase where tracking ROI on AI spend is critical.

Her insights suggest a focus on "vertical AI" and the "last mile" of automation. This means building AI solutions that solve specific, industry-hardship problems, rather than relying solely on general horizontal models. She draws parallels to the early days of e-commerce, where Fortune 500 companies initially struggled to integrate online shopping into their daily workflows. Just as e-commerce had to overcome technological, logistical, and mental friction, AI needs to move from a "shiny object" to a tool that solves mechanical problems.

Luck highlights the increasing role of "forward-deployed engineers" who embed AI within organizations to facilitate adoption and ensure practical, scalable implementation. She also notes that value is being generated across the entire AI stack, from data management to model deployment and user interfaces, not just at the model level. For regulated sectors, accuracy, auditability, cybersecurity, and data provenance are paramount, driving the need for certification standards for AI agents.

Strategies for Sustainable AI Implementation

To navigate these challenges, enterprises are adopting more strategic approaches to AI:

• Aligning AI with Business Goals: Identifying specific, high-impact use cases and mapping them to clear business objectives is crucial for proving value.
• Robust Cost Management: Implementing strategies to track, attribute, optimize, and govern AI spend is essential. This includes monitoring token usage, setting up custom usage alerts, using caching for repeated queries, and optimizing model selection for specific tasks.
• Cloud FinOps: Embracing operational disciplines like Cloud FinOps, which fosters collaboration between technology, finance, and business teams, helps control cloud costs associated with generative AI.
• Workflow Integration: Embedding AI directly into workflows rather than treating it as a separate tool layer is key to achieving measurable outcomes like reduced manual workload, faster execution, and better data quality.
• Focus on Execution, Not Just Adoption: The highest-performing AI organizations tie reclaimed capacity to specific metrics, push AI beyond the suggestion stage into execution, and track quality and defect rates alongside usage.
• Centralized Governance: Establishing a centralized control layer for AI systems, including intelligent model routing, governance policies, and end-to-end observability, can significantly reduce costs and improve efficiency.

The Future of Enterprise AI: From Hype to Practical Value

The current re-evaluation of AI spending marks a maturing phase for enterprise AI. The era of unbridled experimentation and "tokenmaxxing" is giving way to a more disciplined, results-oriented approach. Companies are realizing that true AI value comes not just from deploying the latest models, but from strategically integrating them into core business processes, managing costs effectively, and focusing on measurable outcomes. As Tiffany Luck and other industry experts suggest, the future of successful enterprise AI lies in thoughtful implementation that addresses real-world problems and delivers tangible ROI.

Frequently Asked Questions

What is "tokenmaxxing" and why did it lead to problems?

"Tokenmaxxing" was a trend where companies encouraged employees to maximize their usage of AI tools, particularly large language models (LLMs), often through internal leaderboards. While intended to boost AI adoption and productivity, it led to significant and unexpected cost overruns because LLMs are typically priced based on token consumption (input and output units of data), causing budgets to be exhausted rapidly.

Which major companies faced issues with AI budget overruns?

Several prominent companies have publicly acknowledged or been reported to face AI budget issues due to high usage. Uber famously blew through its entire 2026 AI budget in just four months. Meta discontinued an internal AI usage leaderboard after costs mounted, and Microsoft scaled back its internal Claude Code licenses due to high expenses, directing engineers to its own GitHub Copilot CLI.

What are the main reasons enterprises struggle to achieve AI ROI?

Enterprises struggle with AI ROI due to several factors, including a lack of clear strategic alignment for AI initiatives, underestimation of the total cost of ownership (which goes beyond just API fees to include infrastructure, operations, and talent), skills gaps in AI implementation and management, fragmented data systems that hinder integration, and difficulty in quantifying the intangible benefits of AI into measurable financial returns.

What strategies are companies adopting to manage AI costs and improve ROI?

To manage AI costs and improve ROI, companies are focusing on aligning AI use cases with specific business goals, implementing robust AI cost management practices (like monitoring token usage, caching, and model routing), adopting Cloud FinOps principles, deeply integrating AI into existing workflows for operational efficiency, and shifting focus from mere AI adoption to measurable execution and outcomes.