NEA's Tiffany Luck Says Enterprises Are Still Figuring Out Their AI ROI

Background and Context

Earlier this year, the Silicon Valley technology sector was dominated by a trend that was both humorous and highly indicative of the current market sentiment: "Tokenmaxxing." This term describes the practice of maximizing AI token consumption regardless of the associated costs, driven by a desire to extract maximum value from generative models. During this period, chief executive officers across major tech firms actively encouraged their employees to integrate artificial intelligence tools into their workflows as aggressively as possible. The prevailing logic was that sheer volume of usage would eventually yield technological dividends, leading to a surge in adoption rates that outpaced financial planning. This era was characterized by a sense of urgency and a fear of missing out on the transformative potential of large language models, with leadership teams prioritizing breadth of deployment over depth of optimization.

However, the initial enthusiasm has rapidly cooled as the financial realities of AI integration have come into sharp focus. Reports from media outlets such as TechCrunch highlight a significant turning point in corporate AI spending. Uber, the global ride-hailing giant, reportedly exhausted its entire annual artificial intelligence budget in just a few months. This case serves as a stark illustration of how quickly costs can spiral when usage is not strictly monitored or capped. The incident has become a symbol of the broader industry challenge, where the assumption that AI costs would remain manageable or predictable has been proven false in many high-volume use cases. The gap between projected and actual spend has forced a reevaluation of how these technologies are budgeted for at the enterprise level.

In response to these escalating costs, companies are beginning to implement contraction strategies. Some organizations have started reducing licenses for advanced models like Claude across specific business units, signaling a move away from universal access toward more selective usage. Procurement teams are no longer simply signing multi-year contracts for unlimited compute; instead, they are actively re-evaluating every dollar spent on AI infrastructure. This shift marks a departure from the earlier phase of unrestricted experimentation. The industry is moving from a period of aggressive expansion to one of fiscal discipline, where the focus is shifting from how much AI can be used to how efficiently it can be used. This correction is essential for long-term sustainability, as the previous model of spending was clearly unsustainable for most organizations.

Deep Analysis

The phenomenon of "Tokenmaxxing" can be understood as a form of defensive anxiety and exploratory trial-and-error in the face of technological uncertainty. In the early stages of the generative AI boom, many enterprises suffered from a fear of missing out, worrying that failing to adopt AI comprehensively would put them at a competitive disadvantage. Consequently, management teams adopted a strategy of encouraging high-frequency usage across the organization to rapidly accumulate data, refine workflows, and test the boundaries of the technology. This approach was driven by the belief that widespread adoption was a prerequisite for innovation. However, this strategy overlooked the complex cost structure of large language models, which exhibits significant scale effects and non-linear characteristics. As the volume of concurrent requests increased, inference costs did not rise linearly but often exponentially, particularly when moving beyond simple query-response tasks to more complex operations.

The cost explosion is particularly acute when enterprises transition from basic applications to more demanding tasks such as code generation, long-document analysis, or multi-step reasoning. These complex tasks require significantly more tokens per interaction, leading to consumption levels that far exceed initial estimates. Furthermore, many organizations lacked granular cost monitoring mechanisms during the deployment phase. Without the ability to distinguish between high-value interactions and inefficient or redundant usage, companies found themselves unable to control the bleeding. AI, initially intended as a tool for efficiency, transformed into a budget-draining black hole. The lack of visibility into which specific prompts or workflows were generating value versus those that were merely consuming resources left finance and operations teams blindsided by the final bills.

Tiffany Luck, a partner at the venture capital firm NEA, has highlighted that the core of the current "quantification dilemma" is the absence of robust evaluation models. Most enterprises have not yet established a framework that tightly links AI usage data with key business performance indicators such as code submission rates, customer service resolution times, or content production efficiency. Without a clear definition of return on investment, it is impossible for companies to determine which AI interactions are generating genuine business value and which are simply resource waste. This gap in measurement capability is the primary barrier to rationalizing AI spend. Until organizations can correlate token consumption with tangible output improvements, they will continue to struggle with budget overruns and ineffective allocation of capital.

Industry Impact

This shift from unbridled adoption to cost-conscious evaluation is having profound implications for the competitive landscape and the stakeholders involved. For cloud service providers and large model vendors, the narrative of pure "usage growth" is no longer sufficient to justify high valuations. The market is increasingly focusing on the actual commercial conversion capabilities of AI services. The budget crises faced by companies like Uber send a clear signal to the market that artificial intelligence is not a public good with negligible marginal costs, but rather a commercial infrastructure that requires strict cost control. This pressure is forcing AI suppliers to evolve their business models. They can no longer rely solely on selling compute power; instead, they must pivot toward selling solutions that offer better price-performance ratios, optimized inference efficiency, or even outcome-based pricing structures.

For small and medium-sized enterprises, this trend presents both significant challenges and new opportunities. On the one hand, smaller companies may face higher per-unit cost pressures due to their lack of economies of scale compared to tech giants. They may find it difficult to negotiate favorable rates or absorb unexpected cost spikes. On the other hand, this environment creates an opportunity for those organizations that can establish sophisticated AI governance frameworks early. Companies that can precisely identify high-return-on-investment application scenarios and implement strict cost controls will build significant efficiency barriers against their competitors. The ability to use AI profitably, rather than just extensively, is becoming a key differentiator in the market.

Additionally, this shift is catalyzing the growth of a new market segment focused on AI cost optimization, often referred to as FinOps for AI. A variety of new tools are emerging to help enterprises manage token consumption, optimize prompt efficiency, and monitor for anomalous usage behaviors. These tools provide the necessary visibility and control mechanisms that were missing during the initial boom. By offering granular insights into how AI resources are being utilized, these platforms enable finance and IT teams to enforce budgets and identify waste. The rise of this ecosystem indicates a maturation of the industry, where the focus is shifting from mere deployment to sustainable management. It reflects a broader recognition that long-term success in AI requires not just technological capability, but also financial discipline and operational rigor.

Outlook

Looking ahead, corporate AI strategies are expected to enter a phase of "meticulous cultivation." We anticipate that an increasing number of companies will establish dedicated AI governance committees responsible for setting internal usage policies, cost-sharing mechanisms, and performance evaluation standards. This institutionalization of AI management will be crucial for maintaining control over spending and ensuring that investments align with broader business objectives. The era of ad-hoc adoption is giving way to structured governance, where every AI initiative is subject to rigorous scrutiny regarding its cost-benefit ratio. This shift will help organizations move from a state of reactive budget management to proactive strategic planning.

From a technological perspective, the trend is likely to favor smaller, specialized models over general-purpose large language models in specific vertical domains. These niche models can offer lower costs and higher accuracy for targeted tasks, making them more economically viable for routine operations. Enterprises will increasingly adopt a hybrid approach, using large models for complex reasoning and smaller, cheaper models for high-volume, low-complexity tasks. This stratification of model usage will allow companies to optimize their token spend by matching the capability of the model to the complexity of the task. Furthermore, AI procurement will shift from a "blanket rollout" strategy to a "precision strike" approach. Companies will be more willing to pay for AI solutions that can directly demonstrate efficiency gains or revenue growth in specific use cases, rather than paying for broad, unmeasured access.

A key signal to watch is the emergence of organizations that successfully integrate AI costs into their regular operational metrics. Those that can navigate the transition from the "exploration phase" to a "scaled profitability phase" will likely dominate the next round of market洗牌. For investors and industry observers, the focus must shift from adoption rates to the actual economic benefits generated by AI investments. The race is no longer about who can use the most AI, but who can use it most profitably. This will be a prolonged battle of technological implementation capability and business wisdom, where the winners will be those who can balance innovation with fiscal responsibility. The industry is maturing, and the time for reckless spending is over; the time for measured, value-driven investment has begun.