Reported by 1 source

The short version

  • The rapid adoption of AI agents is driving a projected twenty-four-fold increase in monthly token consumption by 2030, outpacing the decline in individual unit costs.
  • Businesses face significant financial uncertainty because subtle variations in prompts and model behaviors make it difficult to predict total resource usage for specific tasks.
  • Executives warn that current flat-fee personal accounts may become unsustainable as vendors seek profitability, prompting a need for stricter internal controls and more precise prompt engineering.

The economic landscape of artificial intelligence is undergoing a complex shift as major technology firms attempt to monetize services that have historically been offered at little or no cost to consumers. While platforms like ChatGPT, Claude, and Gemini provide accessible entry points for users seeking assistance with routine tasks, the underlying infrastructure requires massive capital investment from developers such as Microsoft, Google, and Anthropic. These companies are now under pressure to recoup hundreds of billions of dollars spent on developing large language models, leading to a push toward paid tiers that offer advanced capabilities for coding, billing, and other specialized functions.

However, establishing a stable pricing structure for these services has proven surprisingly difficult. The fundamental unit of cost in this ecosystem is the token, a mathematical chunk of data processed by the model during both input and output phases. Unlike traditional software licensing, where costs are fixed per user or per seat, AI usage is variable and often unpredictable. Simon Gooch, an executive at identity management firm Saviynt, noted that attempting to lock customers into long-term cost models spanning one to three years is currently illogical because the industry lacks sufficient data to forecast future consumption patterns accurately.

News Journal

The unpredictability stems from the non-deterministic nature of large language models. Subtle changes in a user’s prompt can lead to vastly different responses, and even identical prompts may yield varied results across different runs or models. This variability is exacerbated in agentic systems, where multiple AI agents collaborate to make decisions and execute actions. Such setups significantly increase token usage and complicate cost estimation, as the number of interactions required to complete a task can fluctuate widely depending on how the agents navigate their assigned workflows.

Despite a sharp decline in the price of individual tokens over recent years, total consumption is skyrocketing. Analysis by Goldman Sachs indicates that global token usage could increase twenty-four times between 2026 and 2030, reaching 120 quadrillion tokens per month as organizations transition from experimental use to full-scale deployment of AI agents. This surge in volume threatens to overwhelm budgetary controls, particularly for companies that lack visibility into their real-time consumption until they receive monthly invoices or hit usage caps.

High-profile examples illustrate the risks of unmanaged token expenditure. Reports suggest that Microsoft has restricted its engineers’ access to certain third-party coding tools to curb costs, while Uber reportedly exhausted its annual AI coding budget in just a few months earlier this year. Will Venters, an associate professor at the London School of Economics, explained that companies often struggle to manage these expenses because the value derived from AI is not deterministic. The output quality and resource intensity can vary significantly, making it hard to assign a consistent monetary value to each interaction.

In response to these challenges, some smaller organizations are currently bypassing enterprise pricing structures by using personal accounts with flat fees. Oliver King-Smith, founder of engineering software firm smartR AI, observed that this practice allows smaller entities to operate under the radar, but he warned that it is unlikely to be sustainable. As major AI platforms face increasing pressure from shareholders to demonstrate profitability, they are expected to tighten restrictions on such usage patterns, potentially ending the era of low-cost, unlimited access for business applications.

To mitigate financial risks, experts recommend that companies adopt more rigorous internal controls. Rob Steele, chief financial officer at UK accounting software firm iplicit, emphasized the importance of precise prompt engineering, comparing it to giving detailed instructions to a family member before sending them grocery shopping. Without clear parameters, AI systems may engage in unnecessary processing or generate irrelevant outputs, driving up costs without adding proportional value. Additionally, firms must consider the broader implications of integrating AI into products that serve thousands of users, as scaling these systems can lead to exponential increases in token requirements.

The complexity of managing AI costs extends beyond core development tasks. Managers may find themselves needing tokens for testing, security audits, and implementing guardrails, all of which contribute to the overall expense. Venters pointed out that while expanding a human workforce involves careful deliberation over headcount and hiring processes, deploying additional AI agents can be done with a single click. This ease of expansion poses a significant risk if not accompanied by robust monitoring and cost-management strategies.

Looking ahead, the industry must reconcile the volatility of token economics with the need for predictable business operations. While the immediate future remains uncertain, the trend toward higher consumption and stricter vendor controls suggests that companies will need to invest in better tracking tools and more disciplined usage policies. The ultimate value of AI may still outweigh its costs, but realizing that benefit will require a fundamental shift in how organizations approach budgeting and resource allocation for non-deterministic technologies.

Sources behind this briefing

Go to the original reporting