Enterprises are wrestling with the surge in AI spending as token consumption climbs faster than any clear return on investment, a dilemma that has surfaced across multiple tech giants.
Token spending spirals at big firms
In late 2025, Uber equipped its engineers with Claude Code and introduced internal leaderboards that tracked token usage. By spring 2026 the entire AI coding budget for the year was exhausted, prompting the company to cap spending at $1,500 per employee each month.
Other firms have taken similar steps. Microsoft halted Claude Code licenses in its Experiences and Devices division after questioning the cost. At the language‑learning platform Duolingo, a plan to tie AI usage to performance reviews was abandoned after employee pushback.
Why usage outpaces value
According to a recent forecast, Gartner expects AI agent software spending to near $207 billion this year, up more than 139 % from the previous year. Yet pricing for consumption behaves unlike traditional software costs, varying dramatically from one developer’s session to the next.
Even when every engineer adopts AI tools, companies often cannot demonstrate that the spend translates into more features, fewer bugs, or faster problem resolution. This gap has sparked the term “tokenmaxxing,” describing unchecked consumption without proven benefit.
Infrastructure vs people debate
Noe Ramos, VP of AI operations at Agiloft, argues the issue lies in default infrastructure rather than wasteful habits. He notes that many enterprises let the prompting team pick the model, pushing frontier models onto tasks that cheaper alternatives could handle.
At Promova, head of engineering Dmytro Palaniichuk observed that premium models are pre‑selected on enterprise plans, leading to high‑effort runs for simple jobs like email checks. He is steering his team toward a split across three model families, but stresses that habit and awareness remain obstacles.
One could argue that the real challenge is building transparent cost controls that don’t stifle innovation. If organizations can align model choice with task complexity, the spending surge might settle into a more predictable pattern.
Solutions emerging
Several vendors are deploying routing layers that automatically match tasks to the most cost‑effective model. Databricks unveiled a Smart Routing feature within its Unity AI Gateway, evaluating request intent, length, and complexity before selecting a model.
The router also considers the surrounding harness, acknowledging that the same model can perform differently depending on how it’s used. When a budget ceiling is reached, administrators can either block further calls or fall back to a cheaper compliant model.
SUSE categorizes AI usage into daily work, autonomous agents, and strategic “curve‑jumping” projects, coaching managers to align consumption with impact rather than imposing hard caps.
At Everlaw, per‑person caps trigger a brief notification to the tools team, often resulting in an immediate increase to the limit. The company shared concrete numbers: a Java infrastructure task cost $3,500 in tokens but cut implementation time from 9.5 engineer‑months to 2.5.
Not every experiment succeeds. Everlaw’s attempt to port interface code from Dojo to React generated unusable output, prompting a revised workflow that first documents legacy behavior before generating new code.
Agiloft eventually removed its caps, finding that most users never hit them. Instead, the firm made cheaper models the default and introduced an escalation path to frontier models when necessary, emphasizing routing at the infrastructure level.
Costs keep climbing.
Executives anticipate that token expense will soon be treated like headcount in annual planning. Max Christoff of Everlaw predicts that departments will present both a token budget and a staffing plan, reflecting the growing importance of AI consumption in cost structures.
While specific tactics—such as fixed model mixes or temporary caps—may evolve, the discipline of matching tools to tasks is expected to endure. Companies that embed transparent routing, coach managers on impact‑focused usage, and maintain clear budgeting will likely manage the rising costs more effectively. Max Christoff says departments will present both a token budget and a staffing plan.
