AI Token Costs Are Exploding. Companies Are Scrambling to Manage Them.

The Numbers Are Getting Hard to Ignore

Royal Bank of Canada's CEO disclosed last month that the bank's AI token usage surged 500 percent over six months. At Cisco, a third of employees use an internal AI chatbot daily. CEO Chuck Robbins said on an earnings call that token usage is getting "pretty, pretty crazy." At analytics software firm Amplitude, some top engineers are "spending thousands of dollars a month or more on tokens," according to CEO Spenser Skates.

Those aren't outliers. According to a WIRED review of earnings call transcripts from data provider AlphaStreet, roughly 300 companies addressed AI token costs in April or May of this year. A year earlier, only 93 companies mentioned the word "token" during the same window. Box CEO Aaron Levie called token budgeting "one of the most important" and "heated" topics his company is now navigating.

What a Token Actually Costs You

Tokens are the unit of measure for AI workloads: every word you send a model and every word it sends back counts against your bill. The more complex the task, the more tokens consumed. Multiply that across thousands of employees using AI tools daily for code, emails, customer analysis, and internal documentation, and the invoice compounds fast.

Meta, Uber, and Salesforce have all publicly flagged concerns about rising generative AI costs and have begun introducing usage caps in some cases, according to Wired. Executives at several companies told Wired they are building or buying systems to monitor usage in real time and automatically route prompts to the cheapest model that can handle the job.

One Company That Says It's Working

Not every finance team is panicking. Software company 8x8 says it has saved roughly $5 million annually over the past 18 months by canceling subscriptions to dozens of software and educational tools that Claude, Anthropic's AI model, now replaces. The company's annualized spend on Claude is currently "well below" that $5 million figure, according to Joel Neeb, 8x8's chief transformation and business operations officer, who spoke with Wired.

Neeb expects the gap to close as 8x8 pushes broader adoption and tackles more complex workflows. He didn't share exact total AI spending numbers. But for now, the savings are real and the CFO is satisfied.

Consolidating dozens of SaaS subscriptions under a single AI platform is a legitimate cost-reduction play, and companies that move thoughtfully stand to gain from it. The challenge isn't whether AI can reduce costs. It's whether most organizations have the discipline to manage consumption before the bills outpace the savings.

The Legitimate Counterargument

The case for aggressive AI spending is not irrational. Productivity gains from AI-assisted coding, faster customer response, and automated analysis can compound over time in ways that don't show up cleanly in a quarterly earnings call. Companies that under-invest now risk falling behind competitors who absorbed short-term costs and built durable capabilities. Token costs are also falling as model providers compete on price. What costs $1,000 a month in tokens today may cost $200 in 18 months. Viewed through that lens, current spending is closer to infrastructure investment than runaway waste.

That argument holds if the tools are actually replacing meaningful labor or eliminating real software costs, as 8x8 claims to have done. It breaks down when AI is layered on top of existing headcount and existing software subscriptions as an add-on, rather than a substitute.

Energy Sits Behind All of It

The corporate tokenomics debate is the business-side symptom of a deeper physical constraint. OilPrice.com has reported on the mounting pressure AI infrastructure is placing on the energy grid, with data center power demand driving a renewed push for energy efficiency across the sector. Every token processed anywhere requires electricity. As token volumes triple and quadruple at individual companies, the aggregate demand signal feeding back to utilities and grid operators grows proportionally.

The efficiency conversation companies are having about token costs internally is, at scale, the same conversation grid planners are having about load forecasts. Neither side has solved it yet.

What Comes Next

The unresolved question for most organizations is whether they have the internal monitoring infrastructure to know, in practice, what each department is actually spending, on which models, for which tasks, and whether those tasks are generating measurable output.

According to Wired, many companies are still figuring out the right balance between hiring more people and expanding token budgets. That's not a technology problem. It's a management problem, and it won't be solved by the model providers.