UBS analysts say executives are becoming more conscious of AI-related spending.
“Based on a dozen+ conversations with enterprise IT execs over the prior several weeks, ~60% of enterprises were now in some manner throttling AI spend by putting some degree of guardrails in place,” UBS analysts Karl Keirstead, Timothy Arcuri, and Taylor McGinnis wrote alongside their colleagues in a recent report.
Token spending has become a major concern for companies, especially larger enterprises, as CFOs and CTOs see their AI bills rise. Uber’s operations chief, Andrew Macdonald, said in May that it was getting harder to justify the rising costs given the pretty meager ROI.
The analysts said that their conversations, which began earlier in June, led them to conclude there was a modest “emerging headwind.” More recent conversations, they wrote, have reaffirmed their view of the situation, though the extent of the impact varies.
“Token spend optimization has become a key issue in most organizations, resulting in a big spending speed bump for some organizations, but a smaller speed bump for others that are either too early in their AI deployments or are far deeper but are unwilling to throttle users because they see the offsetting ROI or have an organizational priority in place to drive innovation and hence AI use,” they wrote.
AI model makers, such as OpenAI and Anthropic, are likely to be most exposed to cost-cutting in the near term, the analysts wrote. They specifically called out open-sourced and Chinese models like DeepSeek as the biggest potential beneficiaries, especially for enterprises looking for models for non-coding-related tasks.
Despite the cutbacks, the analysts emphasize that they “are not ringing the alarm bells,” calling the trend “a healthy problem.”
“Some measure of AI spend optimization is normal, no one is hitting the brakes on AI deployment and it is likely that we’re sitting in front of new models trained on next-gen chips that might drive token costs down further,” they wrote.
Leading AI companies have touted the token efficiency of their models and, in some cases, built models aimed at lowering costs. Google has its Gemini 3.5 Flash model. Anthropic rolled out Claude Sonnet 5 on Tuesday, which it said “runs autonomously at a level that just a few months ago required larger and more expensive models.”
One company UBS analysts spoke with said the industry is moving away from the phase of AI experimentation.
“The question isn’t whether to use tokens, it’s how to use them efficiently,” the analysts wrote in an excerpt of the conversation. “As a result, optimization becomes an ongoing engineering discipline rather than a reaction to a budget crisis.”
Another said their CTO went all in on AI early on, but now they are cutting back on the number of tools used.
“We have 5 AI tools internally and all of the LLM products. Like others, we ran into the issue where we have already used most of our token budget for the entire year,” the analysts wrote in an excerpt. “Now we’re only using 2 AI tools and being careful around usage.”
