UBS Finds Enterprises Throttling AI Spending

According to a UBS research note dated June 23, 2026, covered by Business Insider and by Futu/Wall Street CN, UBS analysts Karl Keirstead, Timothy Arcuri, and Taylor McGinnis wrote that, based on conversations with roughly a dozen enterprise IT executives, about 60% of enterprises were "in some manner throttling AI spend" by adding guardrails such as token pooling, model downgrading, waste alerts, and per-user usage limits. Reported examples include one company that exhausted a large share of its annual token budget and cut its internal AI tools from five to two, one user who ran up $35,000 in a single month on AWS Bedrock, and DevOps teams that consistently exceeded weekly token quotas by 100-200%. The analysts called the pattern a modest "emerging headwind" for AI model makers, said "token spend optimization has become a key issue in most organizations," and added they "are not ringing the alarm bells," calling it "a healthy problem."

The primary technical response enterprises are adopting is not blanket rate-limiting but model routing: directing routine tasks to cheaper models and reserving expensive frontier models for complex reasoning, code generation, or long-context work. Per Futu/Wall Street CN's reporting on the same UBS note, the price gap motivating this is real - citing Anthropic's public pricing as an example, Haiku costs a fraction of Opus- or Fable/Mythos-tier pricing per output token, making task-based model selection highly cost-effective. Engineering teams are pairing routing with familiar cost-control patterns: token pooling, batching, caching, context-window pruning, and client-side usage accounting.

The note names open-source and Chinese models - including DeepSeek, Alibaba's Qwen, MiniMax, and Zhipu AI's GLM - as beneficiaries of this shift, with one large global bank reportedly deploying Qwen on-premises to offset spending on premium models like Claude. UBS frames this as normal cost governance during early-stage enterprise AI adoption rather than a sign that AI adoption itself is slowing; a companion UBS survey of roughly 130 companies found only 8% have deployed AI agents at scale in production, suggesting the token-consumption growth this optimization responds to is still in its early stages.

Engineers and ML platform teams should treat this as validation for investing now in routing infrastructure, per-task cost accounting, and usage observability rather than waiting for bills to force the issue - vendors and internal tooling that make routing and cost attribution easy are positioned to benefit regardless of which model wins any given task.

• •Whether AI model vendors respond with tiered pricing or bundled inference options to blunt the shift toward cheaper models.
• •Adoption of token-level observability and chargeback tooling inside enterprises.
• •Procurement shifts toward smaller or fine-tuned models for high-volume, low-complexity tasks.