UBS Survey: 60% of Enterprises Have Capped AI Spending as Chinese Open-Source Models Enter Procurement Options
Claire Weston
A June 23 UBS report finds roughly 60% of enterprises now cap AI token spending. Companies are routing tasks to cheaper models and pulling in Chinese open-source alternatives — reshuffling the AI value chain's winners.
Sixty percent of enterprises capping AI spend — what happened?
As AI agents and coding tools spread, token consumption shifted from occasional chatbot queries to round-the-clock task execution — and the bill landed on the CFO's desk.
Extreme cases from the survey: one company burned through its annual token budget so fast it cut internal AI tools from five to two; a single user on AWS Bedrock ran up $35,000 in one month; DevOps team members hit 100%–200% of their weekly token quota.
Databricks' CEO called it "a big speed bump, not a small one." This means → cost governance has moved from isolated corporate initiative to an industry-wide pattern.
How does "routing tasks to cheaper models" save money?
The key technique is model routing — assigning different tasks to different models so only complex reasoning, critical code, and long-context analysis hit the most expensive tier.
Anthropic's pricing illustrates the gap: Haiku 4.5 output costs $5 per million tokens, Opus runs $25, and the top-tier model reaches $50 — a 10× spread from low to high.
This means → the premium model shifts from "default" to "luxury." In plain terms = teams that once sent every task to the strongest model now ask: does this task really need the most expensive one?
Microsoft's new MAI small-language models hit the same note: a 35-billion-parameter mid-range "Thinking" model and a Code-1 positioned as a low-end frontier option — "good enough but cheaper."
How did Chinese open-source models get onto the enterprise procurement list?
Downtiering isn't limited to a single vendor's lineup. Enterprises are evaluating Chinese open-source models at scale, including Alibaba's Qwen, DeepSeek, MiniMax, Zhipu GLM, and Kimi.
A large global bank, managing token costs, has begun deploying Qwen on-premise to balance its use of high-end models like Claude. This means → the cost structure flips from per-token billing to local hardware provisioning, while sidestepping compliance risks tied to externally hosted Chinese models.
Cloud platforms have already added these models to their standard menus: AWS Bedrock lists MiniMax, Kimi, Qwen, DeepSeek, and GLM; Microsoft offers DeepSeek through Azure AI Foundry. This reflects a shift: Chinese open-source models are no longer a fringe option — they sit squarely on the enterprise cost curve.
Where does the impact hit hardest across the value chain?
The model layer takes the direct hit — customers switching from premium to smaller or open-source models pressure revenue growth at high-end providers first.
Cloud and hardware see limited impact: customers may swap models, but inference still runs on cloud infrastructure, so compute demand doesn't vanish. Cloud vendors' model-API revenue growth may slow, but core capacity spend holds.
Software companies sit in the most complex position: squeezed by customer budget cuts on one side, yet able to reposition as token-optimization platforms on the other. In plain terms = they stand at a fork between "getting their budget cut" and "becoming the tool that saves money."
Who is already running down the "save-the-customer-money" path?
Palantir commercialized AIP Evolve — a tool that picks the right model per task, tunes prompts, and optimizes data calls — roughly a month ago. In one disclosed case, Evolve recommended a model swap that cut token costs by 97%, with 90% adoption within three weeks of launch.
Large seat-based SaaS companies — Salesforce, ServiceNow, Workday — face an especially awkward position: customers are reallocating budgets while these vendors are still pushing the transition from pure seat pricing to "seat-plus-usage" billing.
This reflects the real shape of this optimization wave: AI demand hasn't disappeared, but the winner ranking is being rewritten — low-cost models and routing tools gain; premium-model revenue growth gets squeezed.
Where does AI agent deployment actually stand?
UBS's Evidence Lab surveyed roughly 130 enterprises: only 8% have deployed AI agents at scale in production, 37% use them in production at limited scale, and 29% remain in pilot.
In plain terms = most companies haven't yet woven AI agents into core operations, but they're already worrying about the token bill — optimization is running ahead of large-scale adoption.
Legal-AI firm Harvey shows that optimization and expansion can coexist: its token consumption grew from 1 trillion in January to 12–13 trillion by May. This means → UBS's core call stands: token optimization may temporarily drag on AI revenue growth, but the long-term trend remains strong.
Content is for reference only, not financial advice.