Key points
- Cloud token allocations are being exhausted in weeks instead of months by agentic AI tasks.
- Enterprises are shifting focus toward on-premises infrastructure and open-weight models to manage costs.
- Intel and Dell note a growing demand for central processing units, not just graphics processing units, for AI.
Enterprise artificial intelligence projects are hitting unexpected financial limits as cloud computing costs rise sharply. According to a report by SiliconANGLE, the scalability of these systems is constrained by hard budget walls for many customers. As organizations deploy more agentic workloads, the associated cloud token bills are consuming financial allocations much faster than originally planned. This rapid expenditure is forcing companies to rethink where they run their inference tasks.
By the numbers
- Weeks: The time frame in which some Bay Area customers are exhausting token budgets meant for months.
- Months: The original duration for which those same token budgets were allocated by enterprise customers.
- Three: The number of architectural layers Intel cites for agentic AI, including orchestration, inference, and edge computing.
What the figures mean
Helen O’Sullivan, an AI business development manager at Dell Technologies, stated that many customers are struggling with this immediate cost pressure. She noted that allocations intended to last for months are being used up within a matter of weeks. To address this "token stress," companies are moving on-premises infrastructure and open-weight models back into their deployment strategies. This shift allows them to reduce reliance on expensive frontier models hosted in the cloud.
Tim Wood, a sales director at Intel Corp., discussed these trends with John Furrier of theCUBE Research during the Dell AI Leadership Symposium. Wood explained that enterprises can evaluate deployment options before committing to production environments. Testing open-weight models in sandbox settings helps companies assess what their specific workloads require. This approach allows for better cost management and technical alignment before scaling up operations.
Agentic AI is also changing the balance of processor demand within data centers. O’Sullivan noted that while many buyers remain focused on graphics processing units, the nature of agentic workloads is shifting utilization toward central processing units. She said her team anticipated this rebalance months ago, though many customers have not yet fully adjusted their hardware procurement strategies. The heavier utilization of CPUs for agents is becoming more apparent as workloads evolve beyond simple inference tasks.
What to watch
The conversation around enterprise AI deployment is expanding beyond just cloud-based frontier models. On-premises solutions and open-source alternatives are gaining traction as cost-effective ways to handle increased computational loads. Buyers who remain fixated solely on graphics processing units may miss opportunities to leverage existing CPU infrastructure. Intel’s architecture for agentic AI spans orchestration, inference, and edge computing, suggesting a broader hardware mix is needed.
O’Sullivan emphasized that organizations can no longer pursue AI initiatives without clear corporate goals. Having a specific objective in mind is essential for aligning technology investments with business outcomes. The rapid exhaustion of token budgets highlights the need for strategic planning in AI deployment. Companies must evaluate whether cloud-based inference or local processing better serves their long-term operational and financial needs.
What to do and how to stay safe: Dell
- Audit current cloud token consumption rates to identify workloads that are exceeding budget projections faster than anticipated.
- Evaluate existing central processing unit infrastructure to determine if it can handle agentic AI orchestration and inference tasks locally.
- Test open-weight models in isolated sandbox environments before committing to large-scale production deployments to assess performance and cost.
- Align AI project goals with specific corporate objectives to ensure that computational resources are being used for measurable business outcomes.
Step-by-step guide: Detect Shadow AI: Network Logs, Process Trees and Hidden Signals
General security guidance from the Sandbox Gazette newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Why are enterprise AI budgets being exhausted so quickly?
According to Dell and Intel, agentic workloads are multiplying and consuming cloud token allocations faster than planned, sometimes burning through months of budget in weeks.
What hardware shift are Dell and Intel observing in AI deployments?
They note a rebalance toward heavier utilization of central processing units for agentic AI tasks, even though many buyers remain focused on graphics processing units.
How can companies manage the cost of AI inference?
Companies are moving on-premises infrastructure and open-weight models back into their strategies to eliminate some of the token stress associated with cloud-based frontier models.
