Weekly analysis
Keep alternatives open as AI costs rise
Rising spending and unpredictable bills strengthen the case for testing alternative AI deployments. The evidence supports comparison, rather than a promise that decentralization always costs less.
AI spending is rising. Results remain uneven.
Gartner research reports that technology leaders are spending 18 times more on AI than they did two years ago, while only one in five AI projects are succeeding. It also reports that 16% of enterprises are achieving productivity gains [s3]. Those findings raise a question for leaders: what are we paying for, and which workloads are delivering value?
Billing predictability is part of that question. Splunk’s State of Observability 2026 report, based on a survey of 2,050 ITOps and engineering professionals, says 60% reported operational strain from unexpected AI-related billing spikes [s2]. These are survey findings. They do not establish that dependence on closed models caused the surprises, or that moving to an open model would prevent them.
Compare alternatives on the same workload
Reporting on open-weight models deployed on Nebius infrastructure describes matching the quality of leading closed models at 60% of the cost per answer [s1]. That is a reported result for a particular deployment, rather than a general price guarantee. It is a reason to test an alternative using your own quality requirements and operating costs.
A separate example in the same reporting describes Stripe achieving 92% cost savings in offline testing with NVIDIA GPU-accelerated Ray Data, alongside faster data pipelines [s1]. That example concerns data processing. It should not be treated as a comparison of open and closed language models.
Our view
The case for decentralized AI starts with practical choice. An organization that can evaluate more than one model and deployment environment has a way to question its current supplier’s economics. That ability has value even when the current service remains the best option.
Access to model weights provides only part of that choice. Hardware, cloud services, staffing and data engineering can still create dependencies. Moving a workload to another provider changes those dependencies; it does not make them disappear.
This week’s evidence supports a measured experiment. Select a recurring workload and compare the current service with one viable alternative. Measure the cost of a useful result, including retries, human review and operations. Check the license and document what a future migration would require.
The reports do not settle which deployment will win that comparison. They do give leaders a reason to make the comparison possible before committing more spending.