What Happened
Databricks says its serverless compute usage grew more than threefold over the past year and has introduced smarter cost controls. The available information does not specify the controls or show whether customer spending rose at the same rate [1]. NVIDIA has scheduled a November 17 conference call to discuss results for its fiscal third quarter, which ended October 25. The results have not yet been reported [2].
These updates do not establish a new performance or pricing comparison among NVIDIA, AMD and Intel hardware, or among AWS, Google Cloud, Azure, Databricks, Snowflake and Cloudflare services. Buyers should evaluate those options against their own workloads rather than infer a market winner.
Why It Matters to Businesses
As serverless usage expands, teams need to know which applications, users and jobs consume compute—and whether that consumption produces useful results. For GPU workloads, the lowest advertised instance price may not deliver the lowest cost per completed training run or inference request. Utilization, data movement, latency and operational effort all affect the decision.
Kimbodo Engineering Perspective
Choose a workload before choosing a platform. Managed AI services can reduce operational work but may constrain model selection or cost visibility. Direct GPU access offers more control but requires capacity planning, deployment automation and security operations. Keeping model serving interfaces and data contracts portable is usually more valuable than trying to make every infrastructure component interchangeable.
How We Would Implement It
- Inventory training, batch inference and real-time inference separately; record throughput, latency, data-location and availability requirements.
- Benchmark representative workloads on shortlisted NVIDIA, AMD or Intel configurations and relevant AWS, Google Cloud or Azure services. Measure cost per completed job or successful request, not just hourly rates.
- Evaluate Databricks and Snowflake where workloads depend on their data environments, and Cloudflare where edge placement is a requirement. Verify each candidate’s actual model, hardware and deployment support before committing.
- Deploy through versioned pipelines with workload-level tags, budgets, usage alerts, access controls and rollback procedures. Review idle capacity and serverless consumption regularly.
Risks, Costs and Security
Key risks include unpredictable consumption, GPU capacity constraints, data-egress charges, vendor-specific integrations and sensitive data reaching an unintended service. Set spending limits and ownership for each workload; restrict service identities and network paths; encrypt data; and log model, data and configuration changes. Treat NVIDIA’s forthcoming results as a potential input to procurement planning, not as evidence of future GPU availability or pricing [2].
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our AI Infrastructure & MLOps practice, or Estimate My Infrastructure.
