Amazon Bedrock
Guardrails that cost more than inference
Per-text-unit assessment costs can exceed the inference cost they protect.
What Labs investigates
Labs compares guardrail charges with inference costs to surface the imbalance.
Where the depth comes from.
Research into the infrastructure inefficiencies that billing data alone cannot explain.
Look beyond the obvious.
Four detection examples from the Q1 2026 update. Each starts with a specific behavior and investigates its cost.
Snowflake
Recoverable tables with permanent storage costs
Intermediate tables can be rebuilt, but still incur Fail-safe and extended Time Travel charges.
What Labs investigates
Labs traces table lineage to find data written exclusively from upstream sources.
Azure OpenAI
Rejected requests. Billed input.
Content-safety rejections can leave you paying for input processing with no useful output.
What Labs investigates
Labs examines rejection patterns and their associated token costs.
AWS EC2
CPU credits that change fleet economics
Burstable instances can cost more than a comparable non-burstable fleet.
What Labs investigates
Labs aggregates CPU-credit overages across auto-scaling groups.
Four domains. Deeper questions.
Configuration, usage, and cost analysis across cloud, data, and AI.
InfrastructureCompute, storage, and networking
AWS EC2, EBS, EFS, S3, Lambda, OpenSearch; Azure Storage, Site Recovery, App Configuration; GCP Dataflow, BigQuery, Compute
Platform servicesBehavior behind managed services
ElastiCache provisioned and serverless; SQS standard and FIFO patterns; DynamoDB and RDS Multi-AZ behavior
Data platformsWarehouses, tables, and pipelines
Snowflake warehouses, table lineage, and ingestion; Databricks and BigQuery coverage in active expansion
Production AIInference, models, and GPUs
Bedrock and SageMaker; Azure OpenAI deployments and PTU economics; Routing, caching, guardrails, model selection, and GPU utilization
From observation to detection.
PointFive Labs is the in-house research team behind DeepWaste detections.
Start with customer environments.
Research begins with real configurations, usage patterns, and workload behavior.
Follow the inefficiency below the bill.
Researchers apply an investigative approach drawn from cybersecurity, examining configuration, utilization, and lineage.
Turn findings into detections.
Detections include a cost model to quantify potential impact. Teams validate savings after implementation.
Open research.
Labs also publishes open, reproducible research on how AI infrastructure spend is incurred, with the benchmarks and data behind each finding.
July 2026
Token Reduction Is Not Cost Reduction
An empirical study of end-to-end efficiency in API-based coding agents.
August 2026
Prompt-Induced Waste in Coding Agents
How prompt wording changes reasoning, effort, and end-to-end cost for coding agents.
Benchmarks
Open benchmarks and data
The benchmark and data behind both papers, released so the results can be checked.
What the research uncovered.
The Q1 2026 report on new detections across cloud, data, and AI.
Read the full report- AI detections introduced
- 17
- Initial Snowflake detections
- 7
- Detections in one release
- 32
- Newly addressable annual savings
- ~$24M
Reported for the Q1 2026 customer cohort, including ~$8M in AI opportunities. Addressable savings are estimates, not realized savings or a forecast for every environment.