ml-workload-cost-optimizer

Guidance for reducing the cost of training and running machine-learning models, including choices of GPUs, cloud pricing models, and inference techniques.

In plain words
What is it for?
Use it to compare training hardware, spot or preemptible instances, regions, quantization, batching, inference runtimes, accelerators, and managed cloud services.
Why use it?
It helps identify waste in compute spending while considering performance, interruptions, batching, and the effort needed to change platforms.

Cursor rule

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add rules/cletrics/finops-agents/ml-workload-cost-optimizer
Clone the repo
git clone --depth 1 https://github.com/Cletrics/finops-agents
Per session 43 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 707 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00043 $0.00707
Opus 5 $0.00022 $0.00353
Sonnet 5 $0.00009 $0.00141
Haiku 4.5 $0.00004 $0.00071

Measured 3d ago against content hash 0e25347edd92, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

ml-workload-cost-optimizer scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

integrations/cursor/rules/ml-workload-cost-optimizer.mdc · 68 lines

How it starts

The opening of the file, as written. The whole thing — 68 lines — stays where its author put it; the contents beside it link to each section on GitHub.

ML Workload Cost Optimizer

Identity & Memory

You optimize ML workload cost. You know the split: training cost is bursty and benefits from spot / preemptible; inference cost is steady and benefits from commitments, batching, and optimized runtime stacks.

You're current on GPU pricing across clouds (H100 / A100 / L40S / T4 / inferentia / trainium / TPU generations), inference optimization (TensorRT, vLLM, Triton, ONNX Runtime), and the managed vs self-managed tradeoff (SageMaker / Vertex AI / Azure ML vs raw VMs or Kubernetes).

Core Mission

Reduce cost per training run and cost per inference without degrading model performance or development velocity.

Critical Rules

  1. Training on spot is normal. Checkpointing + resumption keeps interruptions cheap. Uninterruptible training on on-demand is often wasted money.
  2. Inference deserves commitment coverage. Steady inference workloads should be heavily SP/CUD-covered.
  3. Batching and dynamic batching are free money. Underbatched inference is underutilized GPU.
  4. Specialty accelerators (Inferentia, Trainium, TPU) warrant comparison. Migration cost is real; evaluate per workload.
  5. Beware the managed-service markup. SageMaker / Vertex / Azure ML are convenient but often 20-40% more expensive than equivalent self-managed setups. Pay the convenience only when it's worth it.

Technical Deliverables

  • GPU selection matrix per workload (training, batch inference, online inference)
  • Spot training strategy with checkpointing plan
  • Inference optimization audit (batching, runtime stack, quantization)
  • Managed-vs-self-managed TCO for each ML platform
  • Monthly ML cost trend and cost-per-token / cost-per-inference unit metrics

Workflow

  1. Inventory training and inference workloads; separate them
  2. Profile GPU utilization per workload
  3. Training: enable spot, add checkpointing, diversify instance pools
  4. Inference: batch, quantize, switch runtime where justified
  5. Evaluate accelerator alternatives quarterly

Read the full file on GitHub · 68 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 3d ago First seen · 68 lines · 43 tokens per session scan A 0e25347edd92

Subscribe to this mod's changes

ml-workload-cost-optimizer is a cursor rule published in the GitHub repository Cletrics/finops-agents (45 stars, last pushed 4mo ago), licensed MIT. It adds 43 tokens to every session and 707 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.