# The Cost of ML Without a Platform

Most budgets look at GPU bills. The real cost is the team that assembles the pipeline around the GPU — and the months of calendar time that team consumes. This brief models the cost of a single end-to-end model for a team that builds the ML lifecycle manually, without a managed platform, and compares it to the platform-native workflow k3ld Model provides.

> **Scope note:** the comparison targets a team starting from scratch — raw data in S3, no pre-built training infrastructure, no managed ML platform. Teams already using SageMaker Studio, Vertex AI, or Databricks ML will see a narrower gap. (Updated August 2026.)

## The team you need with a manual, no-platform approach

| Role | Headcount | What they actually spend their time on |
|---|---|---|
| Data Engineer | 1–2 | S3 buckets, IAM policies, Glue crawlers, Parquet/JSONL conversions, connectivity debugging |
| ML Engineer | 1–2 | PyTorch scripts, LoRA config, hyperparameters, checkpoint management, GPU debugging (CUDA, OOM, NaN) |
| MLOps / DevOps | 1 | Docker, ECR, SageMaker endpoint config, autoscaling, CI/CD, CloudWatch dashboards |
| Data Scientist | 1 | Data profiling, prompt engineering, evaluation, hand-written drift scripts |
| Engineering Manager | 0.5 | Cost tracking, timeline coordination, vendor negotiation |
| **Total** | **4–5 people** | **≈$700K–$1,000K / year** fully loaded — building infrastructure, *not product features* |

## One model, end to end

- **6–10 weeks** — status quo calendar time for one trained + deployed + monitored model
- **8–14 person-weeks** — ≈ $35K–$65K in salary for a single model
- **2–3 days · 1 person** — with k3ld Model (≈ 4 hours of actual effort)

## Stage-by-stage: manual approach vs. k3ld Model

| Stage | Status quo time | Status quo infra spend | k3ld Model |
|---|---|---|---|
| Connect | 2–4 days | S3 + IAM + Glue, $5–15/mo | 2 minutes |
| Explore | 1–3 days | Athena queries, $2–5 | Instant in catalog |
| Build · from scratch | 2–4 weeks | GPU $800–$2,000 | Hours · a fraction of the GPU bill |
| Build · fine-tune | 3–5 days | GPU $15–50 | 30 min · a fraction of the cost |
| Build · post-train | 1–2 weeks | GPU $30–100 | 30 min · a fraction of the cost |
| Ship | 1–2 weeks | ECR + endpoint $1–3/hr | 2 minutes |
| Monitor | Ongoing · 0.75 FTE | CloudWatch $50–200/mo | Built-in, live |

## Where the money actually goes

**Manual, no-platform approach:** infra + GPU for one model $900–$2,500; salary $35K–$65K per model; monitoring headcount ~$150K/yr (0.75 FTE); retrain after drift = 2–4 weeks of re-work; idle GPU tax (endpoints up 24/7); key-person risk.

**k3ld Model (platform-native):** infra + GPU ~$5–$100 (pass-through AWS costs); ~4 hours of one person; monitoring built in (no separate headcount); retrain ~30 min + deploy; idle GPU tax eliminated (endpoints auto-sleep, wake on demand); the platform is the runbook.

> ⚠️ **Not counted here:** the opportunity cost of 6–10 weeks of calendar time. Every month of delay is a month of lost customer value.

## Honest limits

- The manual approach modeled here is the **worst case** (no ML platform at all). Teams on SageMaker Studio/Vertex/Databricks see a narrower gap. We chose the lower bound because it matches the buyer who needs k3ld most.
- k3ld numbers are **design targets, not guarantees**. "2 minutes to ship" assumes a completed training run with valid checkpoints; real deploys hit GPU capacity constraints, cold starts (5–15 min endpoint provisioning), and artifact misconfiguration. The direction is real; exact times vary by workload.
- Monitoring is a **shared responsibility** — k3ld provides drift dashboards and alerting, but a data scientist still interprets drift and curates retraining data.
- This is **not a comparison to other managed tools** — it's manual DIY vs k3ld Model.

## Methodology & sources (summary)

- Salary ranges (fully loaded): Data Engineer $160–200K, ML Engineer $190–250K, MLOps $170–220K, Data Scientist $170–220K (BLS OEWS May 2025 + 1.25–1.4× loaded multiplier).
- GPU pricing: AWS SageMaker on-demand, us-west-2, Aug 2026 — ml.g5.2xlarge $1.408/hr, ml.g5.4xlarge $2.816/hr, ml.g5.8xlarge $5.632/hr, ml.g5.12xlarge $8.449/hr, ml.p4d.24xlarge $32.773/hr. Spot typically 60–70% cheaper. Verified against the model.k3ld.com live estimator (sources from the AWS Pricing API).
- Timeline estimates from direct experience building ML pipelines on raw AWS without a managed platform.
- k3ld platform times are in-UI interaction time (exclude provisioning/cold-start wall clock).
- Not counted: platform subscription, data transfer, adoption learning curve, production serving (pass-through), and hiring costs.
- Updated quarterly. Last verified 2026-08-10.