22 providers across four categories, evaluated for portability, model integrity, and data security — under the assumption that company data may run outside WWG as long as it stays sealed in a secure, portable, vendor-movable container.
Note: the Word report and Excel matrix are companion files — download both for the complete evaluation. Each file is self-contained and opens on its own.
Recommended build
An open-weight 12–14B instruct model (Qwen3 14B or Phi-4 14B) at 4-bit, served by vLLM, with RAG over notes and Text-to-SQL over tables, embeddings stored inside SQL Server — all in one encrypted, portable container with our own keys. Start with RAG only; fine-tune later only if needed. Validate on a scale-to-zero rented GPU, step up to a confidential-computing VM if a provable guarantee is needed, and buy a server only if usage proves heavy and permanent.
Cost at a glance
~10–25 managers, a few hundred–2,000 queries/day. Embeddings & vector storage are ~free (self-hosted in SQL Server).
Rent · scale-to-zero
$30–55/mo
Cheapest defensible option for light, bursty use. Single-tenant.
Buy · amortized
$140–200/mo
~$5–7k server over 3 yrs. Breakeven vs 24/7 rental ≈ 19 months.
Rent · dedicated 24/7
$320/mo
Always-on single-tenant secure GPU, no capex.
Confidential cloud
~3–4×
Premium over a normal GPU — only when an operator-blind guarantee is required.
1 · Rent-a-GPU (bring your own container)
Provider
H100 $/hr
Cheapest GPU
Data protection
Best for
RunPod
$2.89–3.29
RTX 4090 $0.69
Secure Cloud, SOC 2, region-select
Best cost/flexibility for a small shop
Vast.ai
from ~$2.00
RTX 4090 $0.35
Use vetted “Secure Cloud” hosts only
Cheapest; experiments
Lambda
$3.99 (8×)
V100 $0.79
Private Cloud = single-tenant
Predictable, established
CoreWeave
~$6.16/GPU
node-priced
VPC; $0 egress (best exit)
Committed scale
DO Gradient
$3.39
RTX 4000 $0.76
DO regions, mature compliance
Teams on DigitalOcean
2 · Managed fine-tuning / serving
Provider
FT entry
Download weights?
VPC / self-host
Status
Together AI
$0.48/1M
Yes CLI export
Yes (Enterprise VPC)
Independent
Fireworks AI
$0.50/1M
Yes best-documented
Isolation + BYO bucket
Independent · strong compliance
Modal
GPU-hr only
Yes full ownership
No VPC/on-prem
Independent · container-first
Baseten
GPU-min only
Yes Download Truss
Yes (your VPC)
Independent
OpenPipe
$0.48/1M
Yes + open ART
Yes (on-prem/VPC)
Acquired · CoreWeave ’25
Predibase
per-1M
Yes LoRAX adapters
Yes (your cloud)
Acquired · Rubrik ’25
Replicate
GPU-sec
Partial images
No
Acquired · Cloudflare ’25
3 · Hyperscalers
Cloud
BYO Docker training
Confidential GPU
Export open weights
Watch-out
AWS
Yes (SageMaker)
CPU enclaves GA; GPU roadmap
Yes → your S3
Bedrock FT weights captive
Google Cloud
Yes (Vertex)
Confidential H100 GA
Yes → GCS
Global endpoint ignores residency
Azure
Yes (Azure ML)
Confidential H100 GA (NCC v5)
Yes (inferred)
Azure OpenAI FT weights captive
Oracle OCI
Yes (Data Science)
Not GA for GPU
Yes → Object Storage
$0 egress is a plus
4 · Self-host / open-source stack — the layer that keeps data ours
Tool
Role
License
Cost
Lock-in
vLLM
Inference server
Apache 2.0
Free
Lowest — OpenAI-API compatible
Ollama
Simple model server
MIT
Free
Very low
Unsloth
Fine-tuning library
Apache 2.0
Free
Very low
HF AutoTrain
No-code fine-tuning
Apache 2.0
Free self-host
Very low
NVIDIA NIM
Optimized inference
Proprietary
$4,500/GPU/yr
Highest — NVIDIA GPU + license
Open-weight models — license & availability
Once downloaded under Apache 2.0 / MIT, your copy is perpetual and irrevocable. Prefer the safest tier for a long-lived product.