Self-Hosted LLM vs API: The Break-Even Calculator Your CFO Actually Wants
Self-Hosted LLM vs API: The Break-Even Calculator Your CFO Actually Wants
The real math on when self-hosting an LLM beats paying per-token API fees — hardware costs, breakeven months, and the hybrid strategy that actually works.
BLUF: For most teams under roughly 100 million tokens/month against frontier models, managed APIs are still cheaper once you count the full cost stack. Above that volume — or if you're in a regulated industry with data-residency requirements — self-hosting on your own GPU wins, usually breaking even around month 5–7.
Every CTO has had this conversation: someone on engineering does napkin math — GPU rental is $2/hour, API tokens cost real money, therefore self-hosting saves money. Sometimes they're right. More often they're missing half the cost stack. Here's the actual model.
The Real Cost of Cloud APIs
Pricing varies wildly by provider and tier, which is exactly why teams get this wrong:
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Frontier tier (GPT-5-class, Opus-class) | $5–$45 | $25–$225 |
| Mid tier (Sonnet-class, GPT-4.1-class) | $1.50–$2 | $7.50–$8 |
| Budget tier (DeepSeek-class) | $0.14–$0.50 | $0.28–$2 |
The spread between frontier and budget-tier APIs can be 50–70x for the same token volume. That spread alone changes the entire self-hosting decision — the math that justifies self-hosting against a frontier model often doesn't justify it against a cheap budget API.
The Hardware Side of the Ledger
# Minimal self-hosted inference stack — this is genuinely how simple it's gotten
curl -fsSL https://ollama.com/install.sh | sh
ollama run llama3.3
ollama run deepseek-coder-v2 # coding-specialized model
| Hardware | Price | VRAM | Realistic Capacity |
|---|---|---|---|
| RTX 3090 (used) | ~$700 | 24GB | 7B–13B native, 70B quantized |
| RTX 4090 | ~$1,600 | 24GB | Same class, faster |
| DGX Spark | ~$3,000 | 128GB | 70B at full precision |
| Mac Studio M3 Ultra | ~$8,000 | 192GB | 70B–405B quantized |
Running the Actual Break-Even
For a team spending roughly $500/month on a mid-tier API today, a single RTX 4090 (hardware + electricity) typically breaks even against that spend somewhere around month 5–7. After that point, incremental cost per query approaches the cost of electricity — commonly $30–100/month regardless of volume.
Against a cheap budget-tier API like DeepSeek's, this math often never flips in self-hosting's favor. Cheap APIs are cheap for a reason, and undercutting them with your own hardware plus your own ops time is a losing trade unless data residency forces your hand.
The Hybrid Strategy (This Is What Actually Ships)
Nobody serious runs 100% self-hosted or 100% API in 2026. The pattern that works:
- Route routine, low-risk queries (summarization, classification, drafting) to a local model
- Send complex multi-step reasoning to a frontier API
- Keep an n8n workflow in front of both, with a simple router node deciding which path a request takes based on complexity or data sensitivity
This hybrid pattern commonly cuts total LLM spend by 70–80% while keeping frontier capability available when it's actually needed — and it's the exact architecture I build for automation clients who don't want to gamble their whole stack on one model provider.
The Case Where the Math Doesn't Matter
Healthcare, finance, and legal teams often don't get to run this calculation at all. HIPAA, GDPR data residency, and privilege concerns can make self-hosting (or a private-cloud endpoint) the only compliant option — the cost comparison becomes secondary to the compliance requirement.
Bottom Line
Do the math against your actual provider and actual monthly volume, not a generic "GPUs are cheap" assumption. If you're under the breakeven threshold, stay on the API and consider a hybrid router instead of an all-or-nothing migration.
Need help implementing this? I help teams architect and scale this exact infrastructure. Explore my consulting and freelance services.