The short version: on money alone, running locally does not beat the API. As of September 5, 2026 a used RTX 3090 build costs about $2,700 over three years. That buys over 13 billion tokens from the DeepSeek V4 Flash API, more than one 3090 can generate in three years flat out. Against frontier models such as Claude Sonnet 5 or GPT-5.6 terra, priced around $10 per million output tokens, the break-even sits at 300K to 500K tokens a day, which heavy users do reach. The reasons to buy a card have always been privacy, offline use, fine-tuning and a gaming habit. Here are the three ledgers, line by line.
What are the cost items of a local setup?
Four: the GPU, the rest of the build, electricity and depreciation, with the GPU over 60% of the total.
- GPU: used prices come from the live data on this site. This guide uses three tiers as examples: the RTX 3090 (entry 24GB), the RTX 4090 (mainstream) and the RTX PRO 6000 Blackwell (96GB high end), plus a Mac Studio M4 Max 128GB as a whole-machine comparison.
- Rest of the build: CPU, board, 32GB of RAM, an 850W PSU and a case run about $700 in 2026. Expert offload for MoE models wants 128GB of RAM, another $300 or so. If you already own a desktop, this line is zero.
- Electricity: rated card power × real load factor × hours × price per kWh. Inference typically draws 50% to 70% of the rating; we use 60%, 8 hours a day and residential rates.
- Depreciation: a card is worth 30% to 40% of its price after three years, and old cards like the 3090 fall slowly. The table below does not subtract resale value, so real cost is 10% to 20% lower.
Three-year total for three builds
An entry 3090 build is about $2,700 over three years, a 4090 build about $4,300, and an RTX PRO 6000 build over $18,000. Card prices are the current used medians:
| Model | eBay | Rest of the build | Electricity, 3 years | 3-year total | Per day |
|---|---|---|---|---|---|
 GeForce RTX 3090 | $1,625 | $700 | $3131,840 kWh | $2,638 | $2 |
| GeForce RTX 4090 | $3,200 | $700 | $4022,365 kWh | $4,302 | $4 |
| RTX PRO 6000 Blackwell | $16,985 | $700 | $5363,154 kWh | $18,221 | $17 |
| Mac Studio M4 Max 128GB | — | $0 | $4292,523 kWh | — | — |
Electricity at $0 per kWh, 8 hours a day at 60% of the card's rated power. "Rest of the build" is a flat $700 for CPU, board, RAM, PSU and case; unified-memory machines are complete systems, so it is zero for them. Resale value is not counted.
The Mac Studio is a complete machine, so "rest of the build" is zero. Its 480W rating overstates inference draw, which is closer to 100W, so its electricity line is high; it does not change where it lands in the table.
What does the same usage cost on the API?
API prices spread far wider than hardware: the same million output tokens cost $0.28 from DeepSeek V4 Flash and $10 from Claude Sonnet 5, a 35x gap. September 2026 list prices per million tokens (input / output):
| Model | Input | Output | Closest local tier |
|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 | 32B dense or 30B-class MoE |
| DeepSeek V4 Pro | $0.435 | $0.87 | No home equivalent |
| GPT-5.6 luna | $0.20 | $1.20 | 32B class |
| GPT-5.6 terra | $2.00 | $12.00 | No home equivalent |
| Claude Sonnet 5 | $2.00 | $10.00 | No home equivalent |
Assuming an even input/output split, three years of API at three typical volumes:
| Tokens per day | 3-year total | DeepSeek V4 Flash | GPT-5.6 luna | Claude Sonnet 5 |
|---|---|---|---|---|
| 50K (light personal) | 55M | $12 | $39 | $330 |
| 300K (heavy personal / small team) | 330M | $69 | $231 | $1,970 |
| 1M (batch jobs / product) | 1.1B | $231 | $770 | $6,570 |
Where is the break-even?
Break-even tokens per day = local three-year cost ÷ API price per token ÷ 1,095 days. With the 3090 build at $2,700:
- Against DeepSeek V4 Flash (blended $0.21 per million): 11.7 million tokens a day. A 3090 runs a 32B model at 4-bit near 32 tokens/s, about 900K tokens in 8 hours. It never catches up.
- Against GPT-5.6 luna (blended $0.70 per million): 3.5 million tokens a day, also beyond one card.
- Against Claude Sonnet 5 (blended $6 per million): about 410K tokens a day, four hours of generation. Heavy users get there.
Mind the premise of the third line: you must accept a 32B-class local model in place of Sonnet 5. In most workloads that is a capability step down, and the savings get paid back in rework time.
When is local clearly worth it?
Four cases need no spreadsheet: data that cannot leave the machine, offline operation, fine-tuning and large batch jobs.
- Data cannot leave the machine: legal documents, medical records, unreleased code, financial data. One compliance incident costs more than the card.
- Offline: workshops, ships, air-gapped offices.
- Fine-tuning: a LoRA run on an 8B to 14B model finishes in hours on a 24GB card. API fine-tuning bills per token and never hands you the weights.
- Batch: classifying, embedding or summarizing hundreds of thousands of documents runs overnight on one 3090, where an API means bills and rate limits.
- You were buying a gaming card anyway: the card is paid for, the marginal cost is electricity, and local inference is nearly free.
When should you not buy a card?
Personal use under about 10K tokens a day, people who only want the strongest models, and people without a desktop should not buy a GPU for LLMs.
- A few dozen chat turns and some emails a day cost less per year on the API than the card's electricity.
- If you want Sonnet 5 or V4 Pro-level ability, there is no local equivalent; a card does not solve it.
- Without a desktop the one-time outlay doubles. For the Mac route, see Mac unified memory vs a discrete GPU.
- Budget-limited but decided: start with the cheapest card at every VRAM tier.