Sarah
Indie AI dev · AustinReplicate runs were eating her side-project budget.
Click "Fine-tune" → autopilot picks a cheap 4090 → her existing OpenAI code just works.
~$1.60 for a 4-hour Llama-3 LoRA
The cheapest way to fine-tune & train AI models
Fine-tune Llama-3 for under $1.
Train your own model for under $10.
Wattlend turns idle home GPUs into the world's lowest-cost training infrastructure. LoRA, QLoRA, full SFT, or train-from-scratch — same workflow as Modal or Together, 10–50× cheaper because the GPUs are sitting in someone's gaming PC, not a data center.
No setup · No minimums · Pay per second of GPU time · Adapter weights land in your storage on completion
Pricing
$0.30/hr
for a 4090; A100s ~$1.20/hr
Cold start
60–90s
image pull + container ready
Methods
LoRA · QLoRA · SFT
axolotl under the hood
Output
Yours
adapter weights in your R2 bucket
Same fine-tune job, different bill
Why we're 10–50× cheaper than the incumbents
Modal, OpenPipe, Together, and Predibase rent GPUs from AWS, GCP, or their own data centers — they pay full rack-power prices and pass it on. Wattlend rents from gamers' idle PCs that already paid for themselves. The cost difference is physics, not magic.
| Job | Modal A10G / A100 serverless | OpenPipe managed fine-tune API | Together fine-tune endpoint | Wattlend consumer GPU marketplace |
|---|---|---|---|---|
LoRA fine-tune Llama-3 8B · Alpaca · 3 epochs (~30 min compute) | ~$4 | ~$8 | ~$6 | $0.50 |
QLoRA fine-tune Llama-3 70B · subset · 1 epoch (~3 hr on A100) | ~$12 | ~$25 | ~$18 | $4 |
Train from scratch Phi-3 mini · custom dataset · 5 epochs (~6 hr on 4090) | ~$8 | n/a | n/a | $2 |
Wattlend price: actual, based on $0.30/hr 4090 + image pull time. Competitor prices: directional estimates from public pricing pages (Q1 2026) for an equivalent job. Your exact bill depends on dataset size + epochs.
Universal compute
Any hardware. Any workload.
One marketplace.
Whatever GPU you own, whatever model your buyer needs — Wattlend routes between them in milliseconds. No custom integration, no per-hardware setup, no SDK lock-in.
Any hardware
🎮
RTX 4090 / 3090
24 GB consumer GPUs
⚡
A100 / H100
40-80 GB datacenter
🎯
RTX 4070 / 3080
12-16 GB mid-range
🍎
Mac M2 / M3
Apple Silicon
🔥
AMD MI300
ROCm-compatible
💻
CPU + Ollama
no GPU needed
Any workload
💬
Chat
Llama, Mistral, Phi
🎨
Image gen
SDXL, Flux
🎙
Transcribe
Whisper
🔎
Embeddings
BGE, E5
🧪
Fine-tune
LoRA, full SFT
🔧
Custom Docker
your own image
Any tool you already use
OpenAI SDKLangChainLlamaIndexCursorContinueAiderOpen WebUILibreChatChatboxcurl
Watch it work · 25 seconds
From idle GPU to earning, in 6 steps.
Tom owns an RTX 4090. Sarah wants to fine-tune Llama-3. Watch how Wattlend matches them, runs the workload, and splits the money — auto-plays, pause anytime.
🧑💻
Tom · Ohio
RTX 4090 · 24 GB
⚡ Wattlend
Agent connected
✓ Listed at $0.40/hr
Frame 1 of 6
1. Tom installs the agent on his gaming PC.
Tom's RTX 4090 sits idle 16 hours a day. He pastes a one-line install command. The Wattlend agent auto-detects his hardware, sets a fair hourly rate, and lists the machine — no manual config.
How the marketplace works
Supply meets demand. We're the bridge.
Idle GPUs on one side. AI workloads on the other. Wattlend matches them, handles the plumbing, takes 30%. The seller earns money they wouldn't have. The buyer pays less than they would have. Same GPU. Better outcome for both.
Supply
Idle hardware looking for income
- Gaming PC with an RTX 4090 (idle 16h/day)
- Old crypto mining rigs (post-merge paperweights)
- AI workstations used nights/weekends only
- University lab GPUs idle off-hours
- Small data centers / co-lo with spare slots
Earns $80–$1,200/mo on hardware that would otherwise sit idle.
$$ pays seller
locked hourly · 70% to seller
⚡ Wattlend
Matches buyers + sellers. Handles routing, billing, payouts. Takes 30%.
🖥 compute
OpenAI-compatible API · per-rental URL
Demand
Workloads that need more compute than the laptop has
- Indie devs fine-tuning Llama-3 on side projects
- AI startups burning AWS budget on inference
- Image-gen bursts (1,000 SDXL renders, then nothing)
- Audio transcription jobs (hours of podcasts)
- Data scientists doing weekend analysis
Pays 60% less than AWS. No Docker, no SSH, no instance types — just "run my job."
No one in this picture would have used the GPU otherwise. Tom's 4090 was sleeping. Sarah was about to give up on her side project. Wattlend's 70/30 split funds the platform — the buyer pays less than they would have, the seller earns money they wouldn't have, the platform takes a fair cut for matchmaking.
Who uses Wattlend
Real people on both sides of the marketplace
Two who buy compute, two who sell it — with the numbers that make Wattlend the right call for each.
Buyers
People who need compute they don't own.
Mike
AI startup founder · 3 employeesAWS GPU instances were 60% of his startup's burn.
Same OpenAI-compatible API — flip a base URL, cut compute spend ~60%.
$4,800/mo → ~$1,900/mo
Sellers
People with idle hardware to monetize.
Tom
Gamer · college studentHis $1,800 RTX 4090 sits idle 18 hours a day.
Install the agent, set quiet hours for gaming, earn while you sleep.
$80–$140/mo net after electricity
Dave
Ex-ETH miner · Sacramento32 RTX 3080s/3090s went from earning to paperweights post-Merge.
Bulk-list a fleet of old mining hardware, predictable hourly income.
$400–$1,200/mo across his rigs
Works with your tools
Drop-in for the AI tools you already use
Every rental gives you a private URL and Bearer token. Change two lines in any OpenAI-format client and point it at Wattlend — keep everything else. Your existing code, chat UIs, and dev tools just work.
LLM APIs
- OpenAI SDK✓ native
- LangChain✓ native
- LlamaIndex✓ native
- Anthropic SDK / Claudevia proxy
Chat UIs
- Open WebUI✓ native
- LibreChat✓ native
- Chatbox✓ native
- Our /chat✓ native
Dev tools
- Cursor✓ native
- Continue✓ native
- Aider✓ native
- Custom OpenAI SDK code✓ native
curl https://wattlend.com/v1/rentals/$RENTAL_ID/chat/completions \
-H "Authorization: Bearer $RENTAL_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"model": "llama-3-8b-instruct",
"messages": [{"role":"user","content":"Hello from a rented GPU."}]
}'Each rental's detail page includes step-by-step setup for the most common clients. Same Bearer token, same URL pattern.
For builders
Fine-tune any model. Train your own. Run it anywhere.
LoRA in 10 minutes for $1. QLoRA a 70B for $4. Train Phi-3 from scratch on your data for $2. Adapter weights save to your storage on completion — drop into Ollama, vLLM, or our chat endpoint. Same workflow as Modal or Together, an order of magnitude cheaper.
🧪 Sign up to start fine-tuning →
For GPU owners
Power the cheapest training infra on the planet.
Your idle RTX 4090 or 3090 becomes the supply behind every $1 LoRA buyers ship. One-command install, runs as a background Windows service, you earn 70% of every fine-tune that touches your hardware. Withdraw to your bank on demand via Stripe.
⚡ Sign up to start earning →