Key Takeaways
- Local RTX 4070: $350-500 used + $0.02/hour idle power = $0.02-0.05/hour all-in cost
- Cloud Lambda Labs RTX 4090: $2.50/hour + storage + bandwidth
- Cloud Paperspace A100: $0.60/hour; decent for LLM fine-tuning
- Cloud AWS g4dn.2xlarge (V100): $0.98/hour + compute markup (10-20% premium)
- Breakeven: Local RTX 4070 vs Lambda Labs RTX 4090 = 140-280 compute hours = 4-7 months of weekly use
- For unpredictable workloads: Cloud cheaper (no upfront cost). For consistent 5+ hours/week use: Local is 5-10x cheaper
- Hidden cloud costs: Bandwidth egress ($0.02-0.10/GB), GPU reservation fees, data transfer to/from cloud ($0.05-0.15/GB)
- Local hidden costs: Cooling (included in building), network latency (~100ms), GPU replacement every 5-7 years
📍 In One Sentence
A local RTX 4070 costs $0.02-0.05 per inference hour versus $0.50-2.50/hour for cloud GPUs like Paperspace, Lambda Labs, or AWS -- local is 10-50x cheaper per hour and breaks even within 2-6 months for any consistent workload.
💬 In Plain Terms
Renting a cloud GPU is convenient but adds up fast if you use it regularly. A used RTX 4070 pays for itself in a few months of consistent use and then runs essentially free afterward. Cloud GPUs still make sense for occasional bursts of heavy compute or specialized hardware like H100s that you can't justify owning, but for steady, everyday inference, local hardware wins on cost.
What Is the Hourly Cost: Local vs Cloud?
Local RTX 4070 (used $350): 250W TDP, US electricity $0.14/kWh = $0.035/hour compute cost + $0.008/hour depreciation (5-year lifespan) = $0.043/hour total.
Local RTX 4090 (EOL; used $2,000-2,600): 450W TDP = $0.063/hour compute + $0.053/hour depreciation = $0.116/hour.
Cloud Lambda Labs RTX 4090: $2.50/hour (no depreciation, but includes storage and support). 10-50x more expensive than local.
Cloud Paperspace A100 (80GB): $0.60/hour; reasonable for fine-tuning, still 10-15x more than local RTX 4070.
Cloud AWS g4dn.2xlarge V100: $0.98/hour list price, ~$1.20 on-demand with markup.
When Does a Local GPU Break Even with Cloud Compute?
Local RTX 4070 ($350) vs Cloud Lambda Labs RTX 4090 ($2.50/hr): Breakeven = $350 / ($2.50 − $0.04) = 143 compute hours = 29 weeks at 5 hrs/week.
Local RTX 4090 (EOL; $2,000-2,600 used) vs Cloud Lambda Labs ($2.50/hr): Breakeven = 960-1,200 compute hours = 184-230 weeks at 5 hrs/week (3.5-4.5 years).
Local RTX 4070 vs Cloud Paperspace A100 ($0.60/hr): Breakeven = $350 / ($0.60 − $0.04) = 625 hours = 150 weeks at 5 hrs/week (almost 3 years).
For burst users (5-10 hours/month): Cloud is cheaper. For consistent users (5+ hours/week): Local is cheaper.
How Do Cloud GPU Providers Compare?
Lambda Labs (April 2026): RTX 4090 $2.50/hr, RTX 6000 Ada $3.50/hr, H100 $4.50/hr. No hourly reservation; pay-as-you-go. Excellent for bursts.
Paperspace (April 2026): A100 40GB $0.51/hr, RTX A6000 $0.73/hr. Cheaper than Lambda Labs but older hardware. Good for training.
AWS (April 2026): g4dn.2xlarge V100 $0.98/hr on-demand, ~$0.40/hr reserved (1-year commitment). ec2 g4dn.xlarge cheaper ($0.526/hr) but single V100.
Google Colab Pro: $10/month unlimited (L4 GPU), $50/month with A100. Good value for light users.
RunPod (April 2026): RTX 4090 $0.44/hr, A100 $1.29/hr. Cheaper than Lambda Labs; smaller provider.
What Is the 1-Year Cost of Ownership?
Local RTX 4070 at 20 hrs/week (1,040 hours/year): $350 GPU + (1,040 × $0.03) electricity = $381 total.
Cloud Lambda Labs RTX 4090 at 20 hrs/week: 1,040 × $2.50 = $2,600 total.
Cost ratio: Cloud is 6.8x more expensive than local for this workload.
Local RTX 4090 (EOL; used $2,000-2,600) at 20 hrs/week: $2,300 (est.) + (1,040 × $0.06) = $2,362 total (4-7x more expensive than new RTX 4070).
Cloud Paperspace A100 at 20 hrs/week: 1,040 × $0.60 = $624 total (cheaper than local RTX 4090 for 1 year, but becomes more expensive in year 2).
Frequently Asked Questions
Can I use cloud GPUs for 24/7 continuous inference?
Yes, but costs escalate fast. 24/7 Lambda Labs RTX 4090: $2.50 × 8,760 = $21,900/year. Local RTX 4090 (EOL; $2,000-2,600 used): $460/year depreciation + $551/year power = $1,011/year total.
What about egress bandwidth costs on cloud?
AWS/Google charge $0.02-0.10/GB for data leaving the cloud. Running a local API that returns 100MB/day = $60-300/month egress. Local has zero egress costs.
Does local require a dedicated server or can I use my gaming PC?
Your gaming PC works fine, but it can't serve both gaming and LLM inference simultaneously. Many use underutilized servers or mini PCs instead.
Are cloud GPU prices guaranteed or can they change?
Prices fluctuate (AWS spot instances vary 30-50%). Lambda Labs pricing is stable. Local GPU prices depend on the used market.
What if my local GPU fails mid-inference?
Downtime until replacement. Cloud provides redundancy via multi-region deployments. Local requires backup GPU or failover to cloud.
Can I use cloud GPUs for fine-tuning instead of just inference?
Yes. Fine-tuning is more cost-effective on cloud (better cooling for training stability). Cloud fine-tuning then deploy on local for inference is a common pattern.
What about cost for EU customers who need GDPR compliance?
EU customers comparing cloud options need to factor in compliance: EU GDPR cloud GPU comparison → covers Hetzner (€0.40/hr), Scaleway, OVHcloud, and GDPR-compliant alternatives with full cost breakdowns.
Common Mistakes When Comparing Local and Cloud GPU Costs
- Forgetting depreciation. A local GPU depreciates ~20% per year; include this in total cost.
- Ignoring bandwidth costs. Cloud APIs that output large embeddings/tensors incur egress charges (~$0.02/GB).
- Comparing old GPU prices to cloud. The RTX 4090 is EOL (no longer manufactured); used market prices are $2,000-2,600, making it less economical vs. current-gen GPUs like RTX 5090 or RTX 4080 Super.
- Underestimating infrastructure overhead. Running a local cluster (cooling, redundancy, monitoring) costs 10-20% more than a single GPU.
- Assuming cloud is only for bursts. For unpredictable workloads (spiky traffic), cloud wins. For baseline load, local is cheaper.
Sources
- Lambda Labs GPU pricing: lambdalabs.com/service/gpu-cloud (April 2026)
- Paperspace GPU pricing: paperspace.com/pricing
- AWS EC2 GPU instance pricing: aws.amazon.com/ec2/pricing/on-demand (April 2026)