A100 Price Guide with Performance Metrics and Buying Options

In 2025, AI workloads continue to scale rapidly, and NVIDIA's A100 GPU remains a top-tier choice for deep learning and HPC applications. Introduced in 2020 based on the Ampere architecture, this GPU delivers industry-leading performance—even with the arrival of newer GPUs. In fact, with a 312 TFLOPS mixed-precision throughput, it's still considered one of the best investments for AI infrastructure.

But what does the A100 price look like today? This guide dives into current specs, real-world performance comparisons, and buying options—from outright hardware purchases to cloud rentals via major hyperscalers or optimized GPU cloud providers.

A100 Technical Overview

The A100 comes in two main variants—40 GB and 80 GB HBM2e—with key specs as follows:

CUDA Cores: 6,912

Tensor Cores: 432

Memory: 40 GB (HBM2) or 80 GB (HBM2e)

Memory Bandwidth: 1.6 TB/s (40 GB), 2.0 TB/s (80 GB)

Mixed-precision Performance: 312 TFLOPS

FP64: 19.5 TFLOPS

Peak INT8: 624 TOPS

MIG Support: Up to 7 GPU instances per A100

NVLink: 600 GB/s interconnect in SXM4 models

That explains its adoption across major AI frameworks and data center deployments.

A100 Price—Purchasing Hardware in 2025

For organizations building on-prem GPU servers, here's what you can expect:

A100 40 GB PCIe/SXM4 (new): $7,500–$10,000

A100 80 GB: $9,500–$14,000

eBay/refurbished SXM4 cards: Reports of $2,600 for used A100 40 GB units

Hardware pricing depends on form factor (PCIe vs. server-grade SXM), warranty status, and market conditions. Buying new remains expensive but delivers the highest reliability and form factor compatibility.

Cloud Pricing Overview

Renting A100 GPUs from cloud providers is often more economical for short-to-medium AI projects.

Hyperscaler Clouds (AWS, Azure, GCP):

Google Cloud: a2-highgpu-1g (A100 40 GB) ~$4.27/hr

AWS p4d.24xlarge (8 A100s): combined cost ~$32.77/hr → per-GPU ~$4.10/hr

Azure ND96asr A100 v4: ~$3.40/hr per GPU

Specialized GPU Cloud Providers:

Thunder Compute: A100 40 GB at $0.66/hr

Vast.ai: A100 SXM4 at ~$1.27/hr

Lambda GPU Cloud: A100 40 GB at $1.29/hr

TensorDock: $1.63/hr (on-demand) or $0.67/hr (spot)

Hyperstack: A100 80 GB at $1.35/hr (OD) / $0.95/hr (reserved)

DataCrunch: A100 SXM at $1.12–1.15/hr

Summary Table: A100 Hourly Rates (USD)

The best deals are ~3–6x cheaper than big cloud providers, ideal for training workloads with flexibility

Performance Considerations: 40 GB vs 80 GB

Opting for 40 GB vs 80 GB impacts price and capability:

40 GB is sufficient for most NLP tasks, common model training, and inference

80 GB offers:

Higher memory bandwidth (2.0 TB/s vs 1.6 TB/s)

Larger memory for big models, complex ML workloads

For price-sensitive setups, 40 GB units offer great value. For large, memory-heavy tasks, splurging on 80 GB can be essential.

Real-World Monthly Cost Estimates

Assuming 100 GPU-hours/month:

Thunder Compute: 100 × $0.66 = $66

Lambda: 100 × $1.29 = $129

Hyperstack reserved: 100 × $0.95 = $95

Hyperscalers: $340–427/month

A100 access can cost as little as $66/month or scale to $400+ depending on the provider and usage.

Choosing the Right Option

Budget Constraints?

Pick Thunder Compute or Hyperstack reserved options

Need Performance & Stability?

Choose Lambda, TensorDock, or DataCrunch

Require Enterprise Support & Managed Infrastructure?

Hyperscalers offer robust SLAs, integrated tooling, and reliability

On-Prem or Hybrid Use Cases?

New cards ($7.5K–14K) for full control or refurbished units (~$2.6K used)

Scale & Volume Benefits?

Reserved contracts (e.g., 1–3 years) can lower costs further

Conclusion: A100—High Performance, Multiple Access Paths

The NVIDIA A100 remains one of the best-performing GPUs available today for AI and HPC. You can buy a top-end 80 GB version for $9.5K–14K, or secure cost-effective 40 GB options starting at $7.5K. Renting via cloud varies widely—from under $1/hr with Lightning-fast specialized clouds, to $4+/hr with traditional clouds.

Choose based on your workload scale, performance needs, budget timeframe, and infrastructure footprint. For lean AI teams, providers like Thunder Compute and Lambda offer unbeatable value. Enterprises needing support at scale may still find hyperscalers a better fit.