The NVIDIA Tesla T4 rents from $0.16/hr to $0.59/hr in 2026, an aging but cheap Turing GPU built for inference, video processing, and lightweight machine learning. It carries 16GB of GDDR6 memory and draws just 70W.
This guide covers current T4 cloud pricing, its full specs, buy price, which models fit in 16GB, and how it compares to modern alternatives.
Takeaways
- The T4 rents from $0.16/hr to $0.59/hr, with Vast.ai being the cheapest.
- It has 16GB of GDDR6 on the Turing architecture, delivering 8.1 TFLOPS of FP32.
- Used T4 cards start around $520, with new cards up to about $1,340.
- Most hyperscalers have phased it out in favor of newer GPUs.
- At current prices, newer GPUs offer far better performance per dollar, including Thunder Compute's A6000 at $0.35/hr with 3x the VRAM.
NVIDIA T4 Cloud Pricing
The pricing information in this guide is reviewed weekly.
| Provider | GPU / Instance | On-Demand $/GPU-hr* | Notes |
|---|---|---|---|
| Vast.ai | T4 | $0.16 | Median of 1 verified US/CA host, priced as compute plus 100GB storage; bandwidth billed separately. |
| AWS | g4dn.xlarge | $0.53 | Cheapest US region (us-east-1); ranges $0.53-$5.22/GPU-hr across 28 US options. |
| Azure | Standard_NC4as_T4_v3 | $0.53 | Cheapest US region (eastus); ranges $0.53-$0.63/GPU-hr across 8 US options. |
| Google Cloud | n1-standard-4 + 1xT4 | $0.54 | Cheapest US region (us-central1); ranges $0.54-$0.64/GPU-hr across 10 US options. |
| Modal | T4 | $0.59 | |
| Last reviewed on October 1, 2026. | |||
Thunder Compute offers the RTX A6000 at $0.35/hr, which has 3x the VRAM of the T4 and far more compute.
Methodology
- On-demand only. Pricing reflects public on-demand rates, excluding reserved and spot discounts.
- Comparable instances. Only single-GPU, standard configurations are used for the headline comparison.
- Marketplace caveat. Vast.ai rates reflect live listings, which fluctuate with availability.
- US regions, USD. Hyperscaler pricing is standard on-demand in US regions.
T4 Cost Benchmark
The 10-hour column shows what a single T4 costs across major providers for a typical inference session.
| Provider | On-Demand $/GPU-hr | 10-Hour Cost |
|---|---|---|
| Vast.ai | $0.16 | $1.61 |
| AWS | $0.53 | $5.26 |
| Azure | $0.53 | $5.26 |
| Google Cloud | $0.54 | $5.40 |
| Modal | $0.59 | $5.90 |
NVIDIA T4 Specs at a Glance
The NVIDIA T4 pairs 16GB of GDDR6 at 320GB/s with 2,560 CUDA cores and 320 Turing Tensor Cores, at CUDA compute capability 7.5. It delivers 8.1 TFLOPS of FP32 and 65 TFLOPS of FP16, in a 70W, single-slot, passively-cooled package.
The T4 was purpose-built for inference and high-density servers, not large-model training. It has no NVLink and no FP8 support, so it cannot pool memory across cards or use the FP8 path that accelerates modern transformer inference.
| Specification | NVIDIA T4 |
|---|---|
| Architecture | Turing (TU104) |
| CUDA Cores | 2,560 |
| Tensor Cores | 320 (2nd Gen) |
| Compute Capability | 7.5 |
| GPU Memory | 16GB GDDR6 |
| Memory Bandwidth | 320GB/s |
| FP32 Performance | 8.1 TFLOPS |
| FP16 Performance | 65 TFLOPS |
| INT8 Performance | 130 TOPS |
| NVLink | No |
| TDP | 70W |
| Release Date | September 2018 |
NVIDIA T4 Hardware Price
Used NVIDIA T4 cards start around $520 on eBay, with new units up to $1,340, well below the $2,299 launch MSRP. The T4 is a passive, server-only card, so it needs a chassis with the right airflow rather than a standard desktop.
Cloud rental sidesteps the hardware and chassis requirements entirely: at $0.16/hr on the marketplace, a T4 rents for a fraction of the cost of buying and housing one.
| Option | Approximate Cost1 | Notes |
|---|---|---|
| Used T4 (16GB) | From $520 | eBay used listings; server-only, passive cooling. |
| New T4 (16GB) | Up to $1,340 | eBay and reseller new listings; outliers excluded. |
| Cloud rental | $0.16-$0.59/hr | No hardware to buy or house, instant access. |
| 1 Secondary-market figures vary with condition, seller, and availability. | ||
What Models Fit on a 16GB T4?
16GB runs 7B LLMs in 4-bit or 8-bit and small image models, but not FP16 inference on anything above about 7B. The T4 also lacks FP8, so it cannot use the fastest low-precision path on modern models.
| Model | Type | Setup | Approx VRAM1 | Fits 16GB? |
|---|---|---|---|---|
| Llama 3 8B | LLM | 4-bit | ~6GB | Yes |
| Mistral 7B | LLM | 8-bit | ~8GB | Yes |
| Stable Diffusion XL | Image | FP16 | ~12GB | Yes |
| Llama 3 8B | LLM | FP16 | ~16GB | Tight |
| 1 Approximate VRAM for a single inference at typical settings. LLM figures are weights plus 20-30% for KV cache and activations. | ||||
T4 vs Newer GPUs
Modern GPUs outperform the T4 at similar or slightly higher prices, usually with much more VRAM. The table below compares the T4 to two common upgrades.
| Feature | NVIDIA T4 | NVIDIA RTX A6000 | NVIDIA A100 80GB |
|---|---|---|---|
| Architecture | Turing | Ampere | Ampere |
| GPU Memory | 16GB GDDR6 | 48GB GDDR6 | 80GB HBM2e |
| Memory Bandwidth | 320GB/s | 768GB/s | 1,935GB/s (PCIe); 2,039GB/s (SXM) |
| Released | 2018 | 2020 | 2020 |
| On-Demand Cloud Price | $0.16 - $0.59/hr | $0.35 - $1.89/hr | $1.09 - $5.03/hr |
AWS G4 Instances
AWS offers NVIDIA T4 GPUs in its EC2 G4 Instances, spanning single-GPU to 8-GPU configurations. As usual for AWS, egress fees are $0.09/GB after the first free 100GB per month.
G4 instance characteristics:
- NVIDIA T4 GPUs
- 4-96 vCPUs
- 16-384GB RAM
- 125-1800GB NVMe SSD storage
| SKU | T4 GPUs | vCPUs | RAM | Hourly Price | Price Per-GPU |
|---|---|---|---|---|---|
| g4dn.xlarge | 1 | 4 | 16 GiB | $0.53-$0.63 | |
| g4dn.2xlarge | 1 | 8 | 32 GiB | $0.75-$0.90 | |
| g4dn.4xlarge | 1 | 16 | 64 GiB | $1.20-$1.45 | |
| g4dn.8xlarge | 1 | 32 | 128 GiB | $2.18-$2.61 | |
| g4dn.16xlarge | 1 | 64 | 256 GiB | $4.35-$5.22 | |
| g4dn.12xlarge | 4 | 48 | 196 GiB | $3.91-$4.69 | $0.98-$1.17 |
| g4dn.metal | 8 | 96 | 384 GiB | $7.82-$9.39 | $0.98-$1.17 |
Azure NC T4 Instances
Azure offers the T4 through the NCas_T4_v3-series, compute-optimized VMs built for cost-effective inference and CUDA-accelerated workloads on AMD EPYC processors.
| SKU | T4 GPUs | vCPUs | RAM | Hourly Price | Price Per-GPU |
|---|---|---|---|---|---|
| Standard_NC4as_T4_v3 | 1 | 4 | 28 GiB | $0.59 | |
| Standard_NC8as_T4_v3 | 1 | 8 | 56 GiB | $0.85 | |
| Standard_NC16as_T4_v3 | 1 | 16 | 110 GiB | $1.36 | |
| Standard_NC64as_T4_v3 | 4 | 64 | 440 GiB | $4.91 | $1.23 |
Google Cloud N1 Instances
Google Cloud attaches T4 GPUs to flexible N1 machine types, so the price depends on the CPU and memory you pair with the GPU. The table below is illustrative of the smallest and largest common configurations.
| SKU | T4 GPUs | vCPUs | RAM | Hourly Price | Price Per-GPU |
|---|---|---|---|---|---|
| n1-standard-4 + 1xT4 | 1 | 4 | 15 GiB | $0.54 | |
| n1-standard-96 + 4xT4 | 4 | 96 | 360 GiB | $6.85 | $1.71 |
| n1-highcpu-4 + 1xT4 | 1 | 4 | 3.6 GiB | $0.51 | |
| n1-highcpu-96 + 4xT4 | 4 | 96 | 86.4 GiB | $5.31 | $1.33 |
| n1-highmem-4 + 1xT4 | 1 | 4 | 26 GiB | $0.62 | |
| n1-highmem-96 + 4xT4 | 4 | 96 | 624 GiB | $8.14 | $2.04 |
Run Newer GPUs on Thunder Compute
Thunder Compute offers the RTX A6000 at $0.35/hr, which has 3x the T4's VRAM and far more compute for close to a hyperscaler T4 rate. It comes with per-minute billing, no minimum commitment, and VS Code and Cursor extensions that connect your editor directly to a running instance.
For larger models, Thunder also lists the A100 at $1.09/hr. See current availability and pricing on Thunder Compute →
Last Thoughts on NVIDIA T4 Pricing
The T4 is still functional and cheap, and at $0.16/hr on the marketplace it can make sense for light inference and video work. But it launched in 2018, lacks FP8 and NVLink, and its 16GB of VRAM limits it to small models.
At similar or slightly higher prices, newer GPUs deliver far more performance per dollar. For most 2026 workloads, a modern card like the A6000 or A100 will finish jobs faster and cost less per result, even if the T4's hourly rate looks lower.
FAQ
What Is the NVIDIA T4 Used For?
The NVIDIA T4 is used for inference workloads, video processing, and lightweight machine learning tasks. Its 70W design suits high-density, energy-efficient server deployments.
How Much Does the NVIDIA T4 Cost to Rent?
The NVIDIA T4 rents from $0.16/hr to $0.59/hr across providers in October 2026, with Vast.ai cheapest and hyperscalers around $0.53-$0.59/hr.
How Much VRAM Does the NVIDIA T4 Have?
The NVIDIA T4 has 16GB of GDDR6 memory with 320GB/s of bandwidth.
How Much Does an NVIDIA T4 Cost to Buy?
Used T4 cards start around $520 on eBay, and new units run up to about $1,340. Renting from $0.16/hr avoids the hardware cost entirely.
Is the NVIDIA T4 Outdated?
Yes. While still functional, the T4 launched in 2018 and is being phased out in favor of more powerful, more cost-efficient GPUs.