RTX PRO 5000 48GB Price & Used Prices: Which Local LLMs Can It Run?
Price alert
We'll email you when the eBay (US) price reaches your target.
The RTX PRO 5000 Blackwell 48GB (NVIDIA, 2025) runs up to 35B at 8-bit and 30B-class MoE at about 129 tokens/s. NVFP4 models need a recent vLLM.
eBay
RTX PRO 5000 48GB price history
eBay
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Full RTX PRO 5000 48GB specifications
- Memory48 GB · GDDR7 ECC
- Bandwidth1,344 GB/s
- FP16 / BF16258 TFLOPS
- FP8 / INT8516 / 516 TOPS
- Bus width
- 384-bit
- Architecture
- Blackwell (GB202)
- Cores
- 14,080
- Tensor cores
- 440
- Ecosystem
- CUDA
- Interface
- PCIe 5.0 x16
- NVLink
- Not supported
- Power
- 300 W
- Slots
- 2
- Length
- 267 mm
- Type
- Workstation
- Launch date
- Mar 18, 2025
Best RTX PRO 5000 48GB deals on eBay
Hand-picked deals from trusted eBay sellers. Shop with confidence.
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
UsedNVIDIA RTX PRO 5000 Blackwell 48GB GDDR7 AI Workstation Graphics Card - New OEM$7,299$7,699-5%Buy on eBay
UsedDELL NVIDIA RTX PRO 5000 BLACKWELL GPU 48GB GDDR7 4XDP GRAPHICS VIDEO CARD 21PY4Top Rated$7,300Buy on eBay
UsedNVIDIA RTX PRO 5000 BLACKWELL GPU 48GB GDDR7 4XDP VIDEO CARD 900-5G153-2250-000$7,490Buy on eBay
NewNew NVIDIA RTX PRO 5000 Blackwell 48GB GDDR7, AI Workstation Graphics Card (OEM)$7,590Buy on eBay
NewDELL NVIDIA RTX PRO 5000 Blackwell 48GB GDDR7 PCIe 5.0 Retail Graphics CardTop Rated$7,900Buy on eBay
UsedNVIDIA RTX PRO 5000 Blackwell 48GB GDDR7 Graphics CardTop Rated$7,900Buy on eBay
What LLMs can a single RTX PRO 5000 48GB run?
A single PRO 5000 48GB runs up to 35B at 8-bit with full 256K context; 4-bit makes dense models 60-70% faster. Its ceiling is 70B-class at 4-bit.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 12B12B | 4.7 GB | Runs | 161.7 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 155.9 | 256K |
 Qwen3.6 27B27.8B | 11.8 GB | Runs | 76.3 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 89.5 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 164.9 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 73.7 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 173.1 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 12B12B | 7.1 GB | Runs | 117.4 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 136 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 55.7 | 256K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 56.6 | 256K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 149.5 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 50.3 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 152.2 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 12B12B | 12.7 GB | Runs | 71.7 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 113.4 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 34 | 256K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 33.6 | 256K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 125 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Runs | 29.6 | 202K |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Runs | 128.7 | 256K |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Needs RAM offloadMoE only: expert weights sit in system RAM, so speed is estimated from 70 GB/s RAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
What LLMs can multiple RTX PRO 5000 48GB cards run?
Four PRO 5000 48GB cards (192GB) run DeepSeek V4 Flash at 4-bit at about 86 tokens/s, and 8-bit fits too; two suit gpt-oss 120B. No NVLink.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 2 cards | $15,800 | 96 GB | 130 | 600 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 4 cards | $31,600 | 192 GB | 110 | 1,200 W |
 Hy3295B (21B active) | 107.2 GB | 4 cards | $31,600 | 192 GB | 58.1 | 1,200 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 4 cards | $31,600 | 192 GB | 92.5 | 1,200 W |
| GLM 5.3744B (40B active) | 253.9 GB | 8 cards | $63,200 | 384 GB | 51.7 | 2,400 W |
| Hy4 preview770B (49B active) | 307.2 GBest. | 8 cards | $63,200 | 384 GB | 39.9 | 2,400 W |
 Muse Spark Open816B (42B active) | 311.2 GB | 8 cards | $63,200 | 384 GB | 46.6 | 2,400 W |
 Kimi K2.7 Code1,000B (32B active) | 339.5 GB | 8 cards | $63,200 | 384 GB | 61.2 | 2,400 W |
| Kimi K2.61,000B (32B active) | 340.5 GB | 8 cards | $63,200 | 384 GB | 61.1 | 2,400 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 4 cards | $31,600 | 192 GB | 114.1 | 1,200 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 4 cards | $31,600 | 192 GB | 85.6 | 1,200 W |
| Hy3295B (21B active) | 182.2 GB | 8 cards | $63,200 | 384 GB | 44.7 | 2,400 W |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | 8 cards | $63,200 | 384 GB | 63.4 | 2,400 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 4 cards | $31,600 | 192 GB | 83.5 | 1,200 W |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | 8 cards | $63,200 | 384 GB | 88.4 | 2,400 W |
| Hy3295B (21B active) | 317.7 GB | 8 cards | $63,200 | 384 GB | 31.5 | 2,400 W |
| GLM 5.3 Flash320B (18B active) | 341 GB | 8 cards | $63,200 | 384 GB | 42.6 | 2,400 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Running 27B to 31B dense models on one card
About 50 to 57 tokens/s at 4-bit, and about 30 to 34 at 8-bit
- You want to use models quantized in FP4 (NVFP4)
- Running two services on one card at once
It supports MIG and can be split into two 24GB instances
Not a good fit
- Running models above 70B on one card
70B-class models just fit at 4-bit; for anything larger, look at the 72GB version
- Mostly long documents and batched requests
Its compute is lower than the RTX 6000 Ada's (about 258 vs. 364 TFLOPS of BF16)
Before you buy
How MIG splits the card
MIG can split this card into two 24GB instances, each with its own memory and compute, effectively two 24GB cards: each instance can hold Qwen3.8 27B at 4-bit (16.5 GB), but not at 8-bit (29 GB).
FP4 needs recent software
This generation of workstation cards uses the architecture code sm_120, unlike the data center Blackwell (sm_100). vLLM used to recognize only the latter, so loading an NVFP4 MoE model on this card threw an error (issue #33416, now fixed). Upgrade vLLM to a recent version before using NVFP4 models.
A drop-in replacement for the 6000 Ada
Dual-slot, about 267 mm long, 300W, one 16-pin power connector: all the same as the RTX 6000 Ada, and its blower also exhausts through the bracket. Swapping it into an older workstation usually needs no changes to the case or power supply.
Compared with similar GPUs
RTX PRO 5000 48GB vs RTX 6000 Ada, GeForce RTX 4090 48GB, and RTX PRO 5000 72GB: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | RTX PRO 5000 48GB | RTX 6000 Ada | GeForce RTX 4090 48GB | RTX PRO 5000 72GB |
|---|---|---|---|---|
| Memory | 48 GB | 48 GBSame | 48 GBSame | 72 GB50% more |
| Bandwidth | 1,344 GB/s | 960 GB/s29% lower | 1,008 GB/s25% lower | 1,344 GB/sSame |
| BF16 / FP16 compute | 258 TFLOPS | 364 TFLOPS41% higher | 330 TFLOPS28% higher | 258 TFLOPSSame |
| INT8 compute | 516 TOPS | 729 TOPS41% higher | 661 TOPS28% higher | 516 TOPSSame |
| Qwen3.8 27B speed (2-bit) | 89.5 t/s | 66.3 t/s26% slower | 69.3 t/s23% slower | 89.5 t/sSame |
| Largest dense model (4-bit) | 79B | 79BSame | 79BSame | 122B54% more |
| Power | 300 W | 300 WSame | 450 W50% higher | 300 WSame |
| Interface | PCIe 5.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x16 |
| Released | 2025-03 | 2022-09 | 2022-10 | 2025-12 |
| eBay median |
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
RTX PRO 5000 48GB rental price per hour
On-demand rental of the RTX PRO 5000 48GB starts at $0.8 per hour (Vast.ai), about $195/month at 8 hours a day. Buying one (eBay median $7,900) pays off in about 7.2 years, 3.6 years or 1.2 years at 4, 8 or 24 hours a day.
| Option | Price | 4 h/day | 8 h/day | 24 h/day | View |
|---|---|---|---|---|---|
 Vast.aiRentUpdated 49 min. ago | $0.8/hr | $97/mo | $195/mo | $585/mo | Rent |
 RunPodRentUpdated 49 min. ago | $0.82/hr | $100/mo | $200/mo | $599/mo | Rent |
 eBayUpdated 6 hr. ago | $7,900 | Payback ~7.2 yrsPower $7/mo | Payback ~3.6 yrsPower $13/mo | Payback ~1.2 yrsPower $39/mo | Buy |
Break-even uses the US residential average ($0.18/kWh) and 300 W rated power, excluding resale value and the rest of the build.
Rental links are referral links: we may receive a commission or credit if you sign up through them. As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Amazon.jp
Xianyu