The gaming GPU from Nvidia is almost on par with its data center GPU, with one glaring difference: onboard memory.
Nvidia’s GeForce RTX 5090, its flagship graphics card for absolute top-end PC gaming, is now selling for more than $5,000 in some markets because some buyers are using them as alternatives to enterprise GPUs.
The RTX 5090 originally launched in January 2025 with a $1,999 MSRP, but it’s now is selling for more than $5,000 as AI system builders suck up the limited supply. The Hong Kong news outlet HKEPC first reported that 5090 cards—intended for playing games like Call of Duty—were being put in workstations in place of Nvidia’s enterprise card, the RTX 6000.
The 5090 and the 6000 have a lot in common and have very similar spec sheets. Both are built on the Blackwell architecture; the 6000 has 24,000 cores versus 21,000 in the 5090. The 6000 has a Max performance of 126 TFLOPS versus 104 TFLOPS for the 5090.
Where there is a huge gap is the onboard memory. The 5090 has 32GB of memory on the card versus 96 GB for the 6000, and that makes a huge difference with AI. With a 5090 card, you can’t load a model larger than 32GB, whereas with the 6000, you can load models as large as 96GB. And that absolutely makes a difference.
AI companies traditionally rely on specialized hardware such as Nvidia’s data-center accelerators and professional GPUs. However, the continued expansion of AI workloads has created demand for virtually any high-performance GPU that can deliver substantial compute capability.
And at $5,000 a card, the 5090 is still a bargain. The 6000 has an average price of between $12,000 and $15,000.
For some use cases and some customers, the 5090 might just be fine, says Jon Peddie, president of Jon Peddie Research. “If you’re running small models that will fit in 32GB, and don’t care about error correction, then a 5090 is a cheap AI training solution. That might work for in-house systems but couldn’t be useful to a hyperscaler that has to meet all kinds of workloads,” he said.
Add 32GB is not a lot of room. A relatively small model with 7 billion parameters running on FP16 requires 80 gigabytes of memory, for example.
Nvidia did not respond to a request for comment.
