RTX 5090 Price Matches DGX Spark, Ending Cost Advantage
The RTX 5090 now costs roughly the same as the full DGX Spark, shifting the value balance for local AI users.
Key points
- The RTX 5090 averaged $4,639 in September 2026, matching the DGX Spark's $4,699 list price.
- The RTX 5090 is four times faster than the Spark on the GPT-OSS 20B benchmark.
- The Spark runs 120B models that the 32GB RTX 5090 cannot load.
The price gap between the NVIDIA DGX Spark and the RTX 5090 has effectively vanished. In September 2026, the average US price of the RTX 5090 hit $4,639. This matches the $4,699 list price of the complete DGX Spark system.
Memeburn reports that this shift undermines the long-standing argument for buying the graphics card. The card is no longer a cheaper alternative. Buyers must now weigh raw speed against total system cost.
Market prices erase the discount
NVIDIA launched the RTX 5090 at $1,999. The DGX Spark started at $3,999. By February, NVIDIA raised the Spark price to $4,699 due to memory shortages. Meanwhile, graphics card prices surged.
Pangoly data shows the RTX 5090 averaged $4,639 in September 2026. This price reflects GDDR7 supply issues. The card also requires a separate computer. A full build needs a 1,000W power supply, CPU, and storage.
Speed favors the graphics card
The RTX 5090 has 1,792GB/s memory bandwidth. The DGX Spark offers 273GB/s. Higher bandwidth allows faster token generation. On the Ollama benchmark with GPT-OSS 20B, the RTX 5090 was about four times faster than the Spark.
The Spark’s compute power is comparable to an RTX 5070. It does not match the 5090 in raw speed. If your model fits in 32GB, the card is the faster choice. The Spark wins only when capacity is the bottleneck.
Capacity unlocks specific AI tasks
The DGX Spark includes 128GB of shared memory. This allows it to run large models like GPT-OSS 120B. The RTX 5090 is limited to 32GB. It cannot load these larger models at all.
NVIDIA states that 70B QLoRA fine-tuning fails on 32GB GPUs. The Spark handles this task at about 760 tokens per second. Power usage also differs significantly. The Spark peaks at 240W for the whole system. The RTX 5090 card alone draws 575W.
