NVIDIA’s new RTX PRO 5500 has the same GPU chip as the GeForce RTX 5090, 84GB of memory, and a price tag that has surfaced at $17,999. On paper, that sounds like the world’s most expensive memory upgrade. In practice, it is a useful look at what NVIDIA thinks its best gaming silicon is really worth, and why the answer is no longer a gaming card.
The RTX PRO 5500 Blackwell Workstation Edition appeared on NVIDIA’s product pages in September 2026 and is still listed as “coming soon.” NVIDIA has not announced an official price. The $17,999 figure comes from a US government reseller listing, so treat it as a signal rather than a final sticker. Even so, the gap between that number and the RTX 5090’s $1,999 launch price is the story.
RTX PRO 5500 at a glance
- GPU: GB202, with 170 streaming multiprocessors and 21,760 CUDA cores, the same core count as the RTX 5090.
- Memory: 84GB of GDDR7 with ECC, running at 1,398 GB/s over a 448-bit bus.
- AI hardware: 680 fifth-generation Tensor cores and 170 fourth-generation RT cores, with FP4 support.
- Power: up to 600W, one 16-pin CEM5 connector, dual-slot cooling.
- Workstation features: two isolated 42GB MIG instances, four DisplayPort 2.1b outputs, PCIe 5.0 x16, and air- or liquid-cooled rack options.
- Price: $17,999 in a US reseller listing; NVIDIA’s official price and ship date are still unconfirmed.
Same engine, different truck
The easiest way to understand the RTX PRO 5500 is to stop thinking of it as a graphics card. Think of the GB202 chip as an engine. NVIDIA puts that engine in a gaming truck with 32GB of fast memory, a 512-bit memory bus, and a 575W power limit. The RTX PRO 5500 puts the same engine in a work truck with 84GB of slower memory, a 448-bit bus, a 600W power limit, and drivers built for professional software.
The core count is identical: 21,760 CUDA cores, 680 Tensor cores, and 170 RT cores. That is why the “rebadge” accusation makes sense at first. But the parts around the chip are different. The 5500 uses 28 memory chips, 14 on each side of the board, in a clamshell layout. That is how NVIDIA reaches 84GB without making the board enormous. The trade is bandwidth: the 5090 moves data at 1,792 GB/s, while the 5500 drops to 1,398 GB/s, about 22% less.
NVIDIA also cut the memory speed from 28 Gbps to 25 Gbps and reduced the bus from 512 bits to 448 bits. Some early listings said 416-bit; NVIDIA’s current spec sheet and PNY’s product page say 448-bit. Either way, the point is the same: this card is built for capacity, not for raw memory speed.
So no, NVIDIA did not pull a 5090 off the shelf, unplug its memory, and print a new logo. It built a different board around the same silicon. The distinction matters because the silicon is the scarce part.
Why 84GB matters more than another 10% of speed
For a normal desktop user, 32GB of video memory is already enormous. For someone running a large AI model locally, it can be a hard wall. A model has to fit in memory before it can run. If it does not fit, the software either refuses to load it or offloads part of it to the system, which is much slower. More memory lets you run a bigger model, use a longer context window, or keep more of a project loaded at once.
That is the real pitch for the RTX PRO 5500. It has 2.6 times the memory of the RTX 5090. For AI teams that want to keep sensitive data on their own hardware, or for studios that cannot send huge scenes to a cloud service, that extra room can matter more than a few percent of clock speed. ECC memory also helps in long-running professional workloads by catching certain memory errors instead of letting them corrupt a result.
Multi-Instance GPU support is the other practical feature. The 5500 can be split into two isolated 42GB instances, each with its own memory, cache, and compute. A small lab can give one GPU to two users without them stepping on each other. A rack-mounted workstation can serve more people from the same card. That is a feature the RTX 5090 does not offer.
The catch is bandwidth. The 5500 has less of it than the 5090, so it will not win every task. Workloads that are limited by memory capacity, large models, big datasets, complex scenes, should benefit. Workloads that are limited by memory speed may not. Anyone buying this card should look at benchmarks for their own software, not just the spec sheet.
The price is the awkward part
The RTX 5090 launched at $1,999. The RTX PRO 5500 listing is $17,999. That is nine times the price for the same core count and 52GB more memory. Even allowing for ECC memory, a professional board, MIG, and workstation drivers, the gap is hard to explain on hardware alone.
Supply and demand explain more of it. High-bandwidth memory is expensive and in short supply, and AI demand has made every large-memory GPU more valuable. NVIDIA also earns more per chip when that chip goes into a $17,999 workstation card than when it goes into a $1,999 gaming card. Reuters reported in October that ByteDance, Alibaba, and Tencent had ordered more than three million RTX PRO 5500s at roughly $11,400 each. If those numbers are right, the card was already a massive business before it reached US retail listings. That demand also runs into export controls: Washington decides which advanced chips can be sold to Chinese buyers, and the rules have tightened repeatedly since 2022. Our explainer on how AI export controls work covers the current framework.
There is a strange comparison inside NVIDIA’s own lineup. The 96GB RTX PRO 6000 launched at $8,565, and NVIDIA’s marketplace now lists it at $16,000. The 5500, with fewer cores and less memory, is listed at $17,999, more than the card above it. Launch prices and street prices are different things, especially in a memory shortage, but the inversion shows how quickly the market has moved. Buyers who need this class of card should compare the 5500 with the 6000, the 5000, and cloud rental rates before assuming the newer name is the better deal.
What about the RTX 5090?
NVIDIA insists the RTX 5090 has not been discontinued. In October, NVIDIA executive Jeff Fisher said the card is supply-constrained, not canceled. At the same time, reports have suggested that GB202 supply is shifting toward professional cards, and the 5500 uses the exact 170-SM configuration that previously had one main home: the 5090.
That does not prove the 5090 is going away. It does show that NVIDIA has a choice about where to send each chip, and the professional card pays much better. For gamers, the practical result is the same either way: supply stays tight and prices stay above the launch number. For AI buyers, the 5500 is a way to get the same compute with much more memory, if the price is acceptable.
Who should care, and who should not
AI teams running local models: worth a look if 32GB is the limit and the work cannot move to a cloud provider. The 84GB capacity and two 42GB instances are the main draw.
3D, rendering, and simulation teams: useful for large scenes, ECC protection, and rack-mounted deployments. The professional drivers and display outputs are built for this audience.
Labs and multi-user workstations: the MIG split is genuinely useful. One card can serve two users with isolated memory and compute.
Gamers: this is not your card. The RTX 5090 has more memory bandwidth, a lower launch price, and gaming-focused drivers. The 5500 is slower on memory-heavy games and costs far more.
Small studios on a budget: do the math. If 32GB is enough, an RTX 5090 at or near its launch price is far cheaper per unit of work. If it is not enough, compare the 5500 with cloud GPU time and with NVIDIA’s own RTX PRO 6000 before committing.
What to watch before buying
- Official price and availability: NVIDIA has not confirmed either. Reseller listings can move quickly in a shortage.
- The final memory spec: NVIDIA calls the specs preliminary. Confirm the bus width and bandwidth on the exact model you order.
- Independent benchmarks: the 5500 will not be faster than a 5090 in every task. Check the software you actually use.
- Power and cooling: 600W is a lot. Make sure the workstation, rack, and power delivery can handle it.
- Driver support: professional cards live or die on certified drivers. Confirm your applications are supported before you buy.
Frequently asked questions
Is the RTX PRO 5500 just an RTX 5090 with more memory?
No. It uses the same GB202 chip and the same 21,760 CUDA cores, but it has a different board, 84GB of ECC memory, a narrower and slower memory bus, a higher power limit, professional drivers, and MIG support. It is the same silicon in a different product.
Is the RTX PRO 5500 faster than the RTX 5090?
Not in raw compute. The core count is the same, and the 5500 has less memory bandwidth. It can be faster in workloads that need more than 32GB of memory, because the 5090 may have to slow down or fail to load the task at all. In tasks that fit in 32GB, the 5090 can be just as fast or faster.
Can the RTX PRO 5500 run large AI models?
It can fit larger models than a 32GB card, and it can split into two 42GB instances. Whether a specific model fits depends on its size, precision, and context length. Check the memory requirement for your model and quantization before assuming 84GB is enough.
Why does the RTX PRO 5500 cost $17,999?
NVIDIA has not announced an official price. The figure comes from a reseller listing. It reflects the cost of 84GB of ECC memory, the professional board and drivers, strong AI demand, and limited supply of high-end chips and memory. It is not a simple hardware-cost calculation.
The bottom line
The RTX PRO 5500 is not a scam, and it is not a simple rebadge. It is a specialist tool built around the same chip as the RTX 5090, with the memory and features that professional buyers will pay for. The $17,999 listing is a reminder that, in 2026, the most valuable part of a high-end GPU is not the core count. It is the memory, and the ability to decide who gets it.
For most people, the advice is simple: if you are gaming or working comfortably within 32GB, the RTX 5090 remains the better buy. If your work stops at 32GB, the RTX PRO 5500 is one of the few ways to get 84GB in a single NVIDIA workstation card. Just confirm the official price, the final specs, and the benchmarks before you spend gaming-card money nine times over.
Have a tip? Email [email protected] or use the contact form and choose Editorial. We do not promise anonymity.
Read next: SlipStream: making high-bandwidth flash work for AI chips; Jensen Huang petition tops 180,000 as Musk backs Nvidia return; Claude picked a bombed Iranian school as a target in tests. Browse all Chips coverage.
Sources: NVIDIA; Tom’s Hardware; TechRadar; Reuters.
