RTX 5090 sold out, RTX Pro 5500 with 84 GB: recalculating the purchase

In late August, we measured where our inference lane actually runs: not on the CUDA cores, but on the memory bus (own measurement). Three weeks later, the purchase decision is back on the table, and it can be answered without opinion. The RTX 5090 has disappeared from US online retail, third-party sellers are asking up to 9,500 USD. NVIDIA places the RTX Pro 5500, a workstation card, on the same GB202 chip, with 84 GB instead of 32 GB. Anyone buying now is not comparing two prices. Anyone buying now is comparing two bottlenecks.

What the market is showing right now

The figures, as of mid-September 2026, all from reporting, not from your own cart:

• RTX 5090: No longer available online. Third-party sellers are asking 6,395 to 9,500 USD. The median in June was 4,299 USD, the lowest printed price in early September was 5,199 USD, and a local Micro Center was still around 4,299 USD.
• RTX Pro 5500 Blackwell: same GB202, 170 SM, 21,760 CUDA cores, but 84 GB GDDR7 on a 448-bit bus with 1,398 GB/s, 600 W. The 5090, on the other hand: 32 GB and 1,790 GB/s.
• No official price for the Pro 5500. It sits between the RTX Pro 6000 (96 GB, 15,599 USD) and the RTX Pro 5000 (72 GB, 9,209 USD).

These are US retail prices, observed indirectly. Austrian list or street prices are not verified at this time, and we have not tested a supplier in Tyrol. Anyone using these figures for procurement in the region must inquire about the AT price themselves.

Which resource runs out first

The question about the right card is not "how much VRAM", but: what runs out first in operation. Two work profiles, two answers.

Dense model, decoding. A 27B model keeps all weights in VRAM, and every token pulls them once through the bus. That is exactly why our DRAM measurement was so revealing: KV-cache traffic behaves like streaming HBM access, where the limit is bandwidth, not compute (measurement, operation). On this profile, 1,790 vs 1,398 GB/s is a 22 percent difference that no amount of capacity can fix. The Pro 5500 has less bandwidth than the 5090 on this profile, with the same chip, same core count, and two to three times the median price.

MoE, expert offload, long context. Once a model has to offload parts of its experts or context, memory size is the wall. 32 GB there is not a slower state, but a different operation: offload over the host memory bus costs two bus passes per token, and beyond a certain model size nothing works at all. 84 GB on this profile is a different machine.

The note we kept for ourselves: bandwidth determines how fast a token is generated. Capacity determines whether the model fits on the hardware at all. You cannot buy both in this price range.

The purchase invoice

Our comparison figure is measured, not calculated: one month of agent operation, 18.3 billion tokens, €261 on our own infrastructure vs. €7,640 at API list price (recalculation). That is a difference of €7,379 per month. We compare this against four cards because our lane runs with tensor parallelism across four cards.

Table footnote, calculation assumption in one sentence: simple straight-line depreciation over N months, hardware only, excluding power, rack, and personnel, US list prices in USD versus savings in Euro, no currency conversion applied. This is not a price quote or an offer, these are indirectly observed US prices, and an Austrian end-user price is not verified.

Two things stand out in this table. First: at 9,500 USD, you pay more for 32 GB than the RTX Pro 5000 with 72 GB asks for. The scarcity price of the consumer card is not a quality premium but a panic premium, and you do not get it back when the shelves fill up again. Second: the €261 already includes the depreciation of the current cards. Anyone who factors new cards into the same savings defers a second depreciation onto the same monthly profit. The table is therefore a pain threshold, not an amortization calculation. It only says: anything costing more than five times a monthly profit is too expensive, regardless of the logo on it.

Capacity without kernels

The catch nobody mentions in the datasheet. NVIDIA sells FP4 before the matching kernels are available: Qwen3.8-27B-NVFP4 appeared on September 4, 2026 under Apache 2.0, DeepSeek-V4.1-Flash-NVFP4 on September 16, 2026. One week after vLLM v0.29, the tracker lists exactly what breaks on the Blackwell path in production:

• Issue 57156, September 16: "DeepSeek-V4.1 produces NaN with CUDA graphs on SM120/121"
• Issue 57292, September 17: sparse-MLA with 64-token pages on SM120
• Issue 57209: FlashInfer bmm_fp8 on sm_12x is hard-bound to cuBLAS

The checkpoints exist, and the code path for the card is the one currently failing. Buying FP4 capacity today buys a plan, not capacity: the 84 GB are real, but the path to fill them at the intended precision is not fully working this week.

This is verifiable, and before ordering, not after. We run the test like this: target model, target precision, target KV cache type, CUDA Graphs enabled, long context, same engine version as in production later, using a workload profile with multiple concurrent sequences. What runs cleanly is buyable. What runs in the vendor's demo is the vendor's problem, not the project's. And what only runs with features disabled is a roadmap you should not pay for with your own schedule.

Perspective

We sell no hardware and earn nothing on any card, not even on a recommendation. Our own lane stays with the 5090 as long as bandwidth is the bottleneck and 32 GB suffices for the dense model. The moment we serve MoE with long context in production, we buy capacity. But then the choice is not 5090 vs Pro 5500, but 84 GB vs CPU offload, i.e. the question of whether a card with large memory is the cheapest answer or a host with many channels and a clean offload path (calculation here). We make this decision with a measurement log and not with a launch article, and only when the workload profile exists.

Further Reading