TL;DR: An RTX PRO 6000 Blackwell now costs $13,998–$16,000 against roughly $9,050 for two RTX 5090s — the price parity this matchup was famous for died in August. The ~$5,000 premium buys 32GB more VRAM in one slot, gpt-oss-120b at 196 tok/s with full 131K context, and half the wall power. The dual-5090 rig buys two independent fast cards and a five-figure invoice avoided.

	RTX PRO 6000 (600W)	RTX PRO 6000 Max-Q (300W)	2× RTX 5090
Best for	64–96GB models, batch serving	Same, on a normal PSU	Parallel ≤32GB workloads
Price (Sep 2026)	$13,998–$16,000	~$12,700–$16,400	~$8,400–$10,400
VRAM	96GB, one pool	96GB, one pool	64GB, split 32+32
Power	600W	300W	1,150W combined
The catch	Costs ~1.5× the pair	~23% slower at high batch	70B decode no faster than one big card

Honest take: If the models you actually want live between 64GB and 96GB — gpt-oss-120b resident, 70B at Q8 — pay for the PRO 6000 and skip the multi-GPU tax entirely. Buy two 5090s only if your workload is genuinely two workloads.

When we reviewed the RTX PRO 6000 Blackwell in the spring, the framing wrote itself: the 96GB workstation card cost about $8,500, two RTX 5090s cost about $8,600, same money, pick your architecture. That article aged fine. The prices didn’t.

As of September 2026 the matchup is lopsided in a new way, and the question isn’t “which architecture at the same price” anymore. It’s whether one big card is worth five thousand dollars more than two fast ones. The answer depends almost entirely on one number: how many gigabytes the model you care about actually needs. Check yours against the VRAM calculator before reading further — it will tell you which half of this article is about you.

What happened to the prices

NVIDIA doubled the RTX PRO 6000 Blackwell’s MSRP to $16,000 on August 13, 2026 — a card that opened pre-orders below $8,000 in early 2025 and launched at an $8,565 MSRP. The driver is the same GDDR7 shortage that has frozen NVIDIA’s consumer roadmap: the PRO 6000 carries 96GB of GDDR7 in a clamshell layout, more of the scarce commodity than any other card NVIDIA sells, so it took the steepest repricing in the lineup.

Street prices, per Thunder Compute’s September 2026 tracking: $13,998 at Newegg, $15,499 at B&H, $19,999 at Amazon as of September 12. The 300W Max-Q variant — identical GB202 silicon, identical 96GB, half the power limit — tracked between $12,712 and $16,446 through the summer, and NVIDIA lists both variants at the same $16,000 now. The $8,299 Max-Q deals from late 2025 are gone and not coming back at current memory prices.

The RTX 5090 took its own beating — we covered the $300 board-cost hike in May — but from a lower base. Pangoly’s tracker averaged $4,525 per card on September 13, 2026; Tom’s Hardware’s best-available US price was $5,199, with Micro Center stock near $4,200. Two cards land between $8,400 (patient, Micro Center) and $10,400 (impatient, premium AIB models); Amazon listings run $5,069–$5,799 each. Call the realistic pair ~$9,050.

So the gap between the dual-5090 build and the cheapest PRO 6000 is roughly $4,900 — before you add the second PSU the 5090 pair probably needs.

The spec sheet is a trap

Here’s the number that makes this comparison interesting: 1.79 TB/s. That’s the memory bandwidth of the RTX PRO 6000. It is also, within rounding, the memory bandwidth of a single RTX 5090 (1,792 GB/s). Same 512-bit bus, same GDDR7 generation, nearly the same GB202 die.

Single-stream token generation — one user, one prompt, tokens streaming out — is bandwidth-bound. The GPU spends its time reading weights, not doing math. Which means for any model that fits on both setups, the PRO 6000 and a 5090 generate tokens at roughly the same speed, and — this is the part multi-GPU newcomers get wrong every time — the second 5090 doesn’t make generation faster. Splitting a model across two cards doubles capacity and parallelizes prompt processing; the decode stream still walks through the weights one token at a time. We measured this directly on dual R9700s, where adding a second card took dense-model decode from 24.85 to 24.31 tok/s — slightly backwards — while prefill scaled almost perfectly. Same physics on NVIDIA.

The benchmark record bears it out at 70B scale. Databasemart’s dual RTX 5090 Ollama benchmark (Ollama 0.6.5) measured ~27 tok/s on Llama 3.3 70B Q4 split across two 5090s — genuinely H100-class, the fastest 64GB you can assemble from consumer parts. A single PRO 6000 streams the same model at 30–45 tok/s under vLLM, with the whole 42.5GB of Q4_K_M weights resident on one card. Different engines, so treat the exact figures gently — but the direction is clear: on the marquee dense-70B workload, the single big card is at least as fast as the pair, with none of the coordination overhead.

Where two 5090s genuinely pull ahead is compute: two full-power GB202s against one. Prompt processing on long documents, batched image generation in ComfyUI, two separate models served at once — anything that parallelizes across cards gets close to 2× the throughput. A single 5090 already posts 205 tok/s decode and 5,988 tok/s prefill on Qwen3.5-35B MoE; two of them doing independent jobs is a small render farm.

What 96GB buys that 64GB cannot

The honest version of this section is one sentence: it buys the 64–96GB model window, and that window currently contains the best models a home lab can run.

gpt-oss-120b is the tier’s flagship tenant. The llama.cpp guide for gpt-oss puts it at 196 tok/s generation and 4,503 tok/s prompt processing on an RTX PRO 6000, with the entire 131,072-token context fitting in 68.5GB. On 64GB, the same model at 32K context needs 64.9GB — a sub-gigabyte miss that forces MoE expert layers into system RAM. The --n-cpu-moe flag makes the spill graceful, but you’re managing a workaround where the PRO 6000 owner is not, and you gave up the other 99K tokens of context.

The failure mode looks like this on the dual-5090 rig:

$ llama-server -m gpt-oss-120b-mxfp4.gguf -c 32768 -ngl 99 --tensor-split 1,1
ggml_backend_cuda_buffer_type_alloc_buffer: allocating 32768.00 MiB on device 0: cudaMalloc failed: out of memory
llama_model_load: error loading model: unable to allocate CUDA0 buffer

The fix is spilling a few expert layers to the CPU instead of failing the load:

$ llama-server -m gpt-oss-120b-mxfp4.gguf -c 32768 -ngl 99 --tensor-split 1,1 --n-cpu-moe 3

That works — a 64GB rig only has to offload 2–4 layers, and because just 5.1B parameters activate per token, generation stays in the high 20s tok/s. But it’s a permanent asterisk on every long-context session, and prompt processing pays the PCIe toll. The PRO 6000 loads the same model with -ngl 99 and 27GB to spare.

Dense 70B at real quantizations is the second tenant. 64GB runs Llama 3.3 70B at Q5_K_M (49.9GB). 96GB runs it at Q8_0 (~75GB) — the quantization where quality loss stops being a debate — with cache room left over.

Batch throughput is the third, and the least appreciated. With big batches, capacity converts directly into throughput: a single PRO 6000 pushed 8,425 tok/s aggregate on Llama 3.3 70B AWQ versus 4,570 tok/s for a single 5090 — the extra VRAM holds the KV cache for far more concurrent requests. If you’re serving a coding agent fleet (the workload our sister site lives on), the one-card-per-box math changes completely.

The multi-GPU tax, itemized

Two 5090s are not one 64GB GPU. They’re two 32GB GPUs and a set of obligations:

• Topology. Consumer GeForce cards have no NVLink and NVIDIA disables PCIe peer-to-peer on them; every cross-card transfer bounces through host memory. Tensor parallelism works in vLLM and llama.cpp, but the all-reduce traffic rides PCIe — our NVLink vs PCIe explainer covers what that costs, and the dual-3090 IOMMU/ACS guide covers the BIOS archaeology that multi-GPU boxes still occasionally demand. The PRO 6000 owner’s topology is a single x16 slot.
• Power. Two 5090s are rated 575W each — 1,150W of GPU before the platform. At the EIA’s 18.83¢/kWh April 2026 residential average, that’s about $0.22/hour under sustained load, versus $0.11 for the 600W PRO 6000 and roughly $0.06 for the Max-Q. Run heavy inference 6 hours a day and the dual rig costs ~$240/year more than the Max-Q in electricity — real money, though nobody’s closing a $5,000 gap with it. The larger practical issue is the PSU: dual-5090 builds want 1,600W+ and clean transient behavior, which is its own $400–$600 line item.
• Space and heat. Two triple-slot cards want a full tower and airflow planning. The Max-Q is a two-slot blower designed to live politely in a workstation — it’s the variant NVIDIA builds specifically so four of them can share a chassis.

The Max-Q deserves one honest caveat: 300W is half the power for the same silicon, and at high concurrency it shows. Community power-limit testing collected in the Level1Techs 600W-vs-300W thread and the Pulsed Media wiki found single-user inference loses only ~4% at 300W — decode is bandwidth-bound and rarely pulls full TDP anyway — but 64-concurrent-user batch throughput dropped from 1,558 to 1,206 tok/s, about 23%. Solo user: buy whichever variant is cheaper that day. Batch server: the 600W card earns its cooler.

What to actually buy

Prices as of September 2026, all verified in the comparison above:

Your situation	The machine	Price	Where
Your models live in the 64–96GB window (gpt-oss-120b, 70B Q8)	RTX PRO 6000 Blackwell 96GB	$13,998–$16,000	Check price
Same, but on a normal PSU / multi-card ambitions	RTX PRO 6000 Max-Q 300W	~$12,700–$16,400	Check price
Two genuinely parallel ≤32GB workloads	2× RTX 5090	~$9,050/pair	Check price
Everything you run fits in 24GB	Used RTX 3090	~$1,050	Check price
Not sure the workload justifies five figures	Rent a PRO 6000 first, ~$1.99/hr	pay per hour	RunPod

The last two rows are load-bearing. If your models fit in 24GB, a used RTX 3090 delivers most of this experience for a tenth of the money. And at these prices, renting before buying isn’t a hedge, it’s due diligence: RunPod rents the RTX PRO 6000 at about $1.99/hour secure-cloud — 100 hours of testing your actual workload costs $200 against a $14,000 mistake. The open-source serving stacks you’d run at home (vLLM, llama.cpp, SGLang) are identical in the cloud, so the numbers transfer.

The verdict

The spring version of this article would have been a coin flip. The September version isn’t, because the two options no longer cost the same and no longer even compete for the same buyer.

Buy the RTX PRO 6000 — and at current pricing, shop the Max-Q against the 600W card and take the cheaper one unless you batch-serve — if the point of spending five figures is running models that need more than 64GB. That’s the only thing the extra $5,000 reliably buys, and it happens to be the tier where the current best local models live. One slot, one power connector, 196 tok/s on a 120B MoE with full context, no topology homework.

Buy two RTX 5090s if you’d actually use two GPUs — one serving an LLM while the other renders ComfyUI batches, or two model endpoints at once. As a way to build “one 64GB GPU,” the pair is the weaker product: 70B decode no faster than the single card, sub-1GB short of gpt-oss-120b at 32K, double the wall power, and $9,000 is a lot to pay to almost fit the model everyone wants to run. If that’s the goal and the budget stops at $9K, the honest moves are one 5090 now (a single card carries the ≤32GB class brilliantly) or the 96GB page’s used-3090 quad math — capacity without the Blackwell premium.

And if none of these numbers feel sane: they aren’t, particularly. This is what a memory shortage looks like from the buying end. Rent until it passes.

FAQ

Is the RTX PRO 6000 really the same speed as an RTX 5090 for one user? For models that fit both, close to it — both move ~1.79 TB/s of memory bandwidth, and single-stream decode is bandwidth-bound. The PRO 6000’s advantages are capacity (96GB vs 32GB) and batch throughput (roughly 1.8× a single 5090 on 70B AWQ serving), not single-user speed.

Can dual RTX 5090s run gpt-oss-120b? Yes, with a caveat: at 32K context the model wants 64.9GB, just over the 64GB pool, so you offload 2–4 MoE layers to CPU with --n-cpu-moe. Generation stays acceptable because only 5.1B parameters activate per token. Full 131K context stays out of reach; the 96GB card holds it with room to spare.

Max-Q or 600W Workstation Edition? Same silicon, same 96GB, same bandwidth. Single-user inference loses only ~4% at 300W; heavy concurrent serving loses ~23%. Solo home lab: buy the cheaper listing that day and pocket the power savings. Batch server: the 600W card.

Why not two used RTX 4090s instead? 48GB for roughly $4,300–$4,700 used is the 48GB tier — a fine tier, but it holds a 70B at Q4, not Q8, and misses gpt-oss-120b by 17GB. It’s the value answer to a different question.

Do I need NVLink for tensor parallelism on 5090s? You can’t have it — no GeForce card since the 3090 has NVLink, and PCIe peer-to-peer is disabled on GeForce. vLLM and llama.cpp still split models over PCIe; it works, with prefill and all-reduce paying the toll. The PRO 6000 sidesteps the question entirely by not needing a second card.

Recommended Gear

• NVIDIA RTX PRO 6000 Blackwell 96GB — the one-slot answer to the 64–96GB model window
• NVIDIA RTX PRO 6000 Blackwell Max-Q — same 96GB at 300W, ~4% slower for one user
• NVIDIA RTX 5090 32GB — the fastest ≤32GB card, alone or in pairs
• Used RTX 3090 24GB — still the sane default if your models fit

Sources

• Nvidia doubles RTX PRO 6000 Blackwell’s MSRP to a staggering $16,000 — Tom’s Hardware
• NVIDIA RTX PRO 6000 Blackwell Pricing, September 2026 — Thunder Compute
• RTX PRO 6000 Blackwell Max-Q price tracking, July 2026 — GPU Poet
• GeForce RTX 5090 price trends — Pangoly
• Gaming GPU prices 2026: RTX 5090 tops $5,000 — Tech Insider
• 2× RTX 5090 Ollama benchmark: Llama 3.3 70B — Databasemart
• llama.cpp guide: running gpt-oss — ggml-org discussion #15396
• RTX Pro 6000 600W vs 300W performance — Level1Techs Forums
• NVIDIA RTX Pro 6000 (Blackwell) for LLM inference — Pulsed Media Wiki
• RTX PRO 6000 Blackwell Max-Q Workstation Edition specs — NVIDIA
• Llama 3.3 70B Instruct GGUF quantization sizes — bartowski, Hugging Face
• RunPod pricing guide 2026 — Flexprice

Last updated September 14, 2026. Prices and specs change; verify current rates before purchasing.