DGX Station Alternatives: 4 RTX PRO 6000s Lead Only on MiniMax M3

No DGX Station alternative with a public US price beats the GB300 desktop on price, measured throughput and memory at once, on the prices and independent measurements available on Nov. 4, 2026.

The closest call is four NVIDIA RTX PRO 6000 cards. In StorageReview's tests they beat the Station only on MiniMax M3, a model that barely fits its fast memory, and only from 4 to 16 concurrent streams: 1,181 output tokens per second against 940 at 16, then 759 against 1,041 at 32. The Station's figures there are best-case, because the lab forced its speculative decoder's acceptance length. Two AMD MI350P server cards led on two models at 32 streams, but in AMD's MXFP4 against NVIDIA's NVFP4, and Voltgrid found no public price for them. The baseline price varies too: Central Computers shows $111,999.99 for ASUS's build of the Station as a request-a-quote listing, ASUS publishes no list price, and Exxact's starting price is $98,450.00, with no storage stated.

The four cards hold 384 GB of GDDR7 across four boards, 51% of the Station's 748 GB, which counts 496 GB of slower CPU memory, by Voltgrid's arithmetic, and none of the US builders Voltgrid read publishes a fixed four-card price. Two 512 GB Mac Studio M5 Ultras or four 192 GB Strix Halo mini PCs cost less and add up to more memory on paper, but Voltgrid found no lab that had measured either serving many users as of Nov. 4.

The job is serving a 120B-class open model such as gpt-oss-120b, or a larger mixture-of-experts model, to many users or agents on premises. Voltgrid did not test these machines. Every speed below names who measured it and on which model, says where the lab left quantization or software unstated, and stays inside one test setup. Our DGX Station review covers the platform.

The baseline: 748 GB and 9,280 tokens per second

Every GB300 desktop is built around the same GB300 Superchip design: one Blackwell Ultra GPU with 252 GB of HBM3e at 7.1 TB/s, joined by an NVLink-C2C link that NVIDIA lists at 900 GB/s, and StorageReview labels bidirectional, to a 72-core Grace CPU with 496 GB of LPDDR5X at 396 GB/s. NVIDIA counts the two pools as 748 GB of coherent memory. It lists 1,600 W of total system power and a 20 A circuit; full-load wall power is unpublished.

US prices read on Nov. 4 start at $98,450 for Exxact's Valence DGX Station, which Tom's Hardware found at $94,930 in August. MSI's XpertStation WS300, model WS300T60L with two 2 TB drives, lists at $99,999 at Newegg as a "start from" price on a request-a-quote listing. Puget Systems' GB300 XpertStation is $106,648.80 as configured, and Dell's default Pro Precision build, with a display card and 16 TB of SSDs, is $160,195. TechRadar reported on Sept. 1 that Supermicro had begun selling its GB300 workstation at $92,887.47 before tax and shipping, or about $90,101.05 with a 3% new-customer discount; its headline says almost $91,100. Voltgrid could not find Supermicro's US store page for the model on Nov. 4. These are starting, listed, default or request-a-quote prices, and none is a confirmed delivered price.

StorageReview's vLLM serving test is the yardstick: 512-token prompts and 512-token answers, with concurrent streams doubling up to 128. On gpt-oss-120b, which ships in MXFP4 so every machine runs the same weights, the lab measured 9,280 and 9,294 aggregate output tokens per second at 128 streams on an ASUS ExpertCenter Pro ET900N G3 and an MSI WS300, or 5,023 and 5,038 with 8,192-token prompts and 1,024-token answers. One AMD Instinct MI350P at 450 W, on vLLM 0.32.0 and ROCm 10.1.0, reached 5,057, and one 600 W RTX PRO 6000 Server Edition in a Dell XE7740, on NVIDIA's vLLM 26.08 container, reached 3,064, according to StorageReview's MI350P first look. A DGX Spark-class Acer Veriton GN100 reached 500, a result the lab's ASUS review carries over from its August WS300 testing. Voltgrid found no four-card, Mac, Strix Halo or Tenstorrent result for this setup.

The reason is where the weights sit. Each new token means reading the model's active weights again, so once a model fits, decode speed is limited mostly by memory bandwidth, StorageReview writes, though at many streams compute matters too: in StorageReview's matrix-math test the Station reached about 3.3 times the MI350P's BF16 rate. StorageReview read the Station's HBM at 6,869 GB/s in its own test, against 1,540 GB/s for one RTX PRO 6000 Server Edition, a card its table rates at 1.6 TB/s. A model larger than 252 GB spills into the 396 GB/s LPDDR5X. The rivals' best result came just below that line: MiniMax M3's 233 GB checkpoint barely fits, and four RTX PRO 6000s led it at 4 to 16 streams, while on GLM-5.2, at 433 GB, the Station led the four cards.

Four RTX PRO 6000s: faster at one setting, half the memory

StorageReview's four-card machine was a Dell PowerEdge XE7740 server with 600 W RTX PRO 6000 Server Edition cards of 96 GB each. On MiniMax M3 in NVFP4, with the same EAGLE3 speculative decoder as the Station, the cards matched the Station at one and two streams and led from four streams up to 1,181 against 940 at 16. At 32 streams the Station led, 1,041 against 759. StorageReview's MI350P review adds that the Station, the four cards and two MI350Ps stayed within 10 percent of each other through eight streams, so the four cards' lead from four to eight streams is small. With 8,192-token prompts the four cards also led, about 349 against 271 at four streams and 444 against 103 at eight.

On GLM-5.2, about 433 GB in NVFP4, the order reverses. The Station keeps about 215 GB of experts in Grace memory and served 35 tokens per second to one stream and 139 across 32, best-case figures from MTP speculative decoding; the four cards' decoder setting is not stated. The four cards, each offloading about 30 GB to host memory over PCIe, managed 9 and about 40. Voltgrid found no published four-card run of gpt-oss-120b.

Price is the other gap. None of the builders Voltgrid read, Exxact, BIZON, VRLA Tech, Puget and Central Computers, lists a fixed four-card price: VRLA Tech configures quad builds to the workload, and Puget's Quad GPU Rackmount shows $63,861.73 only with two Max-Q cards. VRLA Tech says four RTX PRO 6000s draw up to 2,400 W and typically need a dedicated 30 A, 208 V circuit. Puget's Threadripper PRO tower with two 300 W Max-Q cards and 256 GB of DDR5 costs $52,412.24 as configured, though those are not the cards StorageReview measured. StorageReview's two 600 W Server Edition cards ran DeepSeek V4 Flash in FP8, with 512-token prompts and answers at 32 streams, at 803 tokens per second in its Sept. 23 ASUS review and 829 in its Nov. 3 MI350P first look, against 1,766 for the Station. Our RTX PRO 6000 review covers the card.

How Speculative Decoding Shapes These Speed Figures

Speculative decoding speeds up token generation. A cheap draft mechanism guesses several tokens, and the full model checks them in one pass and keeps the longest run that matches its own output, NVIDIA's developer blog explains. The average tokens kept per pass is the acceptance length.

StorageReview forced that length to 3 tokens on the Station's MiniMax M3 and GLM-5.2 runs, so the Station's 1,041 and 139 tokens per second are best-case; real serving depends on how often the draft is right. The lab does not say whether the four-card or MI350P runs were forced the same way, so a lead or gap against the Station holds only if the decoder setting matches.

Two AMD MI350Ps: faster in a different format, and unpriced

StorageReview's Nov. 3 first look ran two 144 GB MI350P cards in a Dell PowerEdge R7725 server at 450 W. At 32 streams the pair reached 1,509 tokens per second on MiniMax M3 and 1,462 on GLM-5.3 Flash, against 1,041 and 828 for the Station. Those are not like-for-like: AMD's builds were MXFP4 and NVIDIA's NVFP4, the pair ran ROCm 7.2.3 with older vLLM images while the NVIDIA runs used the vLLM 26.08 container, the Station's MiniMax M3 figure used a forced acceptance length, and StorageReview does not state the pair's decoder setting. At one stream on GLM-5.3 Flash the Station led, 353 to 45. The pair holds 288 GB, two cards of 144 GB. Voltgrid found no public price: AMD's product page lists none, and Dell supports the card in its PowerEdge servers. AMD's deskside Threadripper Halo Station, with up to four MI350Ps, 576 GB of HBM3e and up to 2 TB of CPU RDIMM, 2.6 TB in all by AMD's count, is listed as coming in 2027.

Mac and Strix Halo clusters: more memory for less, speed unmeasured

Apple's Mac Studio M5 Ultra with 512 GB of unified memory at 1.2 TB/s costs at least $22,299, MacRumors reported on Oct. 28, and first orders arrive Dec. 3, according to Digital Trends. Two hold 1,024 GB for $44,598 by Voltgrid's arithmetic, under half Exxact's price. MacRumors reported that Meta's Muse Spark Open model runs on two such Macs with Apple's MLX framework, without a speed. The only independent M5 Ultra figure Voltgrid found is BGR's: its 256 GB unit "generally got closer to 55 tokens per second" on a 4-bit MLX build of Qwen 3.8 27B in LM Studio, an informal single-machine figure with no stated concurrency. Our Mac Studio review and our 512 GB report have the details.

Boxes on AMD's Ryzen AI Max+ PRO 495 cost less still. Framework's 192 GB Desktop is $6,799 without storage, with Batch 3 shipping in December after two sold-out batches. GMKtec's EVO-X5 Pro is $6,799 with 2 TB and in stock. Four of either hold 768 GB for $27,196, four times $6,799 by Voltgrid's arithmetic, though only GMKtec's price includes storage. Each box has 273 GB/s of memory bandwidth, "up to" that figure in GMKtec's wording. Voltgrid found no published multi-user result for a 495 box or a cluster of them. The nearest measured relative, AMD's 128 GB Ryzen AI Halo, which uses the Ryzen AI Max+ 395, served gpt-oss-120b at 222 aggregate tokens per second at batch 64, with 256-token prompts and answers and precision unstated, against 701 for a DGX Spark in StorageReview's July review.

Memory added across boxes is not one pool: a model split across machines sends its work over network links. StorageReview wrote that, in results others have published, per-user throughput on clusters beyond two Sparks can drop into single-digit tokens per second for models large enough to need them, though it did not measure those clusters itself. Until someone does, the Mac and Strix Halo clusters are bigger only on paper.

GB10 pairs: cheaper and measured, but a different class

Two GB10 boxes joined by one 200GbE link are the cheapest priced option with a multi-user measurement Voltgrid found. In StorageReview's May 11 cluster test, gpt-oss-120b in pipeline-parallel mode peaked at 463.97 output tokens per second on two Dell units, 497.26 on two GIGABYTE units and 504.88 on two HP units, at batch 64; the lab did not state precision or prompt lengths. That harness differs from the Station sweep, so its numbers do not sit beside 9,280. By Voltgrid's arithmetic, a GIGABYTE pair at $7,312 a unit (CDW, Nov. 2, 4 TB) costs $14,624, the lowest of the three, an HP ZGX Nano pair at $7,399 (HP Store via Notebookcheck, Oct. 5, 4 TB) $14,798, and a Dell Pro Precision GB10 pair at $7,967 (Dell.com, Nov. 2, 2 TB) $15,934, each with 256 GB. Notebookcheck also lists the HP at $4,759 on Amazon. StorageReview called a two-box cluster a development and learning platform. Our GB10 comparison and DGX Spark review cover the boxes.

What the money buys

Per gigabyte, the cheaper machines win, by Voltgrid's arithmetic. Exxact's $98,450 over 748 GB, 252 GB of HBM3e at 7.1 TB/s plus 496 GB at 396 GB/s, is $131.62 per GB. A $22,299 Mac Studio with 512 GB of unified memory at 1.2 TB/s is $43.55 per GB, and a $6,799 Strix Halo box with 192 GB at 273 GB/s is $35.41. Those are single machines, not clusters, and storage differs: Exxact states none, Framework includes none and GMKtec includes 2 TB.

Measured speed gives two ratios from two different tests. MSI's WS300 at Newegg's $99,999 costs $10.76 per output token per second of its 9,294 on gpt-oss-120b, in StorageReview's 512/512 vLLM test at 128 streams. The HP pair at $14,798 costs $29.31 per token per second of its 504.88, in the May 11 cluster test at batch 64. Both are Voltgrid's arithmetic, and because the tests differ, they are not a ranking.

Renting a GB300 Instead of Buying

Oracle lists a GB300 GPU at $18.00 an hour, a B300 at $15.00 and an RTX PRO 6000 at $4.50, its list on-demand rates as read by Voltgrid's GPU cloud price tracker at 02:00 UTC on Nov. 4. By Voltgrid's arithmetic on those list prices, with no matched throughput, Exxact's $98,450 divided by $18.00 is 5,469 rented hours, about 7.5 months around the clock at roughly 730 hours a month, and Dell's $160,195 is 8,900 hours, about 12.2 months. Four rented RTX PRO 6000s cost $18.00 an hour, four times $4.50, the same as one GB300. The tracker shows on-demand list prices only; committed-use, reserved and spot rates are not shown. Oracle's GB300 is the data-center part, not the desktop chip, the sum leaves out the owned machine's power, space and staff, and a rented GPU is off premises, which rules it out for this job.

Which DGX Station Alternative to Buy

A team serving gpt-oss-120b, or any model that sits well inside 252 GB, to dozens of users has the best-measured option in a GB300 desktop; MiniMax M3, at 233 GB, is under that line, and four cards led it at 4 to 16 streams. StorageReview found the ASUS and MSI builds within 2% of each other on 24 of 28 tests; that says nothing about price, and the lab did not measure Exxact's build. Exxact's $98,450 was the lowest confirmed starting price Voltgrid found on Nov. 4; Supermicro's reported $92,887.47 is lower but unconfirmed.

A team whose model sits close to 252 GB, with MiniMax M3's 233 GB checkpoint as the measured case, that serves 4 to 16 streams and already has rack space and 208 V power can ask builders for a four-card quote. It gets half the memory, and the Station led once a model needed about 215 GB of Grace memory, as GLM-5.2 did.

A developer or small team that needs capacity more than many-user speed can buy a 512 GB Mac Studio for $22,299 or a 192 GB Strix Halo box for $6,799, knowing Voltgrid found no multi-user measurement of either. The first 512 GB Mac measurements should follow the Dec. 3 deliveries.

A team that needs Station-class speed only part of the time can rent a GB300 at $18.00 an hour for up to about 5,469 hours before buying wins, if its data may leave the building. GB10 pairs remain learning machines. Next to watch are StorageReview's 600 W MI350P rerun and DGX Station for Windows, which NVIDIA expects from partners in December.