Nvidia DGX Station GB300 Lists Near $100,000 at OEM Retailers

The price Nvidia wouldn't give is now public

Nvidia never published a price for its GB300-based DGX Station, and its partners have now done it for the company. In August, Exxact posted its Valence VWS-158270643 build online with a starting price of $94,930. That roughly matched the six-figure estimates that had been going around since the machine's Computex 2026 showing.21 According to the reporting, Nvidia representatives at Computex repeatedly avoided saying what the tower would cost. The listing settled the question with a figure close to $100,000.21

Within weeks, other vendors had posted prices of their own. Supermicro began selling its Gold Series ARS-511GD-NB-LCC-01-G2 workstation for $92,887.47 before tax and shipping. New customers could get a 3% discount that brought it to about $90,101. A late-September survey of listings found Supermicro's eStore version starting at $91,812.55, Exxact's configurator at a $96,250 base, and MSI's XpertStation WS300 at $99,999 on Newegg, where it was out of stock.25 Exxact's own product page also shows the $96,250 starting figure.27

The picture is simple. Nvidia is putting rack-class Blackwell Ultra silicon in a box that fits under a desk, and the market has settled on roughly $90,000 to $100,000 for the base model. Fully configured systems cost well over that.

What the money buys

The system is built around the GB300 Grace Blackwell Ultra Desktop Superchip. That pairs a Blackwell Ultra GPU with a 72-core Arm Neoverse V2 Grace CPU over a 900 GB/s NVLink-C2C link.12 Nvidia's current spec sheet lists 252GB of HBM3e at 7.1 TB/s on the GPU and 496GB of LPDDR5X at 396 GB/s on the CPU. Both draw from one 748GB coherent memory pool.12 The company claims up to 20 petaFLOPS of FP4 compute, measured with sparsity, and support for models of up to 1 trillion parameters.17

The memory figure isn't the same everywhere. Earlier partner and press material quoted up to 784GB of coherent memory and up to 288GB of HBM3e at 8 TB/s.1315 Those numbers came from the original announcement. The shipping specification Nvidia now publishes says 252GB and 7.1 TB/s.12 One reference summary notes the conflict directly and suggests it may come from configuration variants.11 The safest reading is that 748GB is what buyers will actually get.

The rest of the machine shows it was designed as data-center hardware. It has a ConnectX-8 SuperNIC rated up to 800 Gb/s, two 400GbE QSFP112 ports, four M.2 Gen 5 slots, and three PCIe Gen 5 slots.17 The whole system draws 1,600W and is air-cooled for office use.18 The Blackwell Ultra GPU has no display output, so buyers add a separate graphics card. Exxact's top manual configuration, at $108,350, includes an RTX Pro 6000 Blackwell with its own 96GB of VRAM.21 The GPU can also be split into up to seven Multi-Instance GPU partitions, which lets several users or jobs share it.19

A long road to a price tag

The timeline in the coverage is messier than the "just launched" headlines suggest. The DGX Station was first announced at GTC 2025. A year later, in March 2026, OEMs including ASUS, Dell, Gigabyte, MSI, and Supermicro began taking orders, while Nvidia still had no official price and no Founders Edition.26 Nvidia formally introduced the GB300 version at GTC 2026.18 At Computex 2026, it added a DGX Station for Windows, developed with Microsoft and due from major manufacturers in Q4 2026.16

That explains how a figure near $97,000 was already circulating in March, tied to an MSI listing at CDW.26 It also explains why one analysis says MSI's published price was $85,000.24 MSI's figures have ranged from $85,000 to $99,999 across retailers and dates. The plausible reading is that pricing moved during a year of tight memory supply, and that no partner has a stable list price to stick to.

Some specialty integrators go much higher. Bizon lists its GB300 AI Station from $122,159 and says its first three batches sold out.23 A managed-services firm advertises DGX Station configurations "from $164,000" and going past $500,000.28 Those figures likely include services, compliance hardening, or extra hardware, and buyers shouldn't treat them as the price of the box itself.

The inference-cost argument

Nvidia's pitch for a $95,000 workstation rests mostly on comparing it with cloud GPU rental, and here the coverage splits widely depending on who is doing the math.

The most cautious analysis comes from a cloud GPU provider, which has a stake in the answer. It takes AWS's published P6-B300 Capacity Block rate of $14.04 per accelerator-hour and an $85,000 purchase price. On that basis, the station pays for itself in about 252 days of round-the-clock use, or about 2.9 years at business hours only.24 Another analysis uses the cheapest Supermicro listing and a $6.69 B200 hourly rate. It gets roughly 13,700 GPU-hours, or about nineteen months of nonstop use, before ownership wins. It adds that the desktop part will be a generation old by the time memory prices ease.25

The most optimistic view comes from a reseller selling the hardware. It compares the station against an eight-GPU H100 AWS instance at about $98 an hour and claims payback in about three months.28 That comparison sets a single-GPU workstation against an eight-GPU cloud node, which flatters the hardware. Readers should treat it as a sales argument.

The more defensible conclusion is that the DGX Station makes financial sense only for teams that would keep it busy most of the time. It also helps when there are reasons beyond cost, such as keeping data on-site or avoiding a queue for shared cluster time. Tom's Hardware put it the same way: for enterprises that want fast iteration and tight security, the price can be justified, especially once the smaller DGX Spark runs out of headroom.21 Power is a modest extra cost. The full 1,600W rating has to be provisioned at the outlet, and the electricity bill comes on top of the purchase price.24

For scale, analysts estimate a full GB300 NVL72 rack with 72 GPUs at roughly $3.7 million to $6.5 million.24 A single B300 GPU sells for about $53,000 through channel listings, and an eight-GPU DGX B300 system runs around $400,000 to $500,000.29 Seen that way, a complete single-GPU workstation with a Grace CPU and nearly 500GB of extra memory for under $100,000 is not out of line. Its Nvidia branding mostly means the company supplies the board and partners build the rest.25

Memory costs are pushing up the whole lineup

The DGX Station's price matters most for what it shows about Nvidia's desktop lineup as a whole. The DGX Spark, Nvidia's small GB10-based machine, launched at $3,999. Nvidia raised it to $4,699 in February 2026, citing supply constraints on its 128GB LPDDR5X memory.1 On October 2, the 128GB model went up again, to $6,950. Nvidia also announced a 64GB version at $4,999, arriving through partners on October 23.39 One analysis calls that roughly a 74% increase over the launch price and links it directly to the 2026 DRAM and HBM shortage.5

That trend affects the Station too. Its 748GB pool is mostly the same kind of memory driving Spark's price up, plus HBM3e, which is even harder to get. Coverage of comparable desktops reflects the same pressure. TechRadar noted that Supermicro's machine costs about seven times as much as a 256GB Mac Studio M5 Ultra at $12,299, though the Mac has far less memory.

The two DGX machines are also very different, despite the shared branding. The Spark has 128GB at about 273 GB/s, while the Station's HBM3e runs at 7.1 TB/s. The price gap of roughly twenty times largely reflects that difference in memory bandwidth.25 For inference, bandwidth largely determines how fast tokens come out. Reviewers have measured the Spark generating text from a 70B model at about 2 to 3 tokens per second.6 Teams that want large models to respond quickly on local hardware have no option between a $7,000 Spark and a $95,000 Station.

The takeaway

Nvidia has left pricing to its partners, and they have landed around $90,000 to $100,000 for a GB300 desktop that matches data-center hardware in memory capacity. This isn't a consumer product, and it isn't really a bargain either. It is an option for organizations that keep it busy, need to keep data in-house, or simply want to own hardware with HBM. The Spark has already gone through two price increases tied to memory supply. That suggests buyers waiting for the Station to get cheaper shouldn't expect it soon.