The DGX Station GB300 price is public in exactly one place: MSI's $85,000 XpertStation WS300. Every other certified vendor selling the same box, Supermicro, Exxact, Dell, ASUS, and HP, takes orders without a posted list price. That gap matters less than the question it obscures: is one of these worth buying at all, when the obvious alternative is renting GB300-class compute by the hour instead?
TL;DR: Is the DGX Station GB300 Worth It Over Renting?
- Price: $85,000 is the only confirmed DGX Station GB300 price, MSI's XpertStation WS300. Other OEMs sell it unlisted.
- One GPU, not a rack: a DGX Station GB300 is one B300 Superchip; a GB300 NVL72 rack holds 72 of them, so the fair comparison is one accelerator-hour, not a rack.
- Break-even: against AWS's $14.04-per-accelerator-hour P6-B300 Capacity Block, the only published rate, $85,000 pays back in ~252 days at 24/7 use, or 2.9 years at business hours.
- The catch: up to 1,600W on a dedicated 20A circuit, which most office desks lack.
- Spheron sells standalone B300 on-demand at $5.40/hr per GPU as of 17 Sep 2026 and reserves GB300 NVL72 by the node. Reserve GB300 NVL72 →
What the DGX Station GB300 Actually Is: One GB300 Superchip, Not a Rack
A DGX Station GB300 is one GB300 Superchip: one B300 Blackwell Ultra GPU paired with one Grace CPU over NVIDIA's own coherent memory fabric, not a slice of rack capacity and not multiple GPUs in one case. NVIDIA's own DGX Station GB300 spec sheet lists 252GB of HBM3e GPU memory running at 7.1 TB/s, plus 496GB of LPDDR5X CPU memory, for 748GB of total coherent memory addressable by both processors at once. That coherent pool is the point of the design: the CPU and GPU share one address space over a 900 GB/s NVLink-C2C link, so a model or dataset doesn't need to be copied across a PCIe bus the way it would on a conventional workstation. Networking runs through a ConnectX-8 SuperNIC at up to 800 Gb/s.
That single-GPU framing matters because it's easy to confuse with NVIDIA's rack-scale GB300 NVL72, which shares the same B300 silicon but at a completely different scale: 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single liquid-cooled cabinet, with 20 TB of pooled GPU memory across those 72 GPUs and a 130 TB/s NVLink fabric tying them together. One box is a desk-side workstation. The other is a data-center rack most teams will never buy outright. Our GB300 NVL72 vs GB200 NVL72 pricing comparison covers the rack-scale specs and the analyst price estimates for that cabinet in full.
How It Differs From DGX Spark (GB10) and a GB300 NVL72 Rack
NVIDIA sells three completely different things under a "DGX" or "GB" name, and conflating any two of them produces a bad buying decision.
DGX Spark runs the much smaller GB10 Grace Blackwell Superchip, a 140W chip built for local development rather than production throughput, with its 128GB configuration priced at $3,999 to $6,950, roughly 12 to 21 times cheaper than the DGX Station GB300. Our DGX Spark alternatives comparison covers the full GB10 lineup across ASUS, Dell, HP, Lenovo, and the rest, and our DGX Spark to GPU cloud pipeline guide covers exactly where a GB10 box hits its ceiling and when to move to rented GPUs.
DGX Station GB300 is this post's subject: one full-power B300 Superchip in a workstation chassis, aimed at teams that need data-center-class single-GPU throughput on a desk rather than a rack.
GB300 NVL72 is the rack: 72 of the same B300 GPUs wired into one NVLink domain for workloads that genuinely need more memory and bandwidth than a single GPU, or even a single 8-GPU node, can offer.
DGX Station GB300 Price: $85,000 Confirmed at MSI, Unlisted Elsewhere
| OEM | Price | Notes |
|---|---|---|
| MSI (XpertStation WS300 / GB300 Ultra) | $85,000 | Published list price |
| Supermicro | Not publicly listed | Certified DGX Station GB300 partner |
| Exxact | Not publicly listed | Certified DGX Station GB300 partner |
| Dell | Not publicly listed | Certified DGX Station GB300 partner |
| ASUS | Not publicly listed | Certified DGX Station GB300 partner |
| HP | Not publicly listed | Certified DGX Station GB300 partner |
MSI re-launched its XpertStation WS300 at a published price of $85,000. That figure is the one hard number in this entire market. Budget at or above MSI's confirmed number and request a direct quote from whichever OEM you're actually buying through; that is the honest procurement reality for this product tier.
MSI's $85,000 figure is its most recently published list price for the XpertStation WS300 and may have changed since; MSI does not publish a price change log. Confirm the current figure directly with MSI or your OEM sales contact before ordering.
Price by Vendor: Supermicro, MSI, Exxact
MSI is the only vendor with a documented, independently-reviewed price and, as of this post, the only one StorageReview has put through hands-on testing. Supermicro and Exxact both sell certified DGX Station GB300 configurations, and both historically operate through enterprise sales channels rather than a retail checkout page, which is consistent with having no public list price. Dell, ASUS, and HP round out the OEM roster the same way. There is no self-serve "add to cart" price for this product outside of MSI's retail listing; every other certified build goes through a sales conversation first.
Full Specs: What One DGX Station GB300 Delivers
The table below is drawn directly from NVIDIA's published DGX Station GB300 specifications.
| Spec | DGX Station GB300 |
|---|---|
| GPU | 1x B300 Blackwell Ultra Superchip |
| GPU memory | 252GB HBM3e at 7.1 TB/s |
| CPU memory | 496GB LPDDR5X |
| Total coherent memory | 748GB (GPU + CPU) |
| CPU-GPU link | 900 GB/s NVLink-C2C |
| Networking | ConnectX-8 SuperNIC, up to 800 Gb/s |
| Power | Up to 1,600W |
| Circuit requirement | 20A |
The spec sheet is one thing; what it actually produces running real models is another, and that's where StorageReview's hands-on testing of the MSI XpertStation WS300 is the only independent data point available. On GPT-OSS-120B, the WS300 measured 5,038 output tokens per second against 872 output tokens per second on an RTX PRO 6000 Blackwell workstation, a gap approaching 6x on that model. StorageReview also measured the WS300 at 14 to 19x the throughput of a DGX Spark across the models it ran on both machines, which is the clearest evidence yet that the price gap between a GB10 box and a GB300 Superchip tracks a real throughput gap, not just a marketing tier.
Two more data points from the same review, on models that push the WS300's memory and quantization harder: DeepSeek v4 Flash ran at 1,766 output tokens per second at concurrency 32 (3,532 total tokens per second across all concurrent requests, on the review's 512-input/512-output workload), and MiniMax M2.7 in NVFP4 hit 4,801 output tokens per second at concurrency 128. Those numbers come from one review unit under one test methodology, so treat them as a directional benchmark rather than a guarantee of what any specific workload will see, but they're the only third-party numbers that exist for this hardware right now. For the chip-level breakdown of the B300 itself, including how it compares to the B200 and H200 on memory and FP8 throughput, see our NVIDIA B300 Blackwell Ultra guide.
What One Unit Can and Can't Do vs a Reserved GB300 NVL72 Slice
748GB of coherent memory on one GPU-and-CPU pair is enough to hold very large single-model inference and fine-tuning jobs entirely locally, offline, with no network dependency and no per-hour meter running. That's the DGX Station GB300's actual advantage over any form of rental, and it's a real one for air-gapped work or for a team that iterates constantly and would rather pay once than watch a clock.
What it can't do is scale past its own NVLink-C2C link. A GB300 NVL72 rack's 130 TB/s NVLink fabric and 20 TB of pooled HBM3e exist specifically for workloads too large for a single GPU, or even a single 8-GPU node, to hold: trillion-parameter pre-training, frontier MoE serving at high concurrency, or reasoning workloads with KV caches too large for one accelerator. The moment a workload needs that kind of multi-GPU NVLink domain, a DGX Station GB300, or ten of them sitting unconnected on separate desks, doesn't get you there. That's the point where reserving rack-scale capacity replaces buying desk-side hardware.
Spheron keeps GB300 NVL72 capacity open to reserve sized from a single node up to a full rack rather than forcing a 72-GPU commitment, which is the direct cloud-side comparison to one desk-side unit once it stops being enough. Worth saying plainly: that reservation runs on a quote, with a typical 24-to-48-hour turnaround on Spheron's own pricing page, not an instant checkout rate, so it can't be plugged into the hourly break-even math below the way AWS's published Capacity Block rate can. For workloads that need GB300-class throughput without committing to a rack reservation, standalone B300 GPUs are available self-serve and billed per minute, at $5.40/hr on-demand and $5.73/hr spot as of 17 Sep 2026.
Pricing fluctuates based on GPU availability. Spheron rates above are live as of 17 Sep 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.
Break-Even Math: DGX Station GB300 Price vs Renting GB300-Class Compute
The Only Published Hourly Rate to Run the Math Against
A DGX Station GB300 is a single accelerator, so the only fair rental comparison is a single accelerator-hour, not a 72-GPU rack reservation. AWS publishes exactly one figure for that: its P6-B300 Capacity Block, listed on AWS's own Capacity Blocks pricing page at $14.04 per accelerator-hour in US regions as of July 2026 (the full 8-GPU p6-b300.48xlarge block runs $112.32/hr, which works out to that same $14.04 per GPU). A Capacity Block is a reservation for a defined future window rather than pure pay-as-you-go, but it's still the only rate anyone has published for GB300-class silicon, which is exactly the gap this post exists to close. For the full rate history and how that figure compares to Spheron's own B300 pricing, see our AWS Capacity Blocks pricing breakdown.
AWS's $14.04 per accelerator-hour P6-B300 rate reflects its published Capacity Blocks price as of July 2026 and may have changed since; AWS updates these rates periodically and without advance notice. Confirm the current figure on AWS's own pricing page before budgeting against it.
Break-Even at 24/7 Use vs Business-Hours Use
$85,000 divided by $14.04 per accelerator-hour is about 6,054 GPU-hours. How long that takes to clock depends entirely on usage pattern.
| Usage pattern | GPU-hours needed | Time to break even |
|---|---|---|
| 24/7 continuous use | ~6,054 hours | ~252 days (~8.3 months) |
| 8 hours/day, 5-day week | ~6,054 hours | ~1,060 calendar days (~2.9 years) |
Run it 24/7 and the DGX Station GB300 catches up with the AWS rate in about eight months. Run it on a normal business schedule, 8 hours a day, 5 days a week, and that stretches to roughly 2.9 years before the purchase has paid for itself against the hourly alternative. Below that usage level, renting wins outright; a GPU that sits idle most of the week never recovers an $85,000 upfront cost against a per-hour rate. Our LLM inference on-premise vs GPU cloud break-even analysis covers the same utilization-threshold logic in more general form, for teams weighing owned hardware against cloud GPUs at any scale.
The Cost Nobody Puts on the Price Tag: Power and Facilities
Every one of those break-even numbers assumes the GPU-hours are free to generate once you own the box. They aren't. NVIDIA's own spec sheet puts the DGX Station GB300 at up to 1,600W, which requires a dedicated 20A circuit, a provision most shared office space and home setups don't have without an electrician involved first. StorageReview's hands-on testing of the MSI XpertStation WS300 measured the B300 GPU alone peaking at 1,292W, against a cooling system rated for 1,400W combined CPU and GPU draw, both comfortably below NVIDIA's 1,600W circuit specification in that unit's real-world testing. That's good news for anyone worried about hitting the circuit ceiling in practice, but the facility still has to be provisioned for the full 1,600W rating, and the electricity bill runs on top of the $85,000 sticker price for as long as the box is plugged in. Our AI inference power and electricity cost guide walks through how to fold a GPU's wattage into a real cost-per-hour figure if you want to add that into your own break-even model rather than treating the purchase price as the whole cost.
Who Should Actually Buy One (and Who Should Just Rent)
StorageReview put it plainly in its review of the MSI XpertStation WS300: "The honest comparison is not against other workstations but against what an AI team is already paying for: reserved cluster time, cloud GPU commitments, and the schedule cost of waiting on shared hardware." Measured against that bar rather than against another desk-side box, the reviewer concluded: "One WS300 is easy to justify against reserved cluster time."
Buy one if your usage pattern comfortably clears the break-even math above, your workload genuinely needs offline, no-network local iteration that a rented GPU can't offer, and you have the 20A circuit and cooling to run it. That describes a team doing heavy, constant single-GPU fine-tuning or large-model local inference, where the schedule cost of waiting on shared cluster time is a real daily tax.
Rent instead if your usage is bursty, if the workload needs more than one GPU's worth of NVLink domain (at which point a GB300 NVL72 reservation, not a second desk-side box, is the only path forward), or if you'd rather not take on a 1,600W circuit upgrade for hardware that might sit idle most weeks. For most teams evaluating GB300-class compute for the first time, that's the more defensible starting point, and it's the one that doesn't require a wire transfer before you know whether the workload needs this much GPU at all.
If the break-even math above doesn't clear for your usage pattern, renting a single GB300-class accelerator by the hour skips the circuit upgrade and the OEM sales call entirely.
Frequently Asked Questions
One. A DGX Station GB300 pairs a single B300 Blackwell Ultra GPU with a single Grace CPU, NVIDIA's own spec sheet describes 252GB of HBM3e GPU memory plus 496GB of LPDDR5X CPU memory addressable by both processors, for 748GB total. That is different from a GB300 NVL72 rack, which NVIDIA's own GB300 NVL72 specifications list as 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single liquid-cooled cabinet. Confusing the two is the single most common mistake in this search: one is a desk-side box, the other is a 72-GPU rack.
$85,000 is the only publicly confirmed price, for MSI's XpertStation WS300 (also marketed as the GB300 Ultra), re-launched at that figure. Supermicro, Exxact, Dell, ASUS, and HP all sell certified DGX Station GB300 systems through their own order channels, but none of them publishes a list price, so budget around MSI's confirmed figure and get a direct quote before committing.
It depends entirely on how many hours a year the GPU actually runs. Against AWS's $14.04 per accelerator-hour P6-B300 Capacity Block, the only published GB300-class hourly rate anywhere, an $85,000 DGX Station GB300 breaks even in roughly 252 days of 24/7 use, or about 2.9 years at 8 hours a day, five days a week. Below that usage level, renting by the hour is cheaper. Above it, and with a genuine need for offline, no-network local iteration, owning one starts to pay for itself.
They are different products at a different order of magnitude. DGX Spark runs NVIDIA's much smaller GB10 Grace Blackwell Superchip and its 128GB configuration is priced at $3,999 to $6,950, roughly 12 to 21 times cheaper than the DGX Station GB300's $85,000 MSI price point. DGX Spark is a development box for models up to roughly 200B parameters with quantization; the DGX Station GB300 is a full-power single-GPU workstation built around the same B300 silicon that goes into GB300 NVL72 racks.
Nobody publishes an official rack price, on either side. Analyst estimates for a GB300 NVL72 rack (72 GPUs) range from roughly $3.7 million to $6.5 million depending on which estimate you believe, against a single DGX Station GB300's confirmed $85,000. That is not a meaningful price comparison, since the rack and the desk-side box serve entirely different scales of workload; it only matters once a workload genuinely outgrows a single GPU's NVLink domain.
