On 24 October 2026, the default builds in three US sellers' online configurators put an HGX B200 server price at $443,278 to $586,947 and an HGX B300 server at $496,639 to $672,806 [22, 26, 30]. Two listings sat outside those ranges: an unlinked Exxact B200 page at $236,076.50 and a Vipera B300 server at $760,000 [21, 31]. Each figure is a seller's asking price on that day. None records a sale.
Each figure prices a different machine. An eight-GPU HGX server is a GPU baseboard in a chassis with two host CPUs, host memory, network cards, storage, power supplies and a warranty. Each seller picks its own default parts, and some defaults carry less host memory than the NVIDIA minimums that Dell and Thinkmate relay [15, 27, 29, 30].
No configurator reviewed prices the GPU board on its own line, so no listing shows the GPUs' share of the price [21, 24, 27].
What an HGX B200 server price includes
NVIDIA describes HGX as a single baseboard with eight Blackwell or Blackwell Ultra SXM GPUs [1]. Server makers build the rest of the system around it.
| Part | What the specifications show | How it appears in a listing |
|---|---|---|
| GPU baseboard | Eight GPUs. NVIDIA lists 1.4 TB of GPU memory for HGX B200 and 2.1 TB for HGX B300 | A selected line with no separate price |
| Host CPUs | Two sockets. Supermicro's two B200 systems take 4th or 5th Gen Xeon (up to 64 cores) or Xeon 6900 (up to 128 cores) | Chosen per build, with upgrade prices |
| Host memory | DDR5 RDIMMs in 24 or 32 slots, with a maximum of 4 TB to 8 TB | Module size and count, priced per option |
| Cluster network cards | Eight ConnectX-8 800 Gb/s ports on B300 platforms. Add-in cards on B200 systems | B300: included. B200: optional, none by default |
| Storage | NVMe U.2 or E1.S drives | Zero to six drives by default (Thinkmate's defaults include none), upgrades priced |
| Power supplies | Six 5,250 W units (Supermicro B200) or twelve 3,200 W units (Dell XE9780) | Included |
| Warranty | Three years of labor and one year of parts as Supermicro's standard for GPU systems | Seller terms of three to five years |
Sources: [1, 11, 12, 13, 14, 15, 21, 24, 27, 28, 29, 31].
NVIDIA states B300 memory two ways: 2.1 TB, or 270 GB per GPU, on its HGX page and datasheet [1, 3], and "8 x 288 GB = 2.3 TB total" in its DGX B300 user guide [7].
Host memory sets a floor under the build
Dell's technical guide for the PowerEdge XE9780 relays NVIDIA's design guide: an eight-GPU server needs more than 1 TB of host memory with B200 GPUs and more than 2 TB with B300 [15]. Thinkmate's configurators relay a different B200 figure, "1.5 TB minimum system memory", along with 2 TB for B300 and at least 48 physical CPU cores per socket [27, 28]. No public copy of the design guide was found, so neither relay can be checked against it.
Several default builds sit under these floors. Thinkmate's B200 default carries 1 TB [27]. Bizon's B200 default carries 512 GB and its B300 default 384 GB, both with two 24-core CPUs [29, 30]. NVIDIA's own DGX B200 ships with 2 TB of system memory [5], and the DGX B300 with 2 TB by default [7].
Raising Thinkmate's B200 build from 1 TB to 2 TB adds $81,600, or 16% of its $504,905 default price [27].
Listed HGX B200 and B300 server prices
| Listing | GPU board | Host CPUs | Host memory | Cluster NICs | Listed price |
|---|---|---|---|---|---|
| Exxact TS4-124903244 (10U) | B200 | 2 x Xeon Platinum 8468 | 1,536 GB | None by default | $236,076.50 |
| Exxact TS4-142487900 | B200 | 2 x EPYC 9554 | 1,536 GB | None by default | $443,278.00 |
| Exxact TS4-128275176 | B200 | 2 x Xeon Platinum 8558P | 2,048 GB | None by default | $482,438.00 |
| Bizon X9000 G4 | B200 | 2 x EPYC 9255 | 512 GB | Optional | $496,728 |
| Thinkmate GXS6-2808-8HGX-B2 | B200 | 2 x Xeon 6740P | 1 TB | None by default | $504,905.00 |
| Bizon X9000 G5 | B300 | 2 x EPYC 9255 | 384 GB | 8 x ConnectX-8 | $496,639 |
| Exxact TS4-131188947 | B300 | 2 x Xeon 6747P | 2,048 GB | 8 x 800 Gb/s | $516,036.40 |
| Exxact TS4-198437339 | B300 | 2 x Xeon 6740P | 3,072 GB | 8 x 800 Gb/s | $564,718.00 |
| Thinkmate GXS6-2808-8HGX-B3 | B300 | 2 x Xeon 6740P | 2 TB | 8 x ConnectX-8 | $664,480.00 |
| Vipera SYS-822GS-NB3RT | B300 | 2 x Xeon 6767P (parts list) | 4 TB (parts list) | 8 x ConnectX-8 | $760,000.00 |
Sources: [21, 22, 23, 24, 25, 27, 28, 29, 30, 31]. Listings viewed 24 October 2026, default builds, in US dollars.
Thinkmate's other HGX models start at $543,070 to $586,947 for B200 (models ending -B2) and $666,915 to $672,806 for B300 (models ending -B3) [26]. In Europe, Servermall listed a Supermicro B300 system at €709,787 with 2,304 GB of memory and €738,668 with 3,072 GB, both before VAT [32].
Exxact's TS4-124903244 page is the outlier at its own seller. On 24 October 2026 it listed $207,201.50 below Exxact's AMD-based B200 build, and Exxact's HGX category page did not link to it [20, 21, 22]. The page gives no reason.
What one line item adds
The same change costs different amounts at different sellers on the same day.
| Change | Seller | Added to the price |
|---|---|---|
| B300, 2,048 GB to 3,072 GB of DDR5-6400 | Exxact | $26,400 |
| B300, 2 TB to 3 TB of DDR5-6400 | Thinkmate | $58,944 |
| B200, 1,536 GB to 2,048 GB (32 x 64 GB) | Exxact 10U | $33,440 |
| B200, eight ConnectX-7 400 Gb/s cards | Exxact | $14,520 ($1,815 each) |
| B200, eight ConnectX-7 400 Gb/s cards | Thinkmate | $17,688 ($2,211 each) |
| B200, two Xeon 8468 to two Xeon 8490H | Exxact 10U | $19,349 |
Sources: [21, 24, 27, 28]. Card totals are eight times the per-card price.
Network cards change the comparison between generations. Dell's guide states: "For AI and HPC deployments, a 1:1 NIC-to-GPU ratio is recommended" [15]. B300 platforms carry eight cluster network ports as standard [13, 16, 24, 28], while B200 configurators default to none [21, 22, 23, 27]. So the two default prices cover different scope.
Line items are seller prices, and they move. TrendForce projected in September 2026 that conventional DRAM contract prices would rise 10% to 15% in the fourth quarter and that server DRAM would stay undersupplied [35].
Software, support and warranty
NVIDIA AI Enterprise is licensed per GPU, at a list price of $4,500 per GPU for one year or $18,000 per GPU for a five-year subscription [9, 10]. NVIDIA states that Blackwell DGX systems need these licenses bought separately [9]. Thinkmate offers it as an add-on at $17,999 per GPU for five years and Exxact at $19,800 [24, 27]. At list price, five years for eight GPUs comes to $144,000 [10].
Supermicro's standard for GPU systems is three years of labor and one year of parts [14]. Exxact includes three years of parts and labor, Thinkmate five years of advanced parts replacement, and Bizon five years of labor with three of parts [21, 27, 29]. Thinkmate's prices exclude shipping and taxes [27].
Price per GPU: server, cluster and deployed cost
Dividing a listing by eight gives $55,410 per GPU for Exxact's AMD-based B200 build and $83,060 for Thinkmate's B300 build (arithmetic on [22, 28]). That figure carries the CPUs, memory and chassis. Epoch AI divides DGX system prices the same way [34].
Fleet figures carry more. On 9 October 2026, Exascale Labs agreed to sell 12 HGX B300 servers with network equipment, management servers, deployment services and 36 months of IT operation services for $8,208,000, exclusive of import duties and taxes [18]. That is $85,500 per GPU (arithmetic on [18]), with no split between equipment and services. Axe Compute bought a company holding 288 B300 servers and their networking by repaying its debt of about $87.8 million and agreeing to pay $715,000 a month for 60 months [19].
IREN's July 2025 purchase of about 2,400 B200 and B300 GPUs for about $130 million included servers, storage, labor, InfiniBand, racks and cabling [17]. The evidence classes note works through its per-GPU figures.
A cluster adds parts no server listing shows. NVIDIA's reference design for a 128-node, four-unit DGX B200 SuperPOD uses 32 leaf and 16 spine InfiniBand switches, with 1,024 node-to-leaf and 1,024 leaf-to-spine cables [8]. The racks that hold them need power and cooling, covered in rack power and liquid cooling, and paying for the systems is the subject of GPU financing structures.
Rillor's GPU systems pricing method models configurations by GPU family, GPU count, memory, CPU, interconnect, cooling, condition, location and seller type, and each observation in its schema keeps its evidence class, so a listing is read as a listing and a quote as a quote. To buy the systems themselves, Rillor specifies, sources and delivers GPU servers, and each quote states the exact build, its condition and the tests run before delivery.
Read the full report
The full report adds the specification sheets part by part, the listings brought to a common build, the European and DGX listings, and the numbered sources.