What an HGX B200 or B300 server price includes

Summary

• On 24 October 2026, the default builds in the online configurators of Exxact, Thinkmate and Bizon listed eight-GPU HGX B200 servers at $443,278 to $586,947 and HGX B300 servers at $496,639 to $672,806 [22, 26, 30]. One unlinked Exxact B200 page listed $236,076.50, and a Vipera B300 listing asked $760,000 [21, 31]. All are listings seen on one day. None records a sale.
• An HGX server price covers the eight-GPU baseboard plus two host CPUs, host memory, network cards, storage, power supplies, the chassis and a warranty. CPU socket, DIMM count and maximum memory change around the same HGX B200 board [11, 12], and no configurator reviewed prices the board on its own line [21, 24, 27].
• Host memory floors relayed from NVIDIA's design guide (more than 1 TB for B200 and more than 2 TB for B300 in Dell's guide, 1.5 TB and 2 TB on Thinkmate's pages) sit above several default builds [15, 27, 28, 29, 30]. Raising Thinkmate's B200 build from 1 TB to 2 TB adds $81,600 [27].
• B300 platforms include eight ConnectX-8 800 Gb/s ports, while B200 configurators leave cluster network cards out of the default build [13, 21, 27]. The same change costs different amounts at different sellers: on 24 October 2026, 2,048 GB to 3,072 GB of host memory on a B300 server added $26,400 at Exxact and $58,944 at Thinkmate [24, 28]. Bringing listings to a common build widens the gap between sellers.
• Per-GPU figures from fleet deals carry networking, services and payment terms. Exascale Labs' sale of 12 HGX B300 servers with networking, deployment and 36 months of operations works out to $85,500 per GPU [18], against $64,505 to $83,060 per GPU for Exxact and Thinkmate B300 listings divided by eight (arithmetic on [24, 25, 28]).

Background

Buyers, finance teams and analysts often need a price for an eight-GPU Blackwell server, to budget a purchase, check a quote or value systems in service.

NVIDIA publishes HGX specifications and no HGX price. Its HGX page describes the platform as "a single baseboard with eight NVIDIA Rubin, NVIDIA Blackwell, or NVIDIA Blackwell Ultra SXMs" [1], and its datasheets list the server options as partner and NVIDIA-Certified Systems [2, 3]. Dell's product page for the PowerEdge XE9780 shows no price and offers "Contact Sales" [16]. Public prices come from sellers with online configurators and product pages [21, 26, 29, 31, 32].

NVIDIA's DGX B200 and DGX B300 are complete systems with published specifications, where system memory is the main option (2 TB, up to 4 TB) [5, 7]. An HGX server is one maker's or seller's build around the same class of GPU board, and every other part can change.

A listing records what a seller asks in public at a point in time. It differs from a quote to a named buyer, from a reported sale and from a verified transaction. The evidence classes note sets out those classes. Listings carry terms: Thinkmate's prices exclude shipping and taxes and "are intended for same-day purchase" [27]. Bizon showed lead times of 9 to 12 weeks, and Vipera 8 to 14 weeks [29, 30, 31].

TrendForce projected in September 2026 that conventional DRAM contract prices would rise 10% to 15% quarter on quarter in the fourth quarter of 2026, and that "Server DRAM is therefore expected to remain undersupplied in 4Q26" [35]. Host memory is a large configurable item, so a listing's date matters.

Method

What was examined. Five groups of sources, all opened on 24 October 2026.

• Specifications: NVIDIA's HGX, DGX and SuperPOD documents [1-8], three Supermicro systems [11, 12, 13] and Dell's PowerEdge XE9780 [15, 16].
• Terms: NVIDIA AI Enterprise licensing and pricing [9, 10] and Supermicro's and Dell's warranties [14, 15].
• Listings: 14 public pages from five sellers. Exxact's HGX category page and five configurators [20-25], Thinkmate's GPU server list and two configurators [26, 27, 28], two Bizon product pages [29, 30], one Vipera product page [31] and two Servermall listing pages in Europe [32, 33].
• Reported sales: three filings with the U.S. Securities and Exchange Commission, from IREN, Exascale Labs and Axe Compute [17, 18, 19].
• Method and market context: Epoch AI's per-GPU price method [34] and TrendForce's memory price forecast [35].

How. For each listing, the default build was recorded part by part, with its price, currency and stated tax treatment, and line items were recorded as shown. Thinkmate and Bizon pages were read in a headless browser, because their line-item prices render only there. Supermicro's pages were read through a fetch tool. Prices carry the evidence class listing or reported sale. Other sources are specifications, terms, method or forecast.

Rules. Prices are never averaged, and every price keeps its date and its build. Derived figures are labelled as arithmetic and cite their inputs. Seller page errors are reported as shown. Unsourced reseller ranges and component splits were left out, as was one HGX B200 listing seen earlier in October 2026 that could not be reopened.

Findings

1. The parts of an eight-GPU HGX server

An HGX server has one fixed part, the GPU baseboard, and a set of parts that vary by maker and by build. NVIDIA's HGX table lists the board's figures: eight SXM GPUs, fifth-generation NVLink with NVLink 5 switches, 1.8 TB/s of GPU-to-GPU bandwidth, total memory of 2.1 TB for HGX B300 and 1.4 TB for HGX B200, and networking bandwidth of 1.6 TB/s for HGX B300 and 0.8 TB/s for HGX B200 [1]. NVIDIA's datasheets give a maximum thermal design power of up to 1,000 W per B200 GPU and up to 1,100 W per B300 GPU [2, 3].

The rest of the server comes from the maker. Four specification sheets show how much it moves.

Sources: [11, 12, 13, 15].

CPU and memory vary around one board. Supermicro's SYS-A21GE-NBRT takes 4th or 5th Gen Xeon on socket LGA-4677 with 32 DIMM slots. Its SYS-A22GA-NBRT takes Xeon 6900 on socket LGA-7529 with 24 DIMM slots [11, 12]. Both carry the same HGX B200 8-GPU board.

Network cards sit inside the B300 platform. Supermicro's HGX B300 system lists "8 OSFP 800 GbE LAN ports (NVIDIA ConnectX-8 SuperNIC)" as standard [13]. NVIDIA's DGX B300 user guide states that "ConnectX-8 networking connects directly to GPUs via PCIe Gen6" and that the "NICs, including cages for connectivity, are housed in the GPU tray" [7]. Dell's product page lists "B300: 8 x CX8 OSFP (default)" [16]. On B200 systems, cluster network cards are add-in PCIe cards. Dell's guide lists up to 12 ConnectX-7 400G cards for its B200 configuration [15], and warns: "For AI and HPC deployments, a 1:1 NIC-to-GPU ratio is recommended to deliver optimal network throughput across multi-server environments" [15].

Facility power sits outside the server price. NVIDIA lists the DGX B200 at about 14.3 kW maximum and the DGX B300 at 14.5 kW [5, 7]. Thinkmate estimated 19,182 W for its default HGX B300 build [28].

NVIDIA's reference builds show a complete specification. The DGX B200 carries two Xeon Platinum 8570 CPUs, 2 TB of system memory configurable to 4 TB, eight ConnectX-7 cards at up to 400 Gb/s, two BlueField-3 DPUs and eight 3.84 TB U.2 drives. It includes DGX OS, lists NVIDIA AI Enterprise as "Recommended", and comes with three-year Enterprise Business-Standard Support [5]. The DGX B300 carries two Xeon 6776P CPUs, 2 TB of system memory by default with up to 4 TB, eight ConnectX-8 cards at 800 Gb/s, two BlueField-3 DPUs and eight 3.84 TB E1.S drives, with three-year business-standard support [6, 7].

GPU memory is stated two ways. For HGX B300, NVIDIA's HGX page, DGX B300 page and Blackwell Ultra datasheet give 2.1 TB, or 270 GB per GPU [1, 3, 6]. Its DGX B300 user guide gives "8 x 288 GB = 2.3 TB total" [7], as do Supermicro and Dell's technical guide [13, 15], while Dell's product page gives 270GB [16]. No source reviewed explains the difference. For HGX B200, NVIDIA's datasheet gives 180 GB per GPU [2] and its technical brief "Up to 192 GB" [4]. Exxact's configurators say 180 GB [21], while Bizon and Exxact's category page say 192 GB [20, 29].

2. Memory and CPU floors for B200 and B300

Dell's technical guide relays NVIDIA's design guide in these words: "NVIDIA design guide require the minimum total memory capacity for Server with 8-way GPU. B200 GPU: the minimum total memory capacity must greater than 1TB per system. B300 GPU: the minimum total memory capacity must greater than 2TB per system" [15].

Thinkmate's configurators relay different figures. Its B200 page states: "Note: For HGX B200 systems, NVIDIA recommends 1.5 TB minimum system memory" [27]. Its B300 page states: "Note: For HGX B300 systems, NVIDIA recommends 2 TB minimum system memory" [28]. Both pages add: "NVIDIA recommends a minimum of 48 physical CPU cores (preferably 56 or more) per socket" [27, 28].

For B200, the two relays disagree: more than 1 TB in one, 1.5 TB in the other. No public copy of the design guide was found, so neither relay can be checked against it.

Five of the ten default builds reviewed meet both relays: the three Exxact B200 builds (1,536 GB or 2,048 GB), Exxact's 3,072 GB B300 build and Vipera's 4 TB parts list [21, 22, 23, 25, 31]. Two B300 builds hold exactly 2 TB, which meets Thinkmate's figure and sits at the line of Dell's "greater than 2TB" [24, 28]. Three fall below both: Thinkmate's B200 build with 1 TB, Bizon's B200 build with 512 GB and Bizon's B300 build with 384 GB [27, 29, 30]. Both Bizon builds also use two 24-core CPUs, half the 48 cores per socket that Thinkmate relays [27, 29, 30].

The DGX B200 ships with 2 TB, above both B200 relays, and the DGX B300 with 2 TB by default, at the B300 line [5, 7].

Reaching the floor costs money. Thinkmate's B200 build goes from 1 TB to 2 TB (32 x 64 GB) for $81,600, or 16% of its $504,905 default price (arithmetic on [27]). Bizon's B200 build goes from 512 GB to 2,048 GB (16 x 128 GB) for $76,915, and its 48-core EPYC 9455 pair adds $6,556 [29]. Bizon's B300 page lists 12-DIMM options for its "AMD/Intel platform", so 12 x 128 GB (1,536 GB) adds $57,686 and still sits below 2 TB. Its 32-DIMM options belong to an "Intel platform" and start at $79,708 for 32 x 64 GB, while a pair of 48-core Xeon 6740P CPUs adds $26,220 [30]. A default price that sits below the relayed floor leaves part of a usable build out of the figure.

3. Public listings, with capture dates and builds

The US configurator and product pages reviewed, each with its default build, viewed on 24 October 2026:

Sources: [21-25, 27-31]. Warranty terms are in Finding 6. Vipera's build is read from the parts list on its page, which gives quantities for 64 systems [31].

Thinkmate's list page shows starting prices for its other HGX models: $543,070 (AMD EPYC), $547,603 (Xeon 6 AP) and $586,947 (5th or 4th Gen Xeon) for B200 (models ending -B2), and $666,915 (AMD EPYC) and $672,806 (Xeon 6 AP) for B300 (models ending -B3) [26].

Europe. Servermall lists systems in euros, with VAT shown separately, shipping included across the EU, the UK and the USA, and a five-year onsite warranty [32, 33].

Sources: [32, 33]. Viewed 24 October 2026.

On the same Supermicro B300 model, adding eight 96 GB modules (2,304 GB to 3,072 GB) adds €28,881 before VAT (arithmetic on [32]). The DGX listings price NVIDIA's complete systems [5, 6].

An outlier at its own seller. On 24 October 2026, Exxact's 10U B200 page TS4-124903244 listed $236,076.50, which was $207,201.50 below Exxact's AMD-based B200 build and $246,361.50 below its Intel-based 8U B200 build (arithmetic on [21, 22, 23]). Exxact's HGX category page listed six B200 and B300 configurators and did not link TS4-124903244 [20]. Its platform text closely follows Supermicro's SYS-A21GE-NBRT (socket LGA-4677, 32 DIMM slots, eight low-profile PCIe 5.0 x16 slots, ten 2.5-inch NVMe bays), but its power supply line reads "8x 3000W (6+2)" where Supermicro lists six 5,250 W units [11, 21]. The page gives no reason for its price, and no source shows whether it would be honored.

Page anomalies, as shown. On the 10U page, the Premium tier ($243,446.50) costs less than Professional ($273,916.50), because Premium carries 1,536 GB of memory and Professional 2,048 GB [21]. A 2,048 GB option through 16 modules of 128 GB costs $20,240, below the $33,440 for 32 x 64 GB [21]. Exxact's CPU lists show errors such as "Intel Xeon Platinum 8470 2 Core" [21, 23]. Bizon's B200 page shows "Starting at $496,728" while its configured total read $496,639 [29].

4. Line-item prices in public configurators

Configurators price each change against the default build. The table lists line items from the configurators reviewed, as shown on 24 October 2026.

Memory moves the total by tens of thousands of dollars. Taking a default build to the relayed floor costs $76,915 to $81,600 at Bizon and Thinkmate [27, 29]. Exxact's B200 defaults already hold 1,536 GB or more [21, 22, 23].

Sellers price the same step differently. Moving a B300 server from 2,048 GB to 3,072 GB of DDR5-6400 adds $26,400 at Exxact and $58,944 at Thinkmate on the same day [24, 28]. One ConnectX-7 400 Gb/s card is $1,815 at Exxact, $2,211 at Thinkmate and $2,613 at Bizon [21, 27, 29].

Network cards are a B200 line and a B300 inclusion. Eight ConnectX-7 cards add $14,520 at Exxact and $17,688 at Thinkmate (arithmetic on [21, 27]). Exxact's B200 configurators accept up to eight, nine or ten network cards and Thinkmate's up to twelve [21, 22, 23, 27]. Bizon's B200 page offers one 400 Gb/s card [29]. On B300 systems the eight 800 Gb/s ports are part of the platform, and extra cards are a separate option [24, 28].

The GPU board has no price of its own. Every configurator reviewed shows the HGX baseboard as a selected line with no separate price [21-25, 27, 28]. Public listings therefore give no source for the share of a server price that the GPUs account for.

5. Bringing listings to a common build

The common builds follow the relays in Finding 2 and Dell's 1:1 network card recommendation [15, 27, 28].

• B200 common build: at least 1.5 TB of host memory, at least 48 cores per CPU, and eight ConnectX-7 400 Gb/s cards.
• B300 common build: 3,072 GB of host memory (above Dell's "greater than 2TB"), at least 48 cores per CPU, and eight 800 Gb/s ports on the platform.

Adjusted prices add the listed line items to the default price. Storage (Thinkmate's defaults include no drive), CPU models, memory speed and warranty still differ [27, 28].

Sources: line items in [22, 23, 24, 25, 27, 28, 29, 30].

The gap between sellers widens. Among the three B200 builds that reach the full common build, the default prices span $61,627. After adjustment they span $146,395 (arithmetic on [22, 23, 27]). Among the four B300 builds, the span grows from $167,841 to $180,987.60 (arithmetic on [24, 25, 28, 30]). Bizon's B300 server goes from below every other B300 default to above both Exxact builds [30].

B300 against B200 at one seller. At Exxact on 24 October 2026, the B300 build with 2,048 GB and eight 800 Gb/s ports listed at $516,036.40. The Intel-based B200 build with 2,048 GB and eight added 400 Gb/s cards came to $496,958, a difference of $19,078.40 (arithmetic on [23, 24]). The two builds use different CPU generations and storage. Bizon's B200 page describes its B300 model as "Same price as B200" [29].

The Exxact 10U B200 page, adjusted the same way, would come to $250,596.50 (arithmetic on [21]). It stays outside the range of every other listing.

6. Software, support, warranty, shipping and tax

Software. NVIDIA AI Enterprise "is licensed on a per-GPU basis", and a license is required for every GPU in a server that hosts its software [9]. List prices for self-managed systems are $4,500 per GPU for one year, $13,500 for three years and $18,000 for five years (five years for the price of four), or $22,500 per GPU for a perpetual license with five years of support [10]. Licenses come included with some earlier GPUs and Hopper-generation DGX systems, and NVIDIA states that "NVIDIA AI Enterprise licenses need to be purchased separately for NVIDIA DGX systems using the NVIDIA Blackwell architecture" [9, 10]. Thinkmate offers a five-year subscription per GPU at $17,999 and Exxact at $19,800, each as an added line [24, 27]. At list price, a five-year subscription for eight GPUs comes to $144,000 (arithmetic on [10]). None of the default builds reviewed includes it.

Warranty and support. Terms differ by maker and by seller.

Supermicro's parts warranty "is subject to availability in Supermicro stock", excludes components or software supplied by third parties, and gives way to special terms set out in an invoice [14]. Dell's standard warranty "does not include software support" [15]. A listing's warranty line gives the seller's terms, which can differ from the maker's standard.

Shipping, tax, tariffs and discounts. Thinkmate's prices "DO NOT INCLUDE SHIPPING AND/OR HANDLING CHARGES OR APPLICABLE TAXES" [27]. Exascale Labs' contract value is "exclusive of import duties and taxes" [18]. Vipera tells buyers to "Ask an agent if import tariffs apply" and requires end-user forms before it shares quotes [31]. Servermall's prices show VAT and include shipping across the EU, the UK and the USA [32]. Exxact and Bizon advertise education and government discounts without amounts [21, 29].

7. Per-GPU figures: server, cluster and deployed cost

A per-GPU figure depends on what is divided. Dividing a server listing by eight spreads the CPUs, memory, chassis, storage and warranty across the GPUs. On that basis, the B200 default builds come to $55,410 to $63,113 per GPU, and the B300 default builds to $62,080 to $83,060, with Exxact's 10U page at $29,510 and Vipera at $95,000 (arithmetic on [21-25, 27-31]).

Epoch AI's hardware price series works the same way, "including all costs of a DGX server node with eight GPUs (as of the GPU's release date), and dividing by eight" [34].

Fleet figures carry more than a server. Three filings show the range.

• Exascale Labs (contract to sell). On 9 October 2026 Exascale Labs entered into an agreement to "sell and provide 12 NVIDIA HGX B300 GPU servers (96 GPUs), together with network equipment and management servers, deployment services and 36 months of IT operation services" for a total contract value of "US$8,208,000, exclusive of import duties and taxes" [18]. Per server that is $684,000, and per GPU $85,500 (arithmetic on [18]). The filing gives no split between equipment and services.
• Axe Compute (purchase of a company holding a cluster). On 30 September 2026 Axe Compute acquired a company whose "assets consist of 288 Supermicro B300 servers (2,304 GPUs) and associated networking equipment" [19]. As consideration it "repaid in full the SPV's pre-existing debt of approximately $87.8 million" and agreed to a deferred purchase price of $715,000 per month for 60 months, which the filing says must be repaid within 12 months of closing or on an earlier financing secured by the cluster [19]. The debt repaid alone comes to about $304,900 per server, and 60 monthly payments would total $42.9 million (arithmetic on [19]). No single server price can be read from this deal.
• IREN (reported purchase). In July 2025 IREN announced the purchase of about 2,400 B200 and B300 GPUs "for ~$130m, including fit-out costs", which it defined as "servers, storage, labor and ancillary equipment such as InfiniBand, racks and cabling" [17]. The evidence classes note works through its per-GPU figures.

A cluster also needs parts that no server listing includes. NVIDIA's DGX SuperPOD reference architecture describes the system as "a combination of DGX systems, InfiniBand and Ethernet networking, management nodes, and storage" [8]. For four scalable units, it lists 128 nodes and 1,024 GPUs, 32 leaf and 16 spine InfiniBand switches, and 1,024 node-to-leaf and 1,024 leaf-to-spine cables [8]. A fleet per-GPU figure can carry all of these, and a server listing carries none.

Implications

For buyers checking a quote. Read the build before the price. In the listings reviewed, the total moves with host memory against the relayed floor, cluster network cards on B200 systems, CPU core counts, storage and the warranty [15, 21, 27, 28, 29]. Ask whether the price includes shipping, taxes, duties and software licenses, and how long it holds [18, 27, 31]. A starting price below the relayed memory floor can rise by $76,915 to $81,600 before network cards [27, 29].

For finance teams budgeting a purchase. B200 builds need network cards added, and NVIDIA AI Enterprise adds $36,000 a year at list price for eight GPUs (arithmetic on [10]). A cluster adds switches, cables, racks, storage and labor [8, 17, 18].

For analysts, lenders and price series. Each figure needs its date, build, currency and evidence class. Default prices from different sellers describe different machines, and the gap between sellers widened once listings were brought to a common build (Finding 5). Averaging listings, or mixing them with reported sales, describes no machine. Rillor's GPU systems pricing method covers system specifications, normalized configurations and price observations for HGX, NVL72 and Instinct systems. The 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.

Limits of this analysis

• Every price here is a listing seen on 24 October 2026 or a contract reported in a filing, never a verified single-server sale. Listings can change daily [27].
• The listings come from five sellers. Dell shows no price for the XE9780 [16], and DGX prices appear only at a European seller [32, 33].
• No source prices the HGX baseboard on its own, so this report gives no GPU share of a server price.
• The host memory floors come from two relays of NVIDIA's design guide, which disagree for B200. No public copy of the design guide was found.
• Adjusted prices in Finding 5 assume that line items add as shown. A quote may price combinations differently.
• Advertised discounts for education, government, volume or resale buyers have no stated amounts, and no price here reflects them [21, 29].
• NVIDIA's GPU memory figures differ between its documents [1, 2, 3, 4, 7]. This report quotes each as published.
• This report describes what public sources show. It is not investment, legal, tax or accounting advice.

Sources

1. NVIDIA HGX Platform. NVIDIA, viewed 24 October 2026. https://www.nvidia.com/en-us/data-center/hgx/
2. Datasheet for NVIDIA Blackwell Architecture. NVIDIA, October 2025. https://dam-cdn.nvd.orangelogic.com/AssetLink/y441155802qub41q118b2852i557jem5.pdf
3. Datasheet for NVIDIA Blackwell Ultra. NVIDIA, October 2025. https://dam-cdn.nvd.orangelogic.com/AssetLink/1k0p832eq8r5ca0u5383ie5o4tp3bst1.pdf
4. NVIDIA Blackwell Architecture Technical Brief. NVIDIA, viewed 24 October 2026. https://dam-cdn.nvd.orangelogic.com/AssetLink/gl2l4l4812s5fw0p614s6i8bv6mi3vx5.pdf
5. NVIDIA DGX B200. NVIDIA, viewed 24 October 2026. https://www.nvidia.com/en-us/data-center/dgx-b200/
6. NVIDIA DGX B300. NVIDIA, viewed 24 October 2026. https://www.nvidia.com/en-us/data-center/dgx-b300/
7. Introduction to the NVIDIA DGX B300 System, NVIDIA DGX B300 User Guide. NVIDIA, viewed 24 October 2026. https://docs.nvidia.com/dgx/dgxb300-user-guide/introduction-to-dgxb300.html
8. DGX SuperPOD Architecture, NVIDIA DGX SuperPOD Reference Architecture with DGX B200. NVIDIA, viewed 24 October 2026. https://docs.nvidia.com/dgx-superpod/reference-architecture-scalable-infrastructure-b200/latest/dgx-superpod-architecture.html
9. NVIDIA AI Enterprise Licensing, NVIDIA Enterprise Licensing Guide. NVIDIA, viewed 24 October 2026. https://docs.nvidia.com/ai-enterprise/planning-resource/licensing-guide/latest/licensing.html
10. NVIDIA AI Enterprise Pricing, NVIDIA Enterprise Licensing Guide. NVIDIA, viewed 24 October 2026. https://docs.nvidia.com/ai-enterprise/planning-resource/licensing-guide/latest/pricing.html
11. SYS-A21GE-NBRT, DP Intel 10U System with NVIDIA HGX B200 8-GPU, specifications. Supermicro, viewed 24 October 2026. https://www.supermicro.com/en/products/system/gpu/10u/sys-a21ge-nbrt
12. SYS-A22GA-NBRT, DP Intel 10U System with NVIDIA HGX B200 8-GPU, specifications. Supermicro, viewed 24 October 2026. https://www.supermicro.com/en/products/system/gpu/10u/sys-a22ga-nbrt
13. SYS-822GS-NB3RT, DP Intel air cooled GPU server with NVIDIA HGX B300 8-GPU, specifications. Supermicro, viewed 24 October 2026. https://www.supermicro.com/en/products/system/gpu/8u/sys-822gs-nb3rt
14. Supermicro Warranty (Standard Server Warranty). Supermicro, viewed 24 October 2026. https://www.supermicro.com/en/support/warranty
15. Dell PowerEdge XE9780 Technical Guide, Rev. A03. Dell Technologies, August 2026. https://www.delltechnologies.com/asset/en-us/products/servers/technical-support/poweredge-xe9780-technical-guide.pdf
16. PowerEdge XE9780 product page. Dell Technologies, viewed 24 October 2026. https://www.dell.com/en-us/shop/ipovw/poweredge-xe9780
17. IREN Expands AI Cloud with 2.4k NVIDIA Blackwell GPUs (Exhibit 99.1 to Form 8-K). IREN Limited, 3 July 2025. https://www.sec.gov/Archives/edgar/data/1878848/000114036125024747/ef20051566_ex99-1.htm
18. Exascale Labs Holdings Inc. Current Report on Form 8-K (purchase and sale agreement). U.S. Securities and Exchange Commission, filed 15 October 2026. https://www.sec.gov/Archives/edgar/data/2109869/000182912626011139/exascalelabs_8k.htm
19. Axe Compute Inc. Current Report on Form 8-K (Equity Purchase Agreement). U.S. Securities and Exchange Commission, filed 5 October 2026. https://www.sec.gov/Archives/edgar/data/1446159/000117184326006386/f8k_100526.htm
20. NVIDIA HGX B300 and B200 Servers (category page). Exxact Corporation, viewed 24 October 2026. https://www.exxactcorp.com/category/NVIDIA-HGX
21. Configure Exxact TensorEX 10U HGX B200 Server TS4-124903244. Exxact Corporation, viewed 24 October 2026. https://configurator.exxactcorp.com/configure/TS4-124903244
22. Configure Exxact TensorEX 8U HGX B200 Server TS4-142487900. Exxact Corporation, viewed 24 October 2026. https://configurator.exxactcorp.com/configure/TS4-142487900
23. Configure Exxact TensorEX 8U HGX B200 Server TS4-128275176. Exxact Corporation, viewed 24 October 2026. https://configurator.exxactcorp.com/configure/TS4-128275176
24. Configure Exxact TensorEX 8U HGX B300 Server TS4-131188947. Exxact Corporation, viewed 24 October 2026. https://configurator.exxactcorp.com/configure/TS4-131188947
25. Configure Exxact TensorEX 8U HGX B300 Server TS4-198437339. Exxact Corporation, viewed 24 October 2026. https://configurator.exxactcorp.com/configure/TS4-198437339
26. GPX GPU-Optimized Servers. Thinkmate, viewed 24 October 2026. https://www.thinkmate.com/systems/servers/gpx
27. GPX GXS6-2808-8HGX-B2 configurator (HGX B200). Thinkmate, viewed 24 October 2026. https://www.thinkmate.com/system/gpx-gxs6-2808-8hgx-b2
28. GPX GXS6-2808-8HGX-B3 configurator (HGX B300). Thinkmate, viewed 24 October 2026. https://www.thinkmate.com/system/gpx-gxs6-2808-8hgx-b3
29. BIZON X9000 G4, 8 GPU NVIDIA B200 server. Bizon, viewed 24 October 2026. https://bizon-tech.com/bizon-x9000-g4.html
30. BIZON X9000 G5, 8 GPU NVIDIA B300 server. Bizon, viewed 24 October 2026. https://bizon-tech.com/bizon-x9000-g5.html
31. Supermicro GPU SuperServer SYS-822GS-NB3RT HGX B300 8-GPU. Vipera, viewed 24 October 2026. https://viperatech.com/product/supermicro-gpu-superserver-sys-822gs-nb3rt-hgx-b300-8-gpu
32. NVIDIA B300 GPU Servers in Europe. Servermall, viewed 24 October 2026. https://servermall.com/sets/servers-with-b300/
33. NVIDIA B200 GPU Servers in Europe. Servermall, viewed 24 October 2026. https://servermall.com/sets/servers-with-b200/
34. Performance per dollar improves around 30% each year (Robi Rahman). Epoch AI, 23 October 2024. https://epoch.ai/data-insights/price-performance-hardware
35. AI Server Demand Sustains Memory Contract Price Increases in 4Q26, While Consumer-Side Pressure Persists, Says TrendForce. TrendForce, 30 September 2026. https://www.trendforce.com/presscenter/news/20260930-13258.html