Suppliers and how buying works

UpStation does not resell: it publishes supplier data, qualifies the buyer and introduces the supplier that fits; the buyer transacts directly. UpStation is paid by suppliers through referral and partner programs where they exist; it adds no markup and takes no part in the transaction. The prices shown are the suppliers' own. Prices in USD as of 2026-10-05; Supermicro lines at eStore B2B net (12% off list, list in brackets); the supplier confirms the final price on order. Continental US; Canada quoted case by case. Ready to buy: book a meeting and UpStation introduces the supplier (how buying works). Every page: llms.txt (map) · homepage (all tables) · catalog.json (data).

How buying works (in order)

• Design and price with your own AI assistant from this catalog (llms.txt or catalog.json). The reference builds and worked designs are computed from the same data, so the arithmetic can be checked.
• When the design is settled, book a meeting or email. UpStation qualifies the requirement in one conversation: end use and end user, ship-to and country, confidential-computing or other platform needs, timeline, budget and payment capacity.
• UpStation introduces the supplier that fits (OEM direct, the Supermicro eStore, a distributor desk, or a system builder), pre-stages the NVIDIA end-user and compliance paperwork every GPU order needs, and stays on the thread until the order is accepted.
• The buyer transacts directly with the supplier at the supplier's price and terms. UpStation adds no markup; where a supplier runs a referral or partner program, UpStation is compensated by the supplier.

UpStation is paid by suppliers through referral and partner programs where they exist; it adds no markup and takes no part in the transaction. The prices shown are the suppliers' own. Book a meeting when a design is ready.

Two kinds of price on this site

• Listed exactly. Listed exactly: the supplier's own public price book (Supermicro eStore, Octoserver, FS.com), re-read monthly.
• Pricing under NDA. The site shows a $25,000 range that contains the quoted price, not the figure. The exact, current quote comes with UpStation's introduction.
• Suppliers with private pricing are the high-value introductions: they quote once a buyer is qualified, and UpStation's role is to bring them that buyer with the paperwork ready.

Suppliers UpStation introduces

NVIDIA Inception hardware discounts (members only; valid 2026-07-27 to 2026-10-25)

Who qualifies: Startups under 10 years old building their own product; not resellers or distributors, cloud service providers, consultancies, crypto companies or public companies. Free to join; approval takes days.

• Requested in the Inception portal and applied by the seller: the server must come from an NVIDIA Partner Network solution provider or an NVIDIA OEM, which quotes the final price (the OEM needs the member's approval number; cost at purchase varies by vendor).
• Hardware bought at member pricing is for the member's internal use (demonstration, evaluation, development, testing, proof of concept; NVIDIA's terms say not production) and may not be resold, transferred or bundled into a commercial offering.
• Caps are lifetime per GPU type: 64 HGX GPUs of one type across x1/x4/x8 systems (H200 HGX counts toward H200, not toward B200); 64 RTX PRO or L40S cards.
• Not available in NVIDIA's excluded countries (China, Hong Kong, Macau, Russia and others). Guidance is revised quarterly; the current table is dated 2026-07-27 to 2026-10-25.

Source: NVIDIA Inception Discount Guidance: NVIDIA Hardware (member portal document, read 2026-10-05).

Market notes (dated; what suppliers said and what the price books show)

• 2026-10-05: Supermicro October price book: Gold Series up 9-13% (4U 4x RTX PRO 6000 Gold $114,440 to $127,673; 2U 2x RTX PRO 6000 Gold $60,683 to $66,680; 5U AMD 2x RTX PRO 6000 Gold $95,755 to $104,670; 2U 24-bay HDD Gold $51,107 to $57,873; 1U Hyper AMD $66,636 to $72,769). UpStation's saved 8x RTX PRO 6000 configuration $191,888 to $196,626 (+2.5%; the 64GB DIMM is now $2,530). B300 Gold $565,384 (Sep 3) to $572,364 (Sep 8) to $568,110 (Oct), and out of stock on the eStore (notify-me). H200 default configuration $319,422 to $323,603 (+1.3%). Configure-to-order chassis base prices moved +0.7% to +4.9% (SYS-422GA-NRT $25,368 to $25,850; AS-5126GS-TNRT $46,168 to $46,495 with 2 GPUs; SYS-521GE-TNRT $48,699 to $49,988 with 2 GPUs; ASG-2015S-E1CR24H $13,462 to $14,116), so the September configurator reads for hosts and nodes are kept with that error band until re-probed.
• 2026-09-22: HGX B200 is the fastest high-end route a system builder could quote in late September: a configured 8x B200 Intel system (2x Xeon Gold 6548N, 1.5TB, 5.8TB NVMe, no fabric NICs) in the $375,000-$400,000 band plus freight on a prepay-wire, NCNR, two-week-validity quote, 11-12 weeks (8 for the ASRock Rack barebone, 3-4 to integrate and test), versus 30 weeks for the same builder's HGX B300 system (2TB, 8x 800G CX-8 on-board) in the $475,000-$500,000 band. B200 is being EOL'd, so treat it as a one-time buy; the Supermicro eStore no longer lists its B200 Gold SKUs. Exact quotes come with an introduction.
• 2026-10-01: HGX B300 8-GPU nodes with 2-3 TB host memory were quoted between about $490k and $615k per node across OEM-direct, distributor and reseller channels in September-October 2026; the spread is set by CPU brand, memory, warranty and how soon the box ships. One OEM could ship Intel from buffer stock before January 2027 and AMD from February; a reseller offered AMD boxes from a cancelled order at 4-6 weeks after Supermicro compliance approval, at the top of the range.
• 2026-09-22: NVIDIA raised DGX pricing on 2026-09-08, the first across-the-board increase one system builder had seen; memory and SSD prices have risen since December 2025 and suppliers expect shortages into 2027. Quotes now carry one-week validity and subject-to-change language, and some suppliers call B300 lead times unguaranteeable.
• 2026-09-30: Lead-time bands for an 8x B300 node: ready-to-ship stock when listed (comes and goes; 1-3 weeks including export review); OEM buffer inventory 8-12 weeks after PO, deposit and signed sales agreement; configure-to-order from NVIDIA allocation 30+ weeks. Payment norms: OEM direct 25% on an NCNR PO and 75% before ship; system builders 50/50; eStore credit card or Net-30; distributors net terms or wire.
• 2026-10-02: Compliance: every NVIDIA GPU server order needs an end-user statement. OEMs run an 8-point review (purchaser, end user, ship-to, headquarters, parent, payer, destination, installation and end-use scenario); distributors will not quote without the end user's name, address and website; the eStore routes GPU orders through export review. Allow a few days and have the ship-to address settled.
• 2026-10-05: NVIDIA Inception: startups under 10 years old that build their own product can request member pricing on NVIDIA hardware once approved; NVIDIA's current guidance (table on this page) is $60,000 off an HGX B300 node, $50,000 off an HGX B200 node, $44,000 off an H200 x8 node and $1,400 off each RTX PRO 6000 or L40S card, with lifetime caps. Apply before the PO: the OEM needs the approval number, and the hardware is for the member's own use, not resale.
• 2026-09-24: Confidential computing: NVIDIA CC is supported on HGX B300; RTX PRO 6000 with Intel TDX runs on 5th Gen Xeon or Xeon 6 (one GPU per confidential VM); HGX B200 with Intel Xeon was still being confirmed by one builder. BIOS versions differ by builder, so ask for TDX/CC pre-enabled and the BIOS version stated on the quote.
• 2026-10-01: Supply alternatives: RTX PRO 6000 SE accessory cards are out of stock at Supermicro (the Gold-plus-loose-cards route is unavailable); Octoserver's passive card is $14,800 in stock. Supermicro's eStore now lists RTX PRO 4500 Blackwell SE entry servers (1U/2U, 1-2 GPUs). H200 NVL PCIe cards carry 35-40 week lead times.
• 2026-10-05: Consumer RTX 5090 in a server: retail 5090s are 3.5-4-slot active cards that do not fit a conventional 4U GPU chassis; the routes are a custom passive card (Octoserver, $5,200, 30 days) or a tall chassis with cable risers and 4-slot trays (OneChassis 7U, $3,000-3,500 before the board). Retail 5090 street prices (median about $6,000, range $5,200-7,450 on 2026-10-05; MSRP $1,999) now sit above Octoserver's passive card, so the DIY saving is in the host and the parts, not the cards: see the DIY worked design. GeForce drivers exclude data-center deployment; buyers treat these as lab or workstation nodes.
• 2026-10-05: Parts market read for build-your-own hosts (eBay sold medians, US): EPYC 7543 32C $595, 7742 64C $680, 7532 32C $318 (SP3); EPYC 9354 32C $1,695, 9654 96C $2,600 (SP5, no second-hand discount yet); 32GB DDR4-3200 RDIMM $240 ($7.50/GB) against 32GB DDR5-4800 RDIMM $800 ($25/GB, DDR5 scarce); PM9A3 3.84TB U.2 $533 median, $1,000-1,100 new. RTX PRO 6000 Blackwell: NVIDIA list $16,000 (from $8,565 at launch), street $15,000-18,000, Server Edition $13,000-14,900. Table on the rate-card page.