- Same 8-GPU DGX B300 / NVL8-class chassis: ~USD 550k factory-class vs ~USD 600k authorized landed vs ~RMB 14.0M regional spot (pending confirmation).
- 3.24× = 14.00 / 4.32 at the source FX; the same ratio appears on the lease track.
- Fit: freeze one corridor. Unfit: mix USD factory with CNY spot, or quote NVL72 against NVL8.
What is the B300 NVL8 quote map?
Quick Answer: The DGX B300 NVL8 quote map is a same-SKU view of one 8-GPU Blackwell Ultra chassis across factory-class, authorized-landed, regional spot, and lease tags. NVIDIA publishes no list. NVL8 is not GB300 NVL72. Soft figures from the 2026-09-05 window are industry estimate / pending confirmation. Fit teams freeze one corridor; unfit teams blend USD factory with CNY spot on one PO.
The DGX B300 NVL8 quote map is a channel-observation equal-SKU sheet: one NVIDIA DGX B300 8-GPU system (NVL8-class whole-system unit) compared across factory-class, authorized landed, regional / non-authorized spot, and lease. It is not a vendor InfoFlash and not an NVIDIA Global Price List. The DGX B300 product page routes buyers to partners; public US/EU reseller bands and this 2026-09-05 map remain industry estimate / pending confirmation.
Form-factor lock: DGX B300 ships 8× Blackwell Ultra SXM. Source shorthand NVL8 means that 8-GPU quote unit. It is not the same SKU as GB300 NVL72 (72 GPU + 36 Vera CPU rack). Mixing NVL8 stickers with NVL72 or Vera Rubin NVL72 stickers on one RFQ is a classification error, not a discount. Typical class specs (pending datasheet): about 288 GB HBM3e per GPU, total GPU memory about 2.1–2.3 TB, NVLink Switch System (2×), ConnectX-8 VPI, BlueField-3 DPU, dual Intel Xeon Platinum 6776P, direct liquid cooling (DLC). CDU / RDHx are facility BOM, not the GPU sticker.
China enterprise use of this page is 政企采购核价: compare corridors, then freeze title / ATP / serial path on a dated RFQ. Export-control language here is buyer risk: DGX B300 is an export-controlled class, so access gates cut some corridors and turn remaining spot into an access premium. This is not a how-to for embargoed SKUs.
- Fit: procurement and finance teams comparing the same 8-GPU DGX B300 / NVL8-class SKU, then freezing one corridor (Incoterms, ATP date, title, serials, DLC/CDU lines).
- Unfit: mixing USD factory-class with CNY gray-spot on one PO; quoting NVL72 against NVL8; treating corridor figures as NVIDIA list; any request to divert or falsify end-user / license paperwork.
- Track 1 anchors: NVIDIA DGX B300, HGX B300, Blackwell Ultra, NVL8, ATP, DLC.
- Track 2 asks: why DGX B300 quotes differ; how to buy B300 8-GPU; authorized vs regional spot; lease vs buy.
Third-party list-class bands conflict on purpose. PCMag citing Financial Times (via GPUSmith) has put US retail DGX B300 near USD 400,000. Older Tech-Insider / iFactory-class roundups still quote USD 300,000–350,000 into September 2026. EU reseller configured systems have been reported at €535,000–€714,000. The 2026-09-05 channel map uses ~USD 550,000 factory-class and ~USD 600,000 authorized landed. Show the conflict; do not pick one as the official price. Adjacent headlines (cloud GB300 NVL72, DRAM forecasts, earnings prints) do not belong on the NVL8 sticker.
Buy tags and lease tags (same 8-GPU unit)
Quick Answer: Four buy tags and three lease tags sit on the same 8-GPU unit. Factory-class about USD 550,000, authorized landed about USD 600,000 (RMB 4.32 million at the source FX), regional spot about RMB 14.0 million. NA long lease about USD 12,500 per month versus regional all-in about RMB 290,000-plus. All pending confirmation.
Observation date ~2026-09-05. FX used by the source for the 3.24×: USD 600,000 ≈ RMB 4.32 million. Do not silently restate FX. All rows below are industry estimate / pending confirmation.
| Buy tag | USD / local | RMB (source) | Who pays |
|---|---|---|---|
| Factory-class / OEM-adjacent | ~USD 550,000 | ~RMB 4.00M | Authorized NA / cloud |
| Authorized landed (NA) | ~USD 600,000 | ~RMB 4.32M | Compliant enterprise PO — 3.24× baseline |
| Regional spot late-2025 | — | ~RMB 4.00M | Early channel; spread not yet open |
| Regional / non-authorized spot summer-2026 | — | ~RMB 14.00M (peak ~14.50M) | Urgent large buyers — 14.00 / 4.32 ≈ 3.24× |
Lease track uses the source formula (not NVIDIA): system price × 1.25 / 60 months.
| Lease tag | /month | Source math | Who pays |
|---|---|---|---|
| NA bare-metal long lease | ~USD 12,500 (~RMB 90k at source FX) | 600,000 × 1.25 / 60 | Overseas compliant tenant |
| Regional 5-year long lease (all-in cited) | ~RMB 290,000+ | 14,000,000 × 1.25 / 60 = 291,700 | Budgeted regional tenant |
| Regional short/spot rent | ~RMB 320,000–350,000 | quoted band | Fragment / burst demand |
3.24× appears twice: buy (14.00M / 4.32M) and lease (29.17 / 9.0) because both use the same ratio. A China RFQ that pastes factory-class USD into a CNY spot column is mixing two title stories, not finding a midpoint.
Why do quotes spread 3.24×?
Quick Answer: The 3.24× is RMB 14.0 million divided by RMB 4.32 million, not an NVIDIA multiple. Four layers explain the gap: export-control access premium, channel hops, no public list, and calendar time after ATP locks. The same ratio appears on the lease track because both use the same system stickers.
Do not fake percentage splits. The thickest qualitative layer for this class is access: export-control gates cut some corridors, so remaining spot behaves like an access premium. Channel hops add spread when whitelist tightening forces extra hands. The information gap is structural because NVIDIA has no public DGX B300 list — buyers who cannot see authorized landed pay the told-you-so quote. Time is calendar interest after large buyers lock ATP.
| Layer | What it does on an equal-SKU sheet | China RFQ action |
|---|---|---|
| Export-control / access premium | Thickest layer; remaining spot is not street NVIDIA | Title / license / end-user check — not circumvention steps |
| Channel hops | Each extra hop adds spread | Ask which hop holds ATP; see the on-site diversion-case twin for public-case hygiene |
| Information gap | No public NVIDIA list | Cite date + corridor name; never paste a channel tag as MSRP |
| Time | Waiters pay calendar interest | Freeze ATP week on the cover before finance averages stickers |
US/EU published bands (USD 300k–400k class, €535k–€714k configured) sit next to this map as a conflict, not as a replacement numerator. If a broker uses the USD 400,000 FT-class print as list and then adds a CNY 14 million spot, you still have two corridors. Freeze columns before you freeze CapEx.
Why some regional spots stayed elevated
Quick Answer: Regional spots stayed elevated because hyperscale training cannot wait, mid-size buyers shift to lease, application teams buy cloud hours, and tail burst pays short rent. Elevated tags are a demand structure, not a license to shop restricted corridors. China RFQs should treat title and ATP as the binding rows.
In the source window, regional spots did not mean-revert to the late-2025 ~RMB 4.00 million tag. Demand structure, not a secret MSRP, explains persistence. Typical moves below are not advice to break controls.
| Buyer | Situation | Typical move |
|---|---|---|
| Hyperscale / large training | Cannot wait | Pays certainty (authorized or, in some regions, elevated spot) |
| Mid-size | Cash-constrained | Lease; some compare NA-class long lease vs regional all-in |
| App companies | Avoid heavy CapEx | On-demand cloud (flexible, often dearest per GPU-hour) |
| Tail / burst | Fragment hours | Short rent RMB 320,000–350,000 / month class |
A China enterprise sheet that treats elevated CNY spot as the 2026 NVIDIA price will mis-size both CapEx and export-control residual. Authorized landed at ~USD 600,000 remains the 3.24× denominator in this map; regional ~RMB 14.0 million is the numerator. Keep them in separate columns with separate title language.
Is USD 12.5k/mo cheap versus RMB 290k+?
Quick Answer: USD 12,500 per month is not official NVIDIA rent. It is the source heuristic of 600,000 times 1.25 divided by 60 months for NA-class bare-metal long lease. Regional all-in near RMB 291,700 uses the same formula on the 14 million sticker. Compare logics, do not average them.
Two pricing logics sit on one formula. NA-class ~USD 12,500 / month is a cost-plus lease on the authorized-landed sticker. Regional ~RMB 290,000+ / month is a scarcity lease on the elevated spot sticker. The 3.24× reappears because 291,700 / 90,000 ≈ 3.24 at the source FX. That is a diagnostic, not a shopping list for restricted corridors.
Cash-constrained inference teams in China often want a 12–24 month lease versus buy on the same 8-GPU unit. Model both columns with identical GPU count, DLC yes/no, and support years. Then ask which corridor can actually deliver title. A cheaper USD rent that cannot clear ATP is not cheaper. Confirm with a dated lease RFQ; do not treat USD 12,500 as official B300 rent.
- Buy track 3.24×: 14.00M / 4.32M
- Lease track 3.24×: 291,700 / 90,000 (source FX)
- Short rent RMB 320,000–350,000 / month is a third logic (burst), not a fourth NVIDIA SKU
Real-world deployment scenarios
Quick Answer: Three China-relevant pictures: authorized 8-GPU pod this quarter, cash-constrained inference lease, and a multi-region CFO slide showing 3.24×. Each freeze SKU, corridor, and ATP date. Never blend USD 550,000 with RMB 14 million into one CapEx number.
Map operational constraints to these three enterprise pictures. They are decision trees, not delivery promises.
A private training pod must land this quarter with DLC already funded. Compare factory-class vs authorized landed only. Add CDU / RDHx and 800 Gb/s fabric lines. Refuse CNY regional spot mixed into the USD column. Title and serial path are the go-live gate, not a cheaper sticker.
Model 12–24 month lease vs buy on the same 8-GPU unit. NA-class ~USD 12.5k/mo vs regional all-in ~RMB 290k+ is a spread diagnostic. It is not a license to shop restricted corridors. If ATP cannot be shown, the lease column is incomplete.
Finance sees 3.24× on a slide and wants a blended budget. Action: freeze SKU (DGX vs HGX vs NVL72), freeze corridor, freeze ATP date, then RFQ. Do not average USD 550,000 and RMB 14 million into one CapEx line. Two columns beat one fake midpoint.
How to freeze a DGX B300 8-GPU (NVL8-class) quote without mixing price corridors (2026)
Quick Answer: Lock SKU (DGX versus HGX versus NVL72), lock one corridor, lock Incoterms plus title and serials, add CDU and fabric lines, then send an equal-BOM note to info@021taohan.com. Soft corridor figures are not nets. ATP proof beats a cheaper sticker without title.
- Step 1 — Lock the SKU. Write DGX B300 8-GPU versus HGX B300 versus GB300 NVL72. Record GPU count, HBM class, and DLC yes/no. Do not mix Rubin/NVL72 stickers into a B300 NVL8 RFQ.
- Step 2 — Lock the corridor. Factory-class / authorized landed / regional spot / lease. One corridor per RFQ column. Never blend USD factory with CNY spot on one line.
- Step 3 — Lock commercial terms. Incoterms, ATP or warehouse proof, title, serials, support years. Soft figures are not nets. Pending confirmation stays on the cover.
- Step 4 — Add facility lines. CDU, RDHx, 800 Gb/s fabric, power. Sticker is not a landed AI factory. DLC typical for dense B300.
- Step 5 — Submit equal-BOM RFQ. Email the equal-SKU sheet to info@021taohan.com for TAOHAN review. Ask for a dated net, not a reprint of this map.
If the pack still mixes corridors after Step 2, stop. A second column is cheaper than a PO that cannot clear title.
Frequently Asked Questions
Quick Answer: Four buyer questions: no NVIDIA list, what 3.24× means, why not budget USD 550,000 against an RMB 14 million broker line, and whether USD 12,500 is official rent. Answers start with the conclusion, then the 2026-09-05 map labeled pending confirmation.
Does NVIDIA publish a DGX B300 list price?
What is the 3.24× on this page?
Can I budget USD 550k if a broker quotes RMB 14M?
Is USD 12,500 per month the official B300 rent?
References
- NVIDIA — DGX B300 product page
- NVIDIA — DGX B300 User Guide (introduction)
- GPUSmith — HGX/DGX B300 price explainer (third-party bands)
- Tech-Insider — Blackwell GPU pricing roundup (industry reporting)
Key Takeaway: Freeze SKU (DGX B300 8-GPU / NVL8, not NVL72), freeze one corridor, freeze ATP and title, then RFQ. 3.24× is 14.00M ÷ 4.32M at the source FX, repeated on lease. Soft tags stay pending confirmation. TAOHAN supplies NVIDIA DGX B300 / HGX B300 8-GPU class systems with competitive pricing and BOM risk review for China enterprise equal-SKU sheets.
Related reading
Freeze one DGX B300 corridor before CapEx
If your sheet still mixes USD factory-class with CNY regional spot, or NVL8 with NVL72, send the equal-SKU BOM to TAOHAN. We help confirm title, ATP date, and facility lines before finance ranks nets. No public street prices on this site.
Email info@021taohan.comEmail: info@021taohan.com