NVIDIA Server GPU 2026 Buying Guide: Specs, Prices, Compliance & Channels

Product Line Classification

NVIDIA's compute GPUs fall into four distinct product lines with vastly different export-control statuses: data-center accelerators (H100/H200/B200/B300/GB systems) — mostly banned from China; China-specific compliant cards (H20/B30) — legally purchasable domestically; professional workstation cards (RTX PRO 4000–6000 series) — unrestricted, available on JD.com; and consumer gaming cards (RTX 5090 etc.) — also unrestricted and widely sold.

Data-Center Flagships: Performance Ceiling, Mostly Unobtainable

H100 (Hopper, 2022)

The "engine of the LLM era" used to train GPT-4, Llama, DeepSeek. 80 GB HBM3, 3.35 TB/s bandwidth, ~3,958 TFLOPS FP8 (sparse), 700 W. Overseas reference price: $25k–40k per card; 8-card system ~$200k+. Banned in China; gray-market units carry customs seizure risk, no warranty, and violate EAR 3A090.

H200 (Hopper refresh, 2024)

Same compute core as H100, but VRAM doubled to 141 GB HBM3e and bandwidth raised to 4.8 TB/s. Identical compute, yet the extra memory significantly benefits long-context inference and large-model training — a clear case of "memory is justice." Overseas price ~$30k–40k. Since Jan 2026, status changed from "presumption of denial" to "case-by-case review," but strict conditions (third-party testing, purchase volume ≤50% of US customer, 25% revenue remittance) mean no Chinese buyer has completed a substantive purchase as of mid-2026.

B200 (Blackwell, 2024)

New architecture with native FP4 support. 192 GB HBM3e, 8.0 TB/s bandwidth, ~10 PFLOPS FP4 dense, 1,000 W. Inference performance jumps orders of magnitude over H100. Overseas price ~$40k–45k; Chinese gray-market quotes ¥350k–600k+ (heavy premium). Banned for China.

B300 / Blackwell Ultra (2025)

Current single-card king: 288 GB HBM3e, FP4 sparse ~20 PFLOPS, up to 1,400 W. One card can hold a 200B+ dense model (FP8); an 8-card node runs trillion-parameter MoE. SemiAnalysis benchmarks show ~35× lower per-token inference cost vs. Hopper generation. Sold mainly as 8-card systems (~$550k overseas, ~¥7M domestically); single cards not retailed. Banned for China.

GB200 / GB300 NVL72 (Rack-Scale)

72 Blackwell GPUs linked via NVLink 5 into a single coherent domain. GB200 NVL72 aggregates 576 TB/s bandwidth, ~120 kW per rack; GB300 pushes total VRAM to 20.7 TB. This is the answer for frontier model training, but entire racks are banned for China, and procurement means buying a full liquid-cooled data-center suite, not just chips.

China-Compliant Cards: Only Two Legal Data-Center Options

H20 (Hopper China-Special, 2024)

Designed to circumvent export controls: compute slashed to ~1/6 of H100 (FP16 ~148 TFLOPS) while retaining 96 GB VRAM and NVLink high bandwidth. Explicitly meant for inference, not efficient training. Specs: 96 GB HBM3 / 4.0 TB/s / 400 W; a 141 GB variant exists. Pricing (single card): 2026 Q1 ~¥90k–100k, Q2 rose to ¥100k–120k; 141 GB system ~¥2.86M–3.0M, 96 GB system ~¥1.9M. Status: ✅ sellable (license granted); sales resumed July 2025, ~400k units shipped in H2 2025.

B30 (Blackwell China-Special, 2025 Q4)

Based on Blackwell but uses GDDR7 to meet compliance. Compute ~75–90% of H20, yet price 30–40% lower and energy efficiency ~30% higher. Tailored for domestic inference market. Price: ~$6.5k–8k per card, domestic ~¥50k–60k; 8-card system ~¥1.4M–1.6M. Status: ✅ sellable (designed for compliance, not subject to advanced-compute controls).

One-line summary: For compliant compute buy H20 (large VRAM, mature ecosystem); for cost-efficiency buy B30 (30%+ cheaper). Both only suit inference and light fine-tuning — do not expect efficient frontier-model training.

Unrestricted Workstation & Consumer Cards: Buy Direct on JD.com

RTX PRO Blackwell Family (Professional Workstation, 2025)

Same Blackwell architecture as B200, supports FP4, ECC VRAM, PCIe 5.0, full CUDA compatibility.

RTX PRO 6000: 96 GB GDDR7, 600 W, desktop flagship, ~¥60k–150k. "Desktop B200" — single card runs 70B–100B quantized models for inference/fine-tuning without rack or data-center power.

RTX PRO 6000 Max-Q: 96 GB, 300 W, low-power variant, same core, enables 4 cards in one workstation.

RTX PRO 5000: 72 GB, 300 W, balanced, ~¥59k (JD.com self-operated) — called "value king."

RTX PRO 4500: 32 GB, 200 W, mid-range, for local fine-tuning / 8K post-production.

RTX PRO 4000: 24 GB, 140 W, entry, ~¥14k (Leadtek self-operated), compact workstation.

Server variants use passive cooling for rack deployment, suitable for small AI servers.

RTX 50 Consumer Series (2025)

RTX 5090: 32 GB GDDR7, ~3,300 AI TOPS, 575 W, domestic ¥28k–34k. Top pick for individual developers/small teams.

RTX 5080: 16 GB, ~¥8k–10k.

RTX 5070: 12 GB, starting ¥4,599.

Local AI developer reality: A single RTX 5090 (¥30k) or RTX PRO 5000 (72 GB, ¥60k) happily runs 70B quantized models; for 100B+ or higher throughput, consider H20/B30 systems or cloud rental.

Comparison Summary (Key Specs & Pricing)

The article provides a detailed comparison table; key takeaways:

H100 SXM: Hopper, 80 GB HBM3, 3.35 TB/s, 989 TFLOPS FP16 sparse, 700 W, $25k–40k, banned.

H200 SXM: Hopper, 141 GB HBM3e, 4.8 TB/s, same compute, 700 W, $30k–40k, case-by-case review.

B200 SXM: Blackwell, 192 GB HBM3e, 8.0 TB/s, 2,250 TFLOPS FP16, 1,000 W, $40k–45k, banned (gray ¥350k–600k).

B300 SXM: Blackwell Ultra, 288 GB HBM3e, 8.0 TB/s, 2,250 TFLOPS, ≤1,400 W, system $550k (~¥7M), banned.

H20: Hopper, 96 GB HBM3, 4.0 TB/s, ~148 TFLOPS, 400 W, ~$10k–12k, domestic ¥90k–120k, ✅ legal.

B30: Blackwell, ~180 GB GDDR7, bandwidth below B300, 75–90% H20 compute, ~700 W, $6.5k–8k, domestic ¥50k–60k, ✅ legal.

RTX PRO 6000: Blackwell, 96 GB GDDR7, 1.79 TB/s, 4,000 TOPS FP4, 600 W, $13k–15k, domestic ¥60k–150k, ✅ unrestricted.

RTX PRO 5000: Blackwell, 72 GB GDDR7, 1.34 TB/s, 2,064 TOPS, 300 W, ~¥59k, ✅ unrestricted.

RTX 5090: Blackwell, 32 GB GDDR7, ~3,300 TOPS, 575 W, domestic ¥28k–34k, ✅ unrestricted.

Rental Market: Renting Often Saves 30–45%

For teams using <50 cards/year, with volatile demand or in pilot phase, cloud rental is cheaper. 2026 rental prices trending up globally but flexibility remains valuable. Sample rates (source: Jisuan, Business New Knowledge GPU pricing, Jul–Sep 2026):

RTX 4090 24 GB: ¥1.96/hr, ¥7,920/mo (8-card) — inference value king.

RTX 4090 48 GB: ¥9,600/mo (8-card) — large VRAM inference.

H20 96 GB: ¥6.20/hr, ¥29,200/mo (8-card) — mainstream domestic adaptation.

H100 80 GB: ¥84,480/mo (8-card, includes NVLink+IB).

Ascend 910B 64 GB: ¥3.58/hr, ¥15,600–22,000/mo — domestic alternative.

B300/3 TB: ¥280k–350k/mo — high-end long-term contract.

Prices fluctuate rapidly with policy and supply; verify real-time quotes before ordering.

Purchasing Channels: Where to Buy Safely

Compliant Data-Center Cards (H20/B30) — Authorized Chain Only

Core authorized distributors: Unisplendour Xintong (Unisplendour, China general distributor for data-center GPUs), CEC Port (official authorized distributor), Digital China (long-standing general distributor), Leadtek (workstation/professional card leader).

Specialized compliant vendors: Hongxin Electronics (Antong, full H20 sales license), Inspur Information, H3C, Nettrix, Huaqin Technology (ODM, mass-producing H20 servers).

Cloud vendors: Alibaba Cloud, Tencent Cloud, Volcano Engine all offer H20 instances and system procurement channels.

Workstation / Consumer Cards — JD.com Self-Operated

RTX PRO 4000/5000/6000, RTX 5090/5080 are uncontrolled; JD.com self-operated and Leadtek official store have ample stock, transparent pricing, no authorization chain needed. RTX PRO 4000 ~¥14k (Leadtek), RTX PRO 5000 ~¥59k.

Cloud Compute Rental — Elastic First Choice

Public clouds: Alibaba Cloud, Tencent Cloud, Volcano Engine, Huawei Cloud (incl. Ascend), Baidu AI Cloud (PAI).

Aggregator/developer platforms: Jisuan (jygpu), AutoDL (per-second billing, rich images), Parallel Tech Paratera (A800/H800 + 100G IB, research training), Featurize (small-scale experiments).

Overseas platforms (require compliance assessment): Lambda Labs, CoreWeave, RunPod, Vast.ai, Azure/Google Cloud.

Gray-Market "Water Goods" — Absolutely Avoid

So-called "water goods" H100/B200 enter via gray channels. Triple risk: customs seizure, no warranty, legal violation of US export controls (EAR 3A090). Since May 2026, even the loophole of Chinese subsidiaries buying via third countries has been closed by BIS. Enterprise procurement must use compliant authorized channels and consult professional compliance advisors.

Decision Framework: Which Card Should You Buy?

Frontier LLM training (70B+ from scratch / large-scale fine-tuning) → Deploy H100/H200/B200 in overseas data centers (if compliant) OR directly adopt domestic Ascend 910C/910D (government/finance bid winner, 8-card ~¥1.1M–1.2M).

Domestic LLM inference / deployment (compliance mandatory) → Budget ample & ecosystem maturity → H20 (¥90k–120k, 96 GB VRAM); Cost-priority → B30 (¥50k–60k, 30%+ cheaper).

Local / small-team AI dev, 70B quantized inference → Extreme value → RTX 5090 (¥30k, 32 GB); Need more VRAM → RTX PRO 5000 (72 GB, ¥60k) / RTX PRO 6000 (96 GB, ¥60k–150k).

Volatile demand, <50 cards/year → Rent cloud compute (Jisuan/AutoDL/Alibaba Cloud), save 30–45%.

Key Trends Summary

Bans normalized, special editions are reality: H100/H200/B200/B300 face "presumption of denial"; only H20/B30 are legal data-center cards in China. Controls upgrade almost monthly (case-by-case review, overseas subsidiary loophole closed, global tiered licensing brewing) — check latest BIS notices before purchasing.

"Memory is justice": Generational leaps focus on VRAM and bandwidth (80 GB → 288 GB) — this dictates whether a model fits and token throughput. H200's unchanged compute but doubled memory delivering significant gains is the best proof.

Local AI democratization: RTX PRO 6000 (96 GB) and RTX 5090 let individual developers run 70B quantized models for tens of thousands of yuan, no rack or data-center needed — the most impactful 2026 shift for ordinary developers.

Rental prices rising in unison: Domestic and overseas rental rates climb together (multiple contracts up 13–200%), "having cards" revalues assets. Yet for short-term elastic needs, renting still beats buying.

Domestic substitution accelerating: Huang Renxun admits US export limits are handing China's market to Chinese firms. Ascend 910C/910D now volume-shipping to government/enterprise/telcos; 2026 capacity stabilizing, becoming the top choice for indigenization.

Data currency: Covers public market conditions Feb–Sep 2026, sourced from NVIDIA official, SMM/Mysteel daily compute reports, Jisuan, Business New Knowledge GPU pricing column, and industry analysis. GPU prices and control policies change extremely fast; prices and statuses here are for reference only — major decisions should rely on latest official announcements and compliance counsel.