GPU价格 — H100, B200 and AI accelerator buy & rental rates
Unit pricing for the accelerators themselves, plus what neoclouds and hyperscalers actually charge to rent them per GPU-hour.
快速回答： 截至 2026-10-14, an H100 rents for a $3.52/GPU-hour advertised median across 41 providers — up ~17.8% year over year — while the module itself changes hands near $25,000.00 on the secondary market. Blackwell B200 rents at $6.69/GPU-hour on-demand at 拉姆达.
There is no exchange for GPUs and no public list price for datacenter parts. Unit levels come from channel and secondary-market reporting; rental levels come from provider rate cards and cross-provider trackers. Every row is flagged 报告值, 计算值 or 估计值 and links to its source. The memory bolted to these parts is tracked on the HBM 价格 page.
最后更新
H100 rental price — advertised on-demand median
Weekly median $/GPU-hour across ~40 providers (GetDeploying dataset via Compute 终端). The 2024–25 glut bottomed near $2.99; the 2026 re-tightening is the story.
来源: Compute 终端 / GetDeploying (CC BY 4.0) · 报告值 · 截至 2026-10-14
On-demand rental rates by GPU — floor, median, ceiling
Advertised $/GPU-hour across tracked neoclouds and marketplaces, 2026-10-14.
- Floor (cheapest provider)
- Median
- Ceiling (most expensive)
Buy — accelerator unit pricing
Channel and secondary-market reference levels. No public list price exists for datacenter GPUs; treat 估计值 rows as order-of-magnitude.
| GPU | Memory | Unit price | 同比 | 来源 |
|---|---|---|---|---|
NVIDIA B200 (SXM6 module) Allocation product sold in systems; unit pricing is channel-报告值, not a public list. HBM3E content is now a large share of the BOM. Blackwell | 180 GB HBM3E | $34,000.00 | +9.70% | Channel-报告值 Blackwell module pricing range ($30k–$40k) 估计值 · 2026-07 |
NVIDIA H200 (SXM module) 上一代。随着 Blackwell 供应的改善而走软——这是加速器书中最明显的下降趋势。 Hopper | 141 GB HBM3E | $27,500.00 | -8.30% | 二级市场及渠道报价 估计值 · 2026-07 |
NVIDIA H100 (SXM5 module) The workhorse of the 2023–24 buildout. Unit pricing is drifting down as H200/B200 capacity ramps, but rental rates re-accelerated through 2026 — used H100s still clear quickly. Hopper | 80 GB HBM3 | $25,000.00 | -16.70% | 二级市场及渠道报价; rental-rate trajectory corroborates 估计值 · 2026-10 |
NVIDIA A100 80GB (SXM4) Two generations back. New-unit sales are effectively over; this is the secondary-market reference level for a part that listed near $15k+ at peak. Ampere | 80 GB HBM2e | $12,000.00 | -20.00% | Secondary-market reseller quotes 估计值 · 2026-10 |
AMD Instinct MI300X Priced below Hopper per unit with more memory per package; HBM content and CoWoS slots are the binding constraints, not logic dies. Instinct | 192 GB HBM3 | $20,000.00 | — | 渠道报告的本能定价 估计值 · 2026-07 |
AMD Instinct MI355X AMD's Blackwell-generation answer. Priced below B200 per unit. Instinct | 288 GB HBM3E | $25,000.00 | +4.20% | 渠道报告的本能定价 估计值 · 2026-07 |
NVIDIA RTX 5090 (consumer) The consumer ceiling. Popular in budget inference fleets — rental markets price it far below datacenter parts. Blackwell | 32 GB GDDR7 | $1,999.00 | — | NVIDIA MSRP; street pricing varies by region 报告值 · 2026-10 |
NVIDIA RTX 4090 (consumer) Previous consumer flagship. The backbone of marketplace rental fleets (Vast.ai, RunPod Community). Ada | 24 GB GDDR6X | $1,599.00 | -11.10% | Launch MSRP $1,599; secondary market tracks below 报告值 · 2026-10 |
Rent — on-demand $/GPU-hour by provider range
Advertised on-demand rates. Reserved, preemptible and spot capacity prices lower — often half the on-demand figure.
| GPU | Floor | Median | Ceiling | Providers tracked | 来源 |
|---|---|---|---|---|---|
NVIDIA B200 Blackwell rental supply is still thin; reserved 1-year cluster rates run above on-demand Hopper. | $5.98 | $6.69 | $8.87 | RunPod, 拉姆达, 芯织, Vast.ai | Provider pricing pages (拉姆达 $6.69 on-demand; $8.87–$9.86/GPU-hr on reserved clusters) 报告值 · 2026-10 |
NVIDIA H200 Sits between H100 and B200 on rate cards; DigitalOcean lists the low end at $3.40. | $3.40 | $3.99 | $6.31 | DigitalOcean, RunPod, Crusoe, 芯织, Oblivus | MLQ.ai cross-provider normalization (188 provider/GPU pairs) 报告值 · 2026-10 |
NVIDIA H100 Median up ~17.8% 同比 — the 2024–25 rental glut has fully cleared. Nebius raised on-demand 16.9% to $4.50 on Oct 1; AWS raised reserved 容量 Blocks 15% to $5.97. | $2.89 | $3.52 | $6.16 | RunPod, 拉姆达, Nebius, Crusoe, Modal, 一起人工智能, DigitalOcean, 芯织 | GetDeploying advertised-median across 41 providers; Memetik 8-cloud ranking 报告值 · 2026-10-14 |
NVIDIA A100 80GB The budget training tier. Marketplace floors under $1.50 reflect older fleet owners clearing capacity. | $1.47 | $1.99 | $2.79 | Vast.ai, RunPod, 拉姆达, Crusoe, 芯织 | Provider pricing pages and MLQ.ai normalization 报告值 · 2026-10 |
AMD MI300X Rents at a persistent discount to H100 despite more HBM per package — the software-ecosystem discount, visible in one number. | $1.91 | $2.39 | $3.45 | DigitalOcean, RunPod, Crusoe | MLQ.ai cross-provider normalization 报告值 · 2026-10 |
NVIDIA L40S Inference workhorse for providers without HBM allocation. | $0.47 | $0.99 | $2.25 | Vast.ai, RunPod, DigitalOcean, 芯织 | MLQ.ai cross-provider normalization 报告值 · 2026-10 |
NVIDIA RTX 5090 Consumer-card fleets price an order of magnitude below datacenter parts. | $0.25 | $0.35 | $0.69 | Salad, Vast.ai, RunPod | MLQ.ai cross-provider normalization 报告值 · 2026-10 |
NVIDIA RTX 4090 The cheapest meaningful inference capacity on the market. | $0.16 | $0.30 | $0.69 | Salad, Vast.ai, RunPod, Genesis Cloud | MLQ.ai cross-provider normalization 报告值 · 2026-10 |
Get the GPU rental and memory moves every Monday — rate-card shifts, HBM allocation news 和 three headlines that moved a price.
广告
最新供应链新闻
所有芯片新闻 ↗- DRAM市场数据
24 Sep spot board: DDR5 16Gb reaches $57.667 while DDR4 16Gb slips to $83.784
DDR5 16Gb 4800/5600 averaged $57.667 (+0.29% session) and eTT reached $25.500 (+0.79%). DDR4 出现分化：8Gb 3200 平均价格为 46.107 美元（+0.47%），而 16Gb 3200 则下跌 0.70%，至 83.784 美元。这些是已发布的主板上的会话变化，而不是 30 天的回报。
DRAMeXchange / TrendForce 现货板 ↗ - DRAM市场数据
17 Sep spot board: DDR5 averages $55.23; DDR4 16Gb retreats 1.85%
DDR5 16Gb 4800/5600 平均价格为 55.233 美元（+0.73%），DDR5 eTT 平均价格为 24.50 美元（+0.82%）。 DDR4 split: 8Gb 3200 rose 0.78% to $46.143 while 16Gb 3200 fell 1.85% to $86.00.该主板的价格仍然处于历史最高水平，但第一个可见的修正出现在选定的旧部件中。
DRAMeXchange / TrendForce 现货板 ↗ - NAND下行风险
由于消费者购买持续疲软，512Gb TLC NAND 跌至 20.083 美元
The 512Gb TLC wafer spot reference fell another 0.31% to $20.083 on 14 September after a 2.71% weekly decline to $20.146. Consumer purchasing remains cautious even while enterprise SSD demand keeps contract pricing elevated.
TrendForce ↗ - DRAM市场数据
Spot board, 1 Sep: DDR5 16Gb $54.08, DDR4 16Gb $92.02, DDR4 8Gb $44.54 — all up under half a percent
The 1 September session closed with DDR5 16Gb 4800/5600 at a $54.083 average (+0.31%), DDR4 16Gb 3200 at $92.024 (+0.27%) and DDR4 8Gb 3200 at $44.544 (+0.40%). Every line is still green, but the daily session moves are now fractions of a percent rather than the multi-point jumps of June and July. The spread is the story: DDR4 16Gb printed a $117.00 high against a $42.20 low in the same session.
DRAMeXchange / TrendForce 现货板 ↗
继续挖掘
FAQ
- How much does an NVIDIA H100 cost?
- As of 2026-10-14, the tracked H100 SXM reference level is $25,000.00 per module on the secondary/channel market, down roughly 17% year over year as Blackwell supply ramps. There is no public list price — datacenter GPUs sell through allocation and system integrators, so every unit figure on this page is flagged 估计值 or 报告值 with its source.
- How much does it cost to rent an H100 per hour?
- The advertised on-demand median across 41 tracked providers was $3.52 per GPU-hour in the week of 12 Oct 2026, up about 17.8% 同比。 Listed rates range from $2.89 (RunPod, PCIe) to $6.16 (芯织, per-GPU on an 8-GPU HGX node). Reserved and preemptible capacity prices well below on-demand.
- Why are GPU rental prices going up again?
- The 2024–25 rental glut has cleared. Neocloud rate cards bottomed near $2.99/H100-hour in early 2026 and have climbed since July; Nebius raised on-demand H100 16.9% to $4.50 on October 1 and AWS raised reserved 容量 Blocks 15% to $5.97. Inference demand is absorbing the older Hopper fleet while new capex shifts to Blackwell.
- What is the cheapest GPU to rent for AI work?
- Consumer cards bottom the table: RTX 4090 capacity rents from about $0.16/GPU-hour on marketplace fleets, and RTX 5090 from $0.25. Among datacenter parts, A100 80GB clears near $1.47–$2.79 and MI300X at $1.91–$3.45 — a persistent discount to H100 despite more HBM per package.
- Is it cheaper to buy or rent GPUs?
- At the tracked H100 levels, a $25,000 module renting at the $3.52 median breaks even against renting in roughly 7,100 GPU-hours — under a year of full utilization. Below ~60% utilization renting wins; above it, ownership or reserved clusters win. That math is why rental medians and secondary unit prices move together.
- How do GPU价格 connect to 内存价格?
- HBM is the largest bill-of-materials line on an AI accelerator after the logic die — six to eight stacks per package. The HBM 价格跟踪器 on this site covers the memory side; this page covers the accelerator itself 和 rental market it feeds.