This panel tracks what it costs to buy or rent the main AI accelerators in China and the US over the past 12 months. US rental prices update automatically every week; China rental prices and purchase prices in both countries are curated by hand, each with a dated source. Everything is converted to US dollars. The source, collection method and processing behind every data point are documented in the provenance section, and you can hover over (or tap) any point in a chart to see its original quote.
Key findings (Oct 2026)
- US rents are up across the board, and the newer the chip, the bigger the rise. Over roughly the past three months, on-demand H100 is up 10%, H200 up 17% and B200 up 20% (B200 1-year reserved up 38%), and RTX 5090 is up 24%. The previous-generation A100 is flat to slightly down.
- China rents are rising slowly. 1-year rents (monthly benchmark) for H100, A100 and H20 are each up about 5–6%; the Huawei Ascend 910B has barely moved. Most SMM indices have been flat since late September.
- Renting is cheaper in China. On a 1-year basis, an H100 costs about $2.15 per GPU-hour in China versus $3.26 in the US (about 0.66×, and the Chinese price includes 13% VAT). For on-demand rent, comparing Mysteel's like-for-like domestic and international indices, H100 and A100 are about 17–19% cheaper in China. The reasons: surplus capacity left over from the 2023–24 build-out of state-backed "intelligent computing centres"; large buyers building their own clusters and being steered toward domestic chips; a discount for embargoed cards, which carry no manufacturer warranty and come with compliance risk; and cheap power and rack space in western China. The domestic 910B costs about $0.51, a quarter of a Chinese H100.
- Chinese cloud-platform list prices corroborate the indices. We scrape public prices from 9 Chinese GPU-cloud platforms every week (the dotted lines in the charts). For the 4090 and 5090, platforms are within ±20% of each other and the median is within 6% of the SMM and Mysteel indices. For the H100, platform prices vary by about 2×, but SMM still sits mid-range (on-demand median ¥17.9 vs SMM ¥15.3; 8-GPU monthly median ¥12.2 vs SMM ¥14.3). Mysteel's monthly price runs higher because it tracks single-GPU monthly rentals, while SMM tracks 8-GPU servers.
- Buying is more expensive in China. An H100 server works out to about $60,000 per GPU in China versus about $41,000 in the US (about 1.5×); an H200 is about $110,000 versus $58,000 (about 1.9×). That is the scarcity premium created by the embargo.
- Consumer cards are being pushed up by memory prices. A new RTX 5090 averages about $4,500 in the US (MSRP $1,999) and about $8,100 in China; individual retail listings reach $7,000–8,600 (the diamond markers). ⚠ The Chinese 5090 figures mix variants (5090D, 5090D V2 and grey-channel cards).
Latest values and 3-month change
| GPU | US short-term $/GPU-hr | US 1-year $/GPU-hr | China short-term $/GPU-hr | China 1-year $/GPU-hr | US purchase $/GPU | China purchase $/GPU |
|---|---|---|---|---|---|---|
| Loading… | ||||||
The figure in brackets is the change versus about three months earlier (the most recent observation 75–135 days before the latest one); it is omitted when that window has no observation. Only each line's main series is used, not the diamond reference points or low-confidence points. Purchase prices: data-centre GPUs are 8-GPU server prices divided by 8 (so they include the rest of the server), consumer GPUs are single-card retail prices. New units are preferred; where there is no new-unit quote, a used price is shown and marked "used".
Charts by GPU
Each GPU has two panels: rental prices on the left (or top) and purchase prices on the right (or below). US lines are drawn in the site's primary colour, China lines in the secondary colour. Solid = short-term or new; dashed = 1-year or used; dotted = cross-platform medians from Chinese GPU clouds.
Methodology and sources
US rents (automatic, weekly). From the GetDeploying open GPU price dataset (CC BY 4.0). Short-term = on-demand price; long-term = 1-year reserved price. Both use the provider median: first the median quote within each provider, then the median across providers. The long-term line is only drawn in weeks with quotes from at least 2 providers. Updated automatically every Sunday at 18:00 UTC.
Long-term is standardised to 1 year. The US uses 1-year reserved contract prices. In China, whole-server leases are quoted in "RMB per server per month plus a term"; the mainstream minimum term is 1 year, and longer terms get lower unit prices. The China 1-year line uses SMM's "8-GPU server monthly benchmark price", which is broadly consistent with explicitly 1-year quotes from the same periods (July index ¥79,500 vs 1-year quotes of ¥76,000–80,000; October ¥82,500 vs ¥87,000, i.e. slightly low). Explicit 1-year quotes are overlaid as diamonds. Multi-year closed contracts and quotes with no stated term are shown as reference points only and are not on the 1-year line. The H200 in China is leased mainly on 3–5-year terms, so there is no 1-year price for it.
Short-term rents. US = on-demand (pay as you go). China = SMM's "single GPU, short-term, retail" index (diamonds are Mysteel's "single GPU on-demand" index and platform list prices). ⚠ The two countries' short-term series are not sampled the same way. For a like-for-like comparison, use Mysteel's paired domestic and international indices (Oct 26: H100 single-GPU on-demand, domestic ¥17.89 ≈ $2.68 vs international $3.23; A100 ¥9.28 ≈ $1.39 vs $1.71; domestic about 17–19% lower).
Chinese cloud-platform list prices (automatic, weekly). Every Sunday a script reads the public price pages or APIs (no login) of 9 independent Chinese GPU-cloud platforms: Gongji Compute (共绩算力), Compshare by UCloud (优云智算), Zhixingyun (智星云), Chenjianyun (晨涧云), Featurize, Xiangongyun (仙宫云), Matpool (矩池云), Yingboyun (英博云) and Jishi Zhisuan (基石智算). Only exactly matching models are used (the 4090 excludes the 4090D and 48 GB variants; the A100 is 80 GB only), and everything is converted to RMB per GPU-hour: on-demand prices as-is, 8-GPU monthly server rent ÷ 8 ÷ 720. For each platform and GPU the cheapest tier is taken, then the median across platforms; a point is drawn only with at least 3 platforms for on-demand or at least 2 for 8-GPU monthly rent. Excluded from the median: quotes with their own "last updated" date older than 180 days, and on-demand prices below the same platform's 8-GPU monthly rate (usually stale listings). Platforms do not state whether prices include tax. The big clouds (Alibaba, Tencent, Volcano Engine, Huawei, Baidu) list at 2–10× the independent platforms and do not publish H100/H200 prices, so they are not included.
China rents and purchase prices in both countries (hand-curated). Every point carries a date and a source and is tagged with its source type, collection method and processing (see the next section). Only real observations are recorded; nothing is interpolated. Chinese 8-GPU monthly rents are converted to GPU-hours by ÷ 8 GPUs ÷ 720 hours; SMM benchmark prices include 13% VAT. Points marked ⚠ come from a single source or have an unclear basis (for example tax, memory configuration, or server vs single card).
Exchange rate. RMB is converted at that week's USD/CNY rate (European Central Bank reference rate, via frankfurter.dev).
Export-control context
The H100, H800 and Blackwell parts (B200 and others) are banned from sale to China; units inside China come from existing stock and grey channels. The H20 can be sold. The H200 has been conditionally allowed since Jan 2026 (Nvidia pays 25% of the revenue to the US government), but actual deliveries have been very small. China-market versions of the RTX 4090 and 5090 are affected by export restrictions (the 4090D and 5090D).
Where the data comes from and how it is processed
Every data point is tagged with three things: its source type (what kind of source it is), its collection method (how we obtained it) and its processing (the conversions applied between the original quote and the value on the chart). Hover over any point in a chart to see its original quote and these three tags.
Processing pipeline
- Collection. US rents are downloaded weekly by script from the GetDeploying open dataset, and exchange rates are fetched weekly from the ECB reference rate. All other points are curated by hand: we open the source page (or its Wayback snapshot, index-history API or chart), copy the original quote verbatim into the
rawfield, and record its date and link. - Unit normalisation. Rents are normalised to "per GPU-hour": 8-GPU monthly rent ÷ 8 GPUs ÷ 720 hours, single-GPU monthly rent ÷ 720. Purchase prices are normalised to "per GPU": 8-GPU server prices are divided by 8 at render time (so they include the rest of the server and are not bare-card prices).
- Ambiguous quotes. Ranges use the midpoint; where only a change is reported, the previous value is back-calculated from it; vague dates such as "before Chinese New Year" or "early September" get an approximate date and are tagged "approximate date"; monthly averages are recorded on the 15th of that month.
- Currency conversion. RMB is converted to USD at the USD/CNY rate for the week of the observation. VAT is not removed (SMM prices include 13% VAT) and there is no inflation adjustment.
- Confidence and line assignment. Every point has a confidence grade (high / medium / low); low-confidence points are drawn hollow and excluded from the 3-month change. Each line connects points from a single main source only; other sources are drawn as unconnected diamond reference points.
- No interpolation, no smoothing. Weeks without an observation stay empty; US weeks with too few providers (fewer than 2 for reserved) are not drawn.
Summary by source type
| Source type | Points | Sources | Main use | Reliability notes |
|---|---|---|---|---|
| Loading… | ||||
Source by source
| Source | Type | Collection | Coverage | Points | Processing |
|---|---|---|---|---|---|
| Loading… | |||||
Tags are generated automatically from each source's domain and keywords in the original quote. The raw data is public: manual.json (hand-curated points), us_rent.json (weekly US rents plus exchange rates) and cn_platforms.json (weekly Chinese platform list prices). Original quotes are kept verbatim, so many are in Chinese.
Version history: v0.4 (Oct 26, 2026) added weekly medians from 9 Chinese GPU-cloud platforms; v0.3 tagged every data point with source type, collection method and processing; v0.2 standardised long-term rent to 1 year; v0.1 first release, Oct 26, 2026.
Adapted from the Spark Science research notes (originally published in Chinese at jimmyresearch.com). Last updated Oct 26, 2026.