[SMM Computing Power Weekly] H100 Western Average Price Rises, A100 Beijing-Tianjin-Hebei Adjusted Upward

Published: Sep 18, 2026 21:11 (GMT+8)
In the week ending September 18, 2 of the 12 SMM computing power price indices in the western region, Beijing-Tianjin-Hebei, and Chengdu-Chongqing were raised, while 10 remained flat. The average price of H100 in the western region rose 1.25% to 81,000 yuan/month (14.06 yuan per card-hour); A100 in Beijing-Tianjin-Hebei was raised 1.56% to 32,500 yuan/month.

I. Market Overview

This week (as of September 18), the compute power leasing markets in west China, Beijing-Tianjin-Hebei, and Chengdu-Chongqing showed a pattern of "mostly consolidating with localized gains": of the 12 SMM price indices across the three regions, 2 were raised and 10 remained flat, with none lowered. The raised indices were SMM-H100-80G-Complete Server-West China-Monthly and SMM-A100-80G-Complete Server-Beijing-Tianjin-Hebei-Monthly, with average prices rising to 81,000 yuan and 32,500 yuan/month, representing monthly rent increases of 1.25% and 1.56%, respectively. On the spot side, H100 complete server spot was newly quoted at 3.4 million yuan/unit, H800 online dropped to 68,000 yuan, hitting a new low for that model, and the relocatable 910B4 in east China fell below 20,000 yuan for the first time.

II. Analysis by Compute Power Leasing Resource Type

SMM-H100-80G-Complete Server-West China-Monthly range moved up. The minimum monthly rent rose from 76,000 to 78,000 yuan (a 2.63% monthly rent increase), and the average price rose from 80,000 to 81,000 yuan/month (a 1.25% monthly rent increase); the minimum per-card-hour rate rose from 13.19 to 13.54 yuan (a 2.65% per-card-hour increase), and the average per-card-hour rate rose from 13.89 to 14.06 yuan (a 1.22% per-card-hour increase). The lower bound moved up while the upper bound held steady at 84,000 yuan, with the overall quote midpoint in west China shifting higher.

SMM-A100-80G-Complete Server-Beijing-Tianjin-Hebei-Monthly was raised. The maximum monthly rent rose from 35,000 to 36,000 yuan (a 2.86% monthly rent increase), and the average price rose from 32,000 to 32,500 yuan/month (a 1.56% monthly rent increase); the average per-card-hour rate rose from 5.56 to 5.64 yuan (a 1.44% per-card-hour increase), while the minimum monthly rent remained unchanged at 29,000 yuan. This was the first increase this month for the mainstream existing model in Beijing-Tianjin-Hebei, creating a regional price spread against the 34,000-yuan anchor of the three-year closed contract in east China.

SMM-H800-80G-Complete Server-Beijing-Tianjin-Hebei-Monthly held flat, while online prices continued to hit new lows. The index remained at 71,250 yuan/month (12.37 yuan per card-hour); online cloud resources dropped to 68,000 yuan/unit-month, 8,000 yuan lower than the 76,000-yuan annual contract in central China (a 10.5% decline), making it the only H-series model in sustained decline.

SMM-910B-64G-Complete Server held flat across the three regions, while delivery premiums widened. Beijing-Tianjin-Hebei stood at 19,500 yuan (3.39 per card-hour), Chengdu-Chongqing at 18,787.5 yuan (3.26 per card-hour), and west China at 13,750 yuan (2.39 per card-hour). Three relocatable 910B4 units in east China fell below 20,000 yuan for the first time, a 33% premium over the 15,000-yuan online price; the 910B3 held at 19,000 yuan, with the two specifications showing a 1,000-yuan price spread for the first time.

A800, H20, and all Chengdu-Chongqing indices held flat. SMM-A800-80G-Complete Server-Beijing-Tianjin-Hebei stood at 32,000 yuan; SMM-H20-96G-Complete Server-Beijing-Tianjin-Hebei at 43,000 yuan (7.47 yuan per card-hour); Chengdu-Chongqing SMM-A100-80G-Complete Server-Chengdu-Chongqing at 37,000 yuan, SMM-A100-40G-Complete Server-Chengdu-Chongqing at 12,000 yuan, SMM-A800-80G-Complete Server-Chengdu-Chongqing at 35,000 yuan, and SMM-H20-141G-Complete Server-Chengdu-Chongqing at 45,500 yuan. The H20 price quotes for the same specification span 18%, making it the model with the most complex caliber.

III. Market Outlook and SMM Views

In the short term, the upper bound of the Beijing-Tianjin-Hebei A100 breaking above 36,000 yuan and the lower bound of the western H100 moving up are in the same direction, with the midpoints of existing most-traded and high-end models rising in tandem. The continued decline of the H800 indicates that demand is concentrating toward high-performance models, and the intra-regional price spread structure will further differentiate.

Three points warrant attention. First, premiums on the sell side are becoming explicit: a South China H20 141G branded full system is quoted at 3.25 million yuan per unit with a three-year warranty, 850,000 yuan higher than the 2.4 million yuan for a 96G full system in the same region. An H200 HGX full system is quoted at 5.66 million yuan, with brand, warranty, and interconnect specifications becoming independent pricing dimensions. Second, the procurement threshold for large-parameter training is rising, with market demand emerging for 16 H200 units (a 128-card cluster) to support pretraining of a 300B large model, where price gives way to delivery certainty. Third, starting September 21, SMM will publish the "SMM Computing Power Leasing Price Composite Index," using standard compute units (SCU) to unify cross-model calibers. At the same time, caution is needed regarding the risks of custody investment models that promise annualized returns above 20% and include repurchase guarantees.

This week's core data summary (2 up, 10 flat):

  • SMM-H100-80G-full system-West-monthly: average price 81,000 yuan/month (monthly rent +1.25%), 14.06 yuan per card-hour (card-hour +1.22%)

  • SMM-A100-80G-full system-Beijing-Tianjin-Hebei-monthly: average price 32,500 yuan/month (monthly rent +1.56%), 5.64 yuan per card-hour (card-hour +1.44%)

  • SMM-A800-80G-full system-Beijing-Tianjin-Hebei-monthly: 32,000 yuan/month, 5.56 yuan per card-hour

  • SMM-H800-80G-full system-Beijing-Tianjin-Hebei-monthly: 71,250 yuan/month, 12.37 yuan per card-hour

  • SMM-H20-96G-full system-Beijing-Tianjin-Hebei-monthly: 43,000 yuan/month, 7.47 yuan per card-hour

  • SMM-910B-64G-full system-Beijing-Tianjin-Hebei-monthly: 19,500 yuan/month, 3.39 yuan per card-hour

  • SMM-A100-80G-full system-Chengdu-Chongqing-monthly: 37,000 yuan/month, 6.42 yuan per card-hour

  • SMM-A100-40G-full system-Chengdu-Chongqing-monthly: 12,000 yuan/month, 2.08 yuan per card-hour

  • SMM-A800-80G-full system-Chengdu-Chongqing-monthly: 35,000 yuan/month, 6.08 yuan per card-hour

  • SMM-H20-141G-full system-Chengdu-Chongqing-monthly: 45,500 yuan/month, 7.9 yuan per card-hour

  • SMM-910B-64G-full system-Chengdu-Chongqing-monthly: 18,787.5 yuan/month, 3.26 yuan per card-hour

  • SMM-910B-64G-full system-West-monthly: 13,750 yuan/month, 2.39 yuan per card-hour

Data Source Statement: Except for publicly available information, all other data are processed by SMM based on publicly available information, market communication, and relying on SMM's internal database model. They are for reference only and do not constitute decision-making recommendations.

For any inquiries or for more information, please contact: lemonzhao@smm.cn
For more information on how to access our research reports, please contact:service.en@smm.cn
Related News
【Flash | Coherent’s PhotonLink Sets the Stage for a Commercial Indium Phosphide (InP) Ramp】
8 hours ago
【Flash | Coherent’s PhotonLink Sets the Stage for a Commercial Indium Phosphide (InP) Ramp】
Read More
【Flash | Coherent’s PhotonLink Sets the Stage for a Commercial Indium Phosphide (InP) Ramp】
【Flash | Coherent’s PhotonLink Sets the Stage for a Commercial Indium Phosphide (InP) Ramp】
Coherent launched PhotonLink, an integrated optics platform for AI infrastructure incorporating Indium Phosphide (InP) lasers, silicon photonics, detectors and optical modules. The company disclosed more than 10 CPO engagements, over 10 NPO engagements and more than five chip-to-chip connectivity projects, with anchor customers and long-term agreements secured for both CPO and NPO. Revenue is expected to begin ramping in Q4 2026. Coherent has shipped over 300 million InP lasers and is expanding its 6-inch InP manufacturing capacity to support the ramp. The customer pipeline provides a clearer commercial demand signal for InP devices and upstream materials, although additional capacity and indium consumption were not disclosed.
8 hours ago
[SMM Flash] SuperX launches AI Token Platform
Oct 01, 2026 01:21 (GMT+8)
[SMM Flash] SuperX launches AI Token Platform
Read More
[SMM Flash] SuperX launches AI Token Platform
[SMM Flash] SuperX launches AI Token Platform
During this year's Singapore Tech Week, SuperX, a full-stack AI infrastructure service provider, officially launched its AI Token Platform and AI collaboration applications globally. Core Logic (Token Economy and Distribution): As enterprises integrate an increasing number of different large models (text, image, video, audio), the management cost of connecting to each provider's API separately is extremely high. The platform enables enterprises to uniformly schedule and access more than 100 AI models from over 10 providers through a standardized, OpenAI-compatible unified API on a token basis. Model Plaza & Intelligent Routing: The platform includes a built-in comparison and evaluation environment (Model Plaza), allowing enterprises to select models based on performance and estimated token costs before sending requests. At the same time, intelligent routing dynamically distributes token requests based on region, latency, service availability, and price, significantly optimizing enterprises' AI operational costs.
Oct 01, 2026 01:21 (GMT+8)
[SMM Flash] Midea Group launches industry-first "Power-Cooling Hyper-Converged" solution
Oct 01, 2026 01:17 (GMT+8)
[SMM Flash] Midea Group launches industry-first "Power-Cooling Hyper-Converged" solution
Read More
[SMM Flash] Midea Group launches industry-first "Power-Cooling Hyper-Converged" solution
[SMM Flash] Midea Group launches industry-first "Power-Cooling Hyper-Converged" solution
During Data Centre World Asia 2026, Midea officially launched the industry-first power-cooling hyper-converged technology for AI computing centers, along with related AI Apps and an AI Token platform, directly addressing the power distribution and cooling bottlenecks caused by high-density energy consumption in current computing centers, and significantly reducing PUE (Power Usage Effectiveness) through system-level integration.
Oct 01, 2026 01:17 (GMT+8)