NVIDIA B200 GPU Leasing Prices Soar 94%! The GPU Leasing Market Enters a "Golden Era"

As AI large models evolve from "instant Q&A" to "long‑horizon tasks," computing power is becoming the scarcest hard currency of the AI age.

I. A Single Piece of News Ignites Computing Power Leasing Stocks

On June 17, 2026, Tuhin Srivastava, CEO of AI inference infrastructure provider Baseten, disclosed news that shook the entire AI industry: his cloud service provider had officially informed him that NVIDIA B200 GPU leasing prices would rise from $2.63 per hour to $5.10 per hour upon contract renewal in October this year – an increase of approximately 94%.

The news triggered an immediate reaction in the capital markets. On June 18, computing power leasing concept stocks rallied strongly in the afternoon. Dongfang Guoxin hit the 20% daily limit, while Zhongjiabochuang and Zhiweizhijie also surged to the daily limit. The Huaan CSI AI ETF rose 5.65% with a turnover of nearly 150 million yuan. Cambricon soared over 14%, and other core computing power chain stocks such as Hygon Information and Zhongji Xuchang followed suit. The CSI AI Index has gained more than 48% over the past six months, with computing power‑related components accounting for over 70% of its weight, fully reflecting the market's high recognition of the computing power industry's prosperity.

II. Behind the 94% Surge: Supply‑Demand Imbalance Reaches a Tipping Point

Why did B200 leasing prices skyrocket by 94%? The answer lies in both supply and demand.

On the demand side, growth has been explosive. Data from the China Academy of Information and Communications Technology (CAICT) shows that in the first quarter of 2026, domestic AI computing power demand surged 417% year‑on‑year, while supply growth was only 128%. As large models move from "instant Q&A" to "long‑horizon tasks" – allowing AI to work continuously for hours and autonomously complete entire complex projects – the consumption of computing power for inference grows exponentially. A Deloitte research report indicates that in 2026, inference tasks will account for two‑thirds of global AI total compute load.

According to the latest statistics from OpenRouter, a leading global AI aggregation platform, China's weekly token calls for domestic large models have ranked first in the world for seven consecutive weeks, reaching 14.19 trillion tokens, up 27.49% month‑on‑month, and surpassing the United States for six straight weeks. Wanlian Securities research suggests that as domestic large models continue to iterate and upgrade, and application scenarios broaden and deepen, token call volumes are expected to keep growing.

On the supply side, constraints remain severe. NVIDIA's Blackwell‑architecture B200/H200 and other high‑end chips are in short supply globally. The delivery cycle for purchasing 1,000 GPUs has now extended to 12–15 months, and new orders are generally scheduled for the second quarter of 2027. Guosheng Securities points out that only top‑tier cloud vendors can obtain adequate supplies of high‑end computing power; second‑tier cloud providers and large model companies face a huge unmet demand for high‑end compute.

This supply‑demand gap has directly pushed up leasing prices. According to the H100 one‑year lease contract price index published by SemiAnalysis, NVIDIA H100 GPU lease contract prices bottomed at $1.70 per GPU‑hour in October 2025 and then surged to $2.35 per GPU‑hour by March 2026 – an increase of nearly 40%. In the domestic market, monthly H100 rents have stabilised in the range of 55,000‑60,000 yuan, with new contract delivery schedules generally pushed to the first half of 2027. In the spot market, the situation is even more extreme – on‑demand leasing capacity for all types of GPUs is fully sold out, and even as prices continue to rise, users who have locked in on‑demand instances are unwilling to release capacity back to the market. Citigroup analysts also noted in a June 14 report that A100 GPU leasing prices rose 0.6% over the past week and have accumulated an 11% increase over six weeks, indicating that AI computing demand is not concentrated solely on the most advanced hardware.

Upstream computing hardware has also entered a broad price‑hike cycle. According to Hualong Securities, Intel and AMD notified customers in late March and early April 2026 respectively of price increases across their entire CPU product lines, with average hikes of 10‑15%. Huatai Securities points out that domestic AI chips are also facing a price‑adjustment window – HBM procurement costs are expected to rise over 50% in the second half of the year, driving up overall AI chip costs by approximately 30‑50%.

III. The Computing Power Leasing Market Enters the 260‑Billion‑Yuan Era

The surge in GPU leasing prices is a microcosm of the booming computing power leasing market.

According to CAICT data, the domestic computing power leasing market reached 68 billion yuan in Q1 2026, up 62% year‑on‑year, and is expected to exceed 260 billion yuan for the full year. Over a longer horizon, China's computing power leasing market grew from approximately 148 billion yuan in 2024 to around 211.6 billion yuan in 2025, and is projected to reach roughly 260 billion yuan in 2026.

In terms of leasing capacity, data from Kexin Consulting shows that China's intelligent computing power leasing market grew from 377 EFLOPS in 2024 to 632 EFLOPS in 2025, and is expected to reach 984 EFLOPS in 2026 – nearly doubling each year. In 2026, computing power leasing orders are expected to be densely released, with total annual leasing volume likely reaching four times that of 2025.

At the same time, billion‑yuan computing power cooperation deals have been emerging frequently in the capital markets, with various companies entering the computing power sector. According to CCTV News, the computing power industry is seen as the digital era's equivalent of high‑end "office building leasing" – companies invest heavily in building data centres and facilities, then lease computing resources to generate returns.

On the policy front, the state is promoting computing power infrastructure development with unprecedented intensity. The "15th Five‑Year Plan" explicitly calls for deepening the "East‑West Computing Power Transfer" project and building a multi‑level computing facility system and a national integrated computing network. The National Data Bureau's "2026 Digital Economy Development Work Priorities" lists "accelerating the construction of a national integrated computing network" as a key task. The Ministry of Industry and Information Technology is promoting computing power infrastructure construction along the "points, chains, networks, and surfaces" framework. It is estimated that direct investment in the computing network will reach the trillion‑yuan level, and it will also drive the accelerated growth of AI, autonomous driving, low‑altitude economy, and other industries.

IV. Internet Giants Race to Build Their Computing Power Moats

In this computing power race, Chinese internet giants are accelerating their computing power deployment at unprecedented speed.

ByteDance is in discussions with Iluvatar CoreX to purchase at least 50,000 AI chips, mainly for large‑model inference workloads. If the deal is completed, Iluvatar will become ByteDance's third GPU supplier, after Huawei and Cambricon. ByteDance has explicitly split its training and inference chip supply chains: large‑scale model training uses Huawei Ascend and Cambricon high‑end training cards; while online C‑side Doubao and enterprise MaaS high‑concurrency inference will adopt Iluvatar's Zhikai series of dedicated inference GPUs.

Beyond ByteDance, other internet giants are also intensively building data centres. Baidu is building 10,000‑card‑level AI computing clusters nationwide; Alibaba's single‑quarter capital expenditure in fiscal year 2026 exceeded 38 billion yuan, with an estimated total of over 380 billion yuan over the next three years to boost cloud and intelligent computing hardware; Tencent is building HCC high‑performance AI clusters in multiple regions across the country, with large‑scale adoption of domestic computing power in the second half of 2026, while also providing mixed‑large‑model MaaS, gaming AIGC, and video generation inference services.

A technical lead in AI infrastructure at Beijing Zhipu AI (Facewall) analysed that internet giants are building multi‑vendor computing power systems mainly for supply certainty, bargaining power, and cost structure – in their computing demand, inference volume far exceeds training, and inference has lower requirements for chips than training, making domestic chips already viable for inference.

V. The Rise of Domestic Computing Power: A "China Solution" for GPU Leasing

Against the backdrop of surging overseas GPU leasing prices and extended delivery cycles, the domestic computing power ecosystem is accelerating, offering AI developers and enterprises a new option for GPU leasing.

Large models deeply integrated with domestic computing power. On June 17, Zhipu AI officially launched and open‑sourced its next‑generation flagship large model GLM‑5.2, focusing on "long‑horizon task" capabilities. It achieved the best performance among globally available models on Code Arena, a front‑end development evaluation system involving millions of users worldwide. GLM‑5.2 is open under the MIT license and free for commercial use. Its online inference is fully supported by multiple domestic computing platforms, achieving full adaptation on Day 0 with eight domestic computing platforms including Huawei Ascend, Alibaba T‑Head, Moore Threads, Cambricon, Kunlunxin, Muxi, Hygon, and Biren. It is expected that when the Ascend 950 supernode is launched in the second half of the year, it will become a powerful computing foundation for GLM‑5.2.

Domestic chips complete trillion‑parameter large‑model training validation. Recently, a project team from the Shenzhen Hetao College AI Training Platform, in collaboration with Harbin Institute of Technology and Huawei, successfully completed full‑parameter post‑training of the 1.6‑trillion‑parameter DeepSeek‑V4‑Pro large model on an Ascend 910C domestic AI computing cluster. The project achieved a model computing utilisation rate exceeding 30% and improved core training operator efficiency by 14%, with all metrics meeting industrial‑grade application standards. This practice validates that domestic AI chips can support world‑class ultra‑large‑parameter model training. Previously, domestic computing power was mostly used for inference and fine‑tuning, and was difficult to apply to full‑parameter deep training – this breakthrough marks that domestic computing power is no longer limited to inference.

Domestic chips receive national certification. In May 2026, for the first time, the state established an AI chip category in its security‑reliability evaluation, with nine domestic chips receiving the highest security‑reliability Level I rating, including Huawei HiSilicon, Alibaba T‑Head, Biren Technology, Hygon Information, Iluvatar CoreX, Muxi, and Moore Threads. Localisation (Xinchuang) has become a mandatory compliance requirement for government, enterprise, and state‑owned business.

Domestic GPU companies accelerate product iteration. On June 17, Suanmiao Technology, which focuses on 3D‑architecture AI cloud‑scale chips, announced that its 3D TokenPU A4E chip for large‑model inference had officially taped out, achieving high‑end AI chip production relying on the domestic supply chain. On June 13, MiniMax open‑sourced its million‑context multimodal large model M3, and Moore Threads announced that its AI training‑inference all‑in‑one computing card MTT S5000 had completed the initial adaptation.

Domestic computing power leasing platforms are on the rise. Take the Omniyq.com computing power leasing platform, under Shenzhen Cloud Engine Technology Co., Ltd., as an example. This company is a professional platform specialising in bulk supply of GPU computing resources and services. The platform brings together a full‑spectrum computing power lineup – from consumer‑grade GPUs like RTX 3090 and 4090 to enterprise‑grade data centre GPUs like A100, A800, H100, and H800 – while also covering NVIDIA's full range and domestic Xinchuang GPU servers. Leveraging massive high‑performance computing cluster infrastructure, Omniyq.com provides flexible and efficient GPU cloud leasing, graphics card and server purchase services for upstream and downstream AI industry clients. In terms of industry influence, Omniyq.com has appeared at both the 2024 and 2025 China Computing Power Conferences, demonstrating its core advantages in the GPU leasing space. Against the backdrop of soaring overseas GPU leasing prices and delivery cycles stretching to 15 months, domestic computing power leasing platforms represented by Omniyq.com – with their cost‑effectiveness and diversified supply systems – are becoming a pragmatic choice for an increasing number of AI developers and enterprises.

VI. From "Selling Compute" to "Selling Tokens": The Upgrade of the Leasing Business Model

Notably, the business model of computing power leasing is undergoing a qualitative leap – evolving from simple bare‑metal compute rental to model‑as‑a‑service or token‑sharing models, i.e., shifting from "selling compute" to "selling tokens." This transformation means that the value chain of computing power leasing is extending, and the industry's ceiling is being pushed higher.

2026 is exactly the inflection point where AI shifts from training‑dominant to inference‑dominant. Agent frameworks such as OpenClaw and ClaudeCode are rapidly gaining traction. Compute applications are moving from training to inference‑driven agent scenarios, from Q&A to agent loops, from single‑turn generation to multi‑step planning and continuous execution. TrendForce data shows that the top five North American cloud service providers expect AI training compute to grow 56% in 2026, while inference compute will surge 122% – more than double the growth rate of training. IDC predicts that by 2029, inference compute in the Chinese market will account for nearly 80% of the total.

The inference era imposes new requirements on computing power leasing: low latency, efficient cache management, and heterogeneous chip collaboration. Dedicated inference chips like the Groq LPU are emerging, working with GPUs in heterogeneous computing to make inference capabilities and efficiency more accessible. This means that GPU leasing is no longer about providing a single type of chip, but about building a diversified computing supply system that covers both training and inference scenarios.

VII. Computing Power Leasing: The "New Oil" of the AI Era

The 94% surge in NVIDIA B200 GPU leasing prices is no accident. It is an inevitable result of the severe supply‑demand imbalance in AI computing, and a strong signal that the computing power leasing market is entering a "golden era."

For AI startups, research institutions, and traditional enterprises, building their own computing infrastructure is becoming an insurmountable barrier – GPU procurement cycles of 15 months, continuously rising unit costs, and accelerating technology iteration. In this context, computing power leasing and GPU leasing are becoming the optimal way to access AI computing power.

From a 260‑billion‑yuan market size, to 417% demand growth, to a 94% rental price increase – all data points to the same conclusion: compute is power, and leasing is the future. Whoever can provide computing power leasing services in a more flexible and efficient way will take the lead in this AI race. As the domestic computing power ecosystem matures and the virtuous cycle of "model – compute – application" accelerates, the GPU leasing market is facing unprecedented opportunities.