On August 5, 2026, as the global large-model race entered a white-hot stage, the AI computing rental market saw another wave of price increases. According to data disclosed by multiple cloud service providers and computing trading platforms, cloud rental prices for Nvidia's flagship H200 GPU rose 25% month-over-month this month, with some edge models seeing increases of over 30%. Behind this phenomenon lies both a direct reflection of the continued shortage in H200 chip supply and a deeper structural shift in the AI industry from training-dominant computing to inference-dominant computing.
Supply gap widens, H200 becomes computing “hard currency”
Since H200 entered full-scale production at the end of 2025, the market had expected supply pressures to gradually ease. However, due to persistent bottlenecks in CoWoS advanced packaging capacity and HBM3e high-bandwidth memory supply, H200's actual delivery time remains as long as 20 weeks. Notices from multiple cloud providers to customers show that queues for new H200 instances have stretched from an initial 4 weeks to 8 weeks, with some hot regions even suspending new orders.
In the secondary market, this supply-demand imbalance is directly reflected in rental prices. According to public quotes from Lambda, the world's largest GPU trading platform, H200's per-card hourly rental price rose from $3.5 in early July to $4.4, an increase of 25.7%. The A100 model, previously favored by SMEs, was also boosted by the “spillover effect,” with prices rising about 12% in tandem. Analysts point out that when head cloud vendors' H200 resources are monopolized by large tech companies, small and medium developers can only turn to secondary computing resources, systematically raising the market's overall price center.
Inference computing demand explodes, market’s “second curve” forms
If the past two years' main theme in the computing market was the “large-model training arms race,” then the keyword for 2026 is undoubtedly “large-scale inference deployment.” The high-frequency calling APIs released by leading large-model companies such as OpenAI and Anthropic, along with the explosion of various AI agent applications, have made inference computing a new “scarce resource.”
According to professional institutions, global inference server shipments in Q2 2026 grew 65% year-over-year, surpassing training servers in total volume for the first time. This shift has had a profound impact on the computing rental market: inference tasks have lower chip utilization and are more fragmented, requiring more GPU nodes to be deployed for the same computing power, further intensifying demand for high-performance chips like H200. The rise of edge inference scenarios has also brought lower-power, more cost-effective H100 and AMD MI300 series back into market focus.
Cloud vendors face cost pressure, price hikes pass through to end users
The surge in computing rental prices is challenging cloud vendors' business models. Although AWS, Microsoft Azure, and Google Cloud have massive self-built computing pools, they have had to adjust pricing strategies in the face of rising hardware procurement costs and electricity expenses. It is reported that the three cloud giants successively raised the pricing rates for AI cloud server instances in July, and indirectly raised prices by shortening annual discount windows.
For startups that rely on cloud computing resources for model fine-tuning, this wave of price increases is a mixed blessing. On one hand, rising computing costs have raised the industry's entry barrier and accelerated the clearance of small and medium AI companies. On the other hand, companies that can lock in long-term computing contracts early are expected to consolidate their competitive moats. It is reported that multiple venture capital firms have made “computing reserve capability” a core indicator in due diligence for AI projects.
Industry outlook: supply bottlenecks may last until 2027
Based on Nvidia's current capacity plans, the new-generation Blackwell Ultra platform is scheduled to ramp up in Q4 2026, but its initial capacity will still prioritize core large customers. The market generally expects H200 supply tightness to last at least until early 2027, and before then, computing rental prices are likely to remain high and volatile.
For investors, amid the current price rises in the computing market, attention should be paid not only to the short-term game of supply and demand, but also to the industrial transformation brought by AI's shift from training to inference. Whoever secures stable computing resources first and utilizes existing computing power more efficiently will have a chance to take the lead in the coming wave of AI applications. And with the boost of capital, the story of this “gold mine” in the computing rental market may just be beginning.
