Google CEO Pichai says new computing power leasing demand in 2027 will be extremely difficult to meet, and the industry is entering a high-prosperity period worth hundreds of billions

Main text

Google CEO on the 2027 compute rental market: New demand is already extremely difficult to satisfy

According to Cailian Press, Google CEO Pichai revealed in a podcast interview that Google even briefly rented external compute last year to meet the needs of some customers. He said that even if customers now want compute for 2027, it is already extremely difficult for Google to meet demand, and companies across the industry all wish the scale they originally planned had been twice as large. Around the new model Gemini 4 Argon, he also discussed agent commercialization, the Apple partnership, and space data centers. Pichai acknowledged that supply chain bottlenecks and macroeconomic shocks could lead to overbuilding, but he remains confident in long-term compute demand.

Compute rental is the core path for filling the AI compute supply gap with an asset-light model featuring pay-as-you-go pricing and elastic scaling. Compute rental refers to service providers renting server compute resources equipped with high-performance chips such as GPUs to enterprises or institutions with AI training and inference needs through cloud-based, pay-as-you-go, elastically rented models. It is a key bridge connecting upstream compute hardware and downstream AI applications.

In terms of market size, according to data from the China Academy of Information and Communications Technology, the compute rental market is expected to exceed 260 billion yuan in 2026, and Founder Securities estimates that the domestic compute rental market is expected to approach 1 trillion yuan by 2030.

CSC Financial pointed out that compute rental is accelerating its climb, ushering in simultaneous rises in volume and price and entering a golden period of high prosperity at the hundred-billion-yuan level. The industry is leapfrogging from extensive expansion to high-end premiums and refined operations, and the Matthew effect is driving resources to concentrate among leading players. The business model is upgrading from bare metal rental to Token revenue sharing and integrated Token factory services. Combined with long-term order locking and TaaS elastic billing, this breaks the bottleneck of the heavy-asset cycle. On financing models, compute infrastructure construction is comprehensively entering a stage of financialization and high leverage, and capital and supply chain innovation have become core barriers.

Shanxi Securities believes that severe domestic AI compute shortages are driving the continued high prosperity of the compute rental industry. On the demand side, the explosive growth of AI inference demand, combined with the accelerated deployment of enterprise-level AI Agents, further raises the slope of the demand curve through the multiplier effect of per-customer compute consumption. On the supply side, Nvidia's high-end AI chips face multiple capacity bottlenecks. Market concentration in the compute rental industry is expected to continue rising. Participants have now differentiated into three tiers. The first tier has completed the construction of large-scale compute resource pools and locked in long-term orders. Small and medium-sized manufacturers face difficulties from depreciation and fixed interest costs due to low rental rates, and combined with a lack of strong business endorsement that pushes up financing costs, they will likely be gradually cleared out. The business model is undergoing a transition from compute resource rental to Token revenue sharing. Under a neutral assumption, by 2028 Token revenue sharing income for compute rental vendors is expected to reach 59.2 billion yuan. It is recommended to pay attention to vendors with supply chain advantages.

Kaiyuan Securities believes that as AI large models accelerate their evolution from training to inference, rapidly growing inference-side compute demand has become the core source of the supply-demand gap. In Q1 2026, domestic AI compute demand grew 417% year on year, while effective supply grew only 128%. The traditional IDC construction cycle of more than three years makes it difficult to match the rapid iteration of AI compute, and it faces the dilemma of becoming obsolete as soon as it is completed. Training-side compute remains the fundamental pillar of large model vendors' core competitiveness. H100 rental prices rose from $1.70 to $2.35, a cumulative increase of nearly 40%. New B-series products cannot yet fully fill the high-end training compute gap, and high-quality compute has entered a scramble-for-goods mode. The Matthew effect in the compute rental industry continues to stand out. GPU acquisition capability is the core survival barrier. Highly certified vendors can obtain direct supply quotas from original manufacturers that are 30% to 50% higher than ordinary partners, with delivery cycles shortened by 6 to 12 months. Leading vendors continue to seize the high-end market, while the survival space for tail-end small and medium-sized vendors narrows.

II. List of stocks of institutional interest

Reference research report sources:

Shanxi Securities - "Computer Industry: Severe Persistent Supply-Demand Mismatch, Business Model Transitioning to Token Revenue Sharing - Compute Rental Special Report" - October 14, 2026

CSC Financial - "Computer Industry: Long-Term Contract Closed Loop, Service Evolution and Financialization Breakthrough - Compute Rental Under High Prosperity" - August 30, 2026

Kaiyuan Securities - "Computer Industry: Compute Supply-Demand Gap Continues to Widen, Compute Rental Industry Prosperity Rising - In-Depth Report" - August 25, 2026

(Article source: 21st Century Business Herald) Middle of main text Embedded advertisement