After Morgan Stanley, Goldman Sachs Asia team also lowered its AI server delivery forecast this year.

Following Morgan Stanley's expectation that the cloud spending cycle may peak this year, and the GB200 delivery forecast this year has been greatly lowered, the Goldman Sachs Asia team also lowered its AI server delivery forecast this year. However, the difference is that the Goldman Sachs Asia team believes that this does not mean that the industry has reached the peak of the cycle, but that the market is undergoing structural changes. The short-term adjustment is more affected by factors such as product upgrades, and it will usher in better performance in 2026.

Recently, Goldman Sachs Asia team wrote in the report that they have updated the global server market scale, and some AI server production plans will be postponed from 2025 to 2026. This is mainly influenced by GB200, GB300 product iteration and DeepSeek, which leads to the redistribution of budgets of different AI servers and the adjustment of production bases.

Goldman Sachs Asia team predicts that 31,000 rack-level AI servers will be delivered in 2025 (with 72-GPU equivalent calculation), down from the previous expectation of 38,000. In 2026, it is estimated that 66,000 rack-level AI servers will be delivered, which is higher than the previous expectation of 58,000.

Specifically, looking at the performance in 2025, the above team predicts that the shipment of ODM rack-level AI servers will fluctuate in 2025: the delivery volume is small in the first quarter, climbing in the second quarter, product iteration in the third quarter, and rising again in the fourth quarter.

Looking at the AI training server market, the above team expects that the production plan will also be postponed to 2026 (calculated by 8-GPU equivalence):

In 2025/26, the shipments of AI training servers are expected to increase by 42%/42% year-on-year, reaching 716,000 units/1.016 million units respectively.

The revenue of AI training server is expected to increase by 46%/39% in 2025/26, reaching $179 billion/$248 billion respectively.

Among them, the value growth in 2025 is faster than the shipment, mainly due to the new product form and the unit price increase brought by GPU upgrade.

Regarding the revenue outlook of AI server market, Goldman Sachs Asia team expects:

The revenue of AI training server is expected to increase by 46%/39% year-on-year in 2025/26, reaching $179 billion/$248 billion respectively.

The revenue of AI inference server is expected to increase by 132%/14% in 2025/26, reaching $18 billion/$21 billion respectively.

Shipments of general-purpose servers are expected to increase by 5%/4% year-on-year in 2025/26. The demand recovery in this market is mainly driven by the server replacement cycle, and its market value is expected to increase by 9%/6% in the same period.

It is worth mentioning that the Goldman Sachs Asia team believes that with the AI application scenario and landing, the AI reasoning server will usher in a higher popularity in 2025.

The recently released low-cost DeepSeek R1 model further reduces the cost of large-scale model training and reasoning, so the shipment of AI reasoning servers is expected to increase by 58%/24% in 2025/26, and the market scale will increase by 132%/14%.

Risk warning and exemption clause

The market is risky and investment needs to be cautious. This paper does not constitute personal investment advice, nor does it take into account the special investment objectives, financial situation or needs of individual users. Users should consider whether any opinions, viewpoints or conclusions in this article are in line with their specific situation. Invest accordingly at your own risk.