In the tumultuous era of technological advancement, the realm of artificial intelligence (AI) servers has emerged as a critical sector, garnering significant attention from both industry experts and the general public. Recently, influential financial institutions like Morgan Stanley have made headlines by forecasting that cloud spending cycles may peak this year, prompting them to substantially revise their shipment expectations for the key GB200 server. Following suit, Goldman Sachs' Asian team quickly shared a pessimistic outlook, triggering renewed discussions across market circles.
Goldman Sachs’ thorough report presents an updated assessment of the global server market size, highlighting critical findings. Notably, production schedules for certain AI servers have been pushed from 2025 to 2026. This shift is influenced by a myriad of complex factors, most notably the product iteration of GB200 and GB300 servers. Each new iteration represents a significant technological breakthrough, coupled with enhanced performance optimizations. However, these advancements necessitate enormous investments of time and resources for research and development, rigorous testing, and comprehensive adjustments within production lines to meet the escalating demands for high-performance technology.
The competitive landscape of the AI server market is continuously evolving, evidenced by the emergence of innovative technologies and products like DeepSeek. In this fierce competition, companies are revisiting their AI server budgets and strategically repositioning their production bases, striving to achieve greater efficiency while concurrently lowering costs, thus enhancing their market competitiveness.
Regarding specific delivery forecasts, particularly in terms of 72-GPU equivalent computing, Goldman Sachs provided clear estimates. They predict that in 2025, the delivery of rack-level AI servers will reach approximately 31,000 units, a notable decrease from earlier market expectations of around 38,000 units. Conversely, the forecast for 2026 projects an increase in delivery to about 66,000 rack-level AI servers, surpassing previous anticipations of 58,000 units. This juxtaposition in numbers illustrates the distinct developmental phases within the AI server landscape. The downward revision for 2025 signifies numerous challenges and uncertainties that the market currently faces—ranging from technological bottlenecks to fluctuations in demand and disruptions in supply chains. Meanwhile, the forecasted rise in 2026 suggests a recovering market that may capitalize on matured technologies and a more stable supply chain.
Examining the quarterly performance of 2025, the shipment volume of ODM rack-level AI servers reflects pronounced volatility. The onset of the first quarter saw reduced delivery volume due to traditional holiday influences and insufficient market demand. However, by the second quarter, activity began to pick up, with various sectors gradually increasing their acquisition of AI servers, resulting in a steep increase in shipments. Yet, as the third quarter approached—a critical phase for product iterations—companies were compelled to redirect more energy and resources toward developing and promoting new products, which temporarily hampered shipment totals. Fortunately, with the fourth quarter's arrival, new products gained traction in the market, leading to renewed growth in demand and rising shipment volumes.
From a macro perspective, Goldman Sachs’ Asian team also anticipates a similar deferral for the production plans of AI training servers, using an 8-GPU equivalent measure. For 2025/26, the shipment numbers for AI training servers are expected to rise by 42% in both years, reaching 716,000 units and 1.016 million units, respectively. In terms of revenue, their forecast estimates a 46% increase in revenue for 2025 to $179 billion, followed by a 39% increase to $248 billion in 2026. Notably, the faster growth rate in revenue compared to shipments in 2025 can be attributed to the innovative introduction of new product forms and the continuous enhancement of GPU capabilities. These elements collectively drive significant increases in product pricing, which contributes robustly to revenue growth.
Looking ahead at revenue projections within the AI server market, Goldman Sachs has crafted a comprehensive outlook. The anticipated revenue for AI training servers is set to continue its upward trajectory in 2025/26, with growth rates of 46% and 39%, respectively, resulting in figures of $179 billion and $248 billion. Revenue from AI inference servers is likewise predicted to explode, with anticipated increases of 132% in 2025 and 14% in 2026, tallying figures of $18 billion and $21 billion. Additionally, the general server market anticipates modest growth at rates of 5% and 4% for the same periods, driven primarily by the server replacement cycle, accompanying a respective 9% and 6% increase in total market value. Particularly noteworthy is Goldman Sachs’ observation that as the application of AI becomes increasingly prevalent, the market for AI inference servers in 2025 is poised for unprecedented interest. This is exemplified by the recent launch of the cost-effective DeepSeek R1 model, which has significantly reduced expenses related to training and inference in large models, consequently catalyzing the demand for AI inference servers. Expectations for shipment volumes in 2025/26 predict astounding increases of 58% and 24%, propelling market growth by 132% and 14%.