In 2026, AI servers remain one of the sectors with the clearest demand signals in the semiconductor industry. Recently, Terry Gou, Chairman of Hon Hai Precision Industry (Foxconn), revealed externally that the company's AI server capacity for at least the next year and a half is fully booked. Data from TrendForce indicates that the total capital expenditure of nine major cloud service providers in 2026 is estimated to exceed USD 957 billion, a year-on-year increase of 106%, with investment in AI servers and related computing infrastructure being a primary driver.
The surge in AI server demand has also propagated upstream. Extended lead times and sold-out capacities for GPUs, HBM, enterprise-grade SSDs, and MLCCs are the most direct signals in the supply chain recently. While the tightness in these categories stems from specific supply-side causes, the common driving force on the demand side is the concentrated release of AI server system demand. This demand propagates upstream along the AI server bill of materials (BOM), manifesting simultaneously as supply constraints across four categories, forming the current basic landscape of the AI server supply chain.
01、Core Component Capacity Crunch Points to the Same Demand Source
While extended delivery cycles for GPUs are no longer news in 2026, the magnitude of waiting times continues to set new records. Tuhin Srivastava, CEO of Baseten, revealed in June that the delivery cycle for procuring 1,000 GPUs has extended to 12–15 months, with new orders generally scheduled for delivery in Q2 2027.
Recently, foreign media reported that some internet clients' procurement demand for specific high-performance computing accelerator cards far exceeds suppliers' global planned production for the coming year, with actual deliveries potentially satisfying less than10% of orders. This supply-demand gap intuitively demonstrates that against the backdrop of surging AI inference and development needs, the global GPU supply chain faces severe technical and capacity bottlenecks.
Overall, AI server GPUs remain in short supply. Despite continuous capacity expansion, bottlenecks in advanced packaging, HBM supply, and final system delivery continue to limit shipment rates.
As a key component tightly integrated with GPUs through advanced packaging, High Bandwidth Memory (HBM) performance directly determines data throughput efficiency for large model training and inference. According to recent statements from the three major original manufacturers, the state of HBM supply tightness has substantively extended beyond 2027.
SK Hynix stated that combining signed long-term agreements and ongoing customer order negotiations, capacity schedules for DRAM and HBM are effectively sold out through 2029, with declared customer procurement demand reaching more than twice the existing supply capability.
Samsung management confirmed that HBM capacity for 2027 is fully sold out. The memory business has signed long-term supply agreements with 10 customers, while another 10 are negotiating multi-year contracts, aiming to lock in 60% to 70% of memory capacity through long-term agreements.
Micron publicly stated that its 2026 HBM capacity is fully sold out, while over 75% of 2027 capacity has already been allocated. Remaining capacity will be dynamically adjusted based on customer needs. Additionally, industry reports suggest that Micron's 2027 DRAM and HBM capacity allocations have substantively been completed.
Regarding new capacity, short-term expansion relies primarily on converting existing fabs and upgrading production lines. Industry experts note that although new production lines from major manufacturers are progressing, substantive capacity ramp-ups are largely expected only after late 2027 or into 2028 due to construction cycles and equipment lead times. This means HBM supply will remain tight from late 2026 through the first half of 2027. Although total output in 2027 will increase, demand will grow synchronously, likely extending the shortage until around 2028.
The capacity siphoning effect of HBM does not only affect GPUs. When manufacturers allocate more DRAM wafers to HBM, the supply space for general-purpose DRAM narrows accordingly, putting pressure on the AI server system BOM.
Beyond computing and bandwidth, AI's demand for storage is spreading from training clusters to broader data center segments.
AI servers heavily utilize high-capacity, high-endurance SSDs. However, building new wafer fabs takes considerable time to reach stable mass production, making it difficult for effective supply of enterprise-grade SSDs to catch up quickly in the short term. Manufacturers are mainly increasing supply through upgrades to older production lines, such as increasing the proportion of NAND layers above 200 and expanding high-capacity QLC products, but incremental gains in 2026 remain limited.
In comparison, MLCCs are the segment with relatively lower unit prices and often overlooked among the four, yet the signal of extended lead times cannot be underestimated.
Recently, Norio Nakajima, President of Murata Manufacturing, stated during an earnings call that a single AI server requires 15,000 to 25,000 MLCCs, approximately 10 times the quantity needed for a smartphone. Murata plans to invest CNY 1 trillion equivalent (JPY1 trillion) over five years, averaging about JPY 200 billion per year, roughly double previous levels. Approximately JPY 350 billion is allocated for construction projects, including two new MLCC and ceramic material production plants in Okayama Prefecture, expected to be completed in 2028 and 2030 respectively. Overseas, plans include adding new facilities at the Thailand plant by 2030. Looking at the timeline, new plant operations still require several years, meaning short-term supply tightness cannot be rapidly alleviated through new capacity. [Related Reading]
The four categories vary in the degree of tightness, but the largest common driving force is no longer traditional consumer electronics recovery or general server replacement, but rather the cross-category occupation of semiconductor capacity by AI infrastructure.
GPU deliveries are constrained by HBM and advanced packaging; HBM expansion squeezes general-purpose DRAM; enterprise-grade SSD increments are limited by wafer fab construction cycles; and MLCC capacity expansion will not gradually release until after 2028. Shortages do not occur in isolation within a single component but propagate layer by layer along the AI server BOM. Therefore, understanding this round of shortages requires looking back at the demand side: AI servers themselves.
02、Fully Booked Capacity and Surging Revenue: Signals from the AI Server Demand Side
Returning to the demand side, the pre-booking of OEM capacity is not an isolated phenomenon. Taking Foxconn as an example, Chairman Terry Gou recently stated that the company's AI server capacity for at least the next year and a half is fully booked. Behind this tightness in the ODM sector lies structural changes occurring in AI server delivery forms, customer composition, and computing loads.
The most intuitive change occurs in delivery form. AI servers are shifting from piecing together individual devices to integrated delivery in "rack" units. To pursue extreme computing power, manufacturers highly integrate dozens of GPU nodes into a single rack, directly causing power consumption per rack to break through traditional air-cooling limits, making liquid cooling standard rather than optional.
Rack-level delivery means customers are no longer buying isolated compute nodes but a complete system including high-speed interconnects, massive power supplies, and complex thermal dissipation. This hardware upgrade not only significantly raises the price per system but also implies that the physical concentration of core components like GPUs, HBMs, enterprise SSDs, and MLCCs within a single rack increases exponentially. The consumption of various semiconductors per system is far greater than that of past general-purpose servers.
Beyond tight ODM capacity, brand vendors' financial reports also signal "surging revenue." Supermicro's report published in early November showed quarterly revenue increased nearly threefold year-on-year, with AI solutions accounting for 71% of revenue. Previously, Dell disclosed that its backlog of AI server orders had reached tens of billions of dollars. These performance jumps from two brand vendors confirm a shift in industry procurement models: besides top-tier cloud providers directly customizing with ODMs, an increasing number of second-tier cloud providers and enterprise customers are beginning to procure ready-to-use systems through brand vendors.
Beyond more buyers, the nature of computing demand itself is changing. As large models transition from "training" to "deployment/inference," server configuration logic is rewritten. Previously, building training clusters essentially meant determining how many GPUs to purchase. Inference-oriented clusters, however, require not only sustained operation of compute chips but also larger system memory to support data throughput, more enterprise SSDs for caching, and a more balanced CPU-to-GPU ratio. This change transforms AI server procurement from simply "buying computing power" to matching overall capabilities across computing, storage, networking, and cooling.
The more integrated the hardware, the greater the difficulty for supply chain delivery. The combination of rack-form upgrades and increased inference loads makes today's AI servers a system where pulling one hair moves the whole body. If capacity for any supporting component within the rack (such as a specific capacitor or memory chip) fails to keep up, the delivery of the entire rack is delayed.
From tight capacity at leading ODMs, to surging revenues at Supermicro and Dell, to configuration upgrades driven by rack-based delivery and inference demand, signals from these three dimensions point to the same judgment: AI server demand growth is no longer merely an accumulation of node counts, but involves more complex system configurations per device, more semiconductor components involved, and dual expansion in both customer base and load types.
03、Market Value Estimate Jumps to Nearly USD 920 Billion: Changes and Constants in the 2027 AI Server Supply Chain
Looking ahead to 2027, the AI server market shows distinct changes in shipment growth and value growth rates. TrendForce estimates that AI Server shipments will grow 32.2% year-on-year in 2026, and nearly 36% in 2027. In terms of market value, the 2027 AI Server market value is estimated to jump to nearly USD 920 billion, growing over 90% year-on-year, rising to account for approximately 81% of the total Server market value. [Related Reading]
In 2026, GPUs remain the mainstay of AI server chips, with Blackwell ramping up and the Rubin platform entering early delivery stages. Meanwhile, the shipment share of ASIC AI Servers is increasing. Industry consensus expects NVIDIA to maintain dominance in the GPU market in 2027, with Rubin succeeding Blackwell as the main shipment driver. Concurrently, CSP self-developed ASICs and domestic Chinese chips are accelerating their rise, liquid cooling penetration continues to improve, and HBM4 volume production drives up HBM prices and revenue. The situation where a single GPU segment dominates supply pacing is shifting towards parallel tracks involving GPUs, ASICs, HBM4, and liquid cooling.
From the perspective of 2027 supply and demand dynamics, there are no clear signs yet that HBM's encroachment on general-purpose DRAM capacity will ease. The memory segment may remain the hardest part of the AI server supply chain to loosen.
While the form of supply constraints is changing, the shortage itself may not necessarily disappear. Focus points for the 2027 AI server supply chain may no longer be limited to GPU lead times alone, but need to simultaneously observe ASIC capacity ramps, HBM4 volume pacing, and liquid cooling deployment progress. These intertwined segments make the evolution direction of the AI server industry even more noteworthy.