Generative AI server market projected to reach $1.88 trillion by 2035: computing power infrastructure enters a phase of structural restructuring

The global generative AI server market is undergoing a round of expansion that is not cyclical. According to estimates by Precedence Research, the market was valued at $101 billion in 2025, is expected to reach $135.34 billion in 2026, and is projected to grow to approximately $1.885 trillion by 2035, with a compound annual growth rate of 34% from 2026 to 2035.

The significance of these figures lies not in the growth rate itself, but in the structural change they reflect in how enterprises use computing resources: servers are no longer merely general-purpose compute carriers, but are becoming an infrastructure category redesigned around training, inference, and agentic AI workloads.

Industry Context

The definition of generative AI servers has already diverged from that of traditional enterprise servers. They are systems specially optimized for large language model training and inference, composed of high-density GPU or accelerator clusters, high-bandwidth memory, and specialized interconnects. Traditional servers cannot support the deployment of foundation models in terms of memory bandwidth and interconnect density, which has redefined the meaning of the word “server” within the industry.

Demand-side momentum comes from three directions:

First, hyperscale infrastructure expansion. Hyperscale cloud providers continue to purchase GPU racks for both training and inference, which is the primary driver of current global server demand.

Second, model scale continues to balloon. The growth in foundation model parameters and context sizes is forcing infrastructure suppliers to plan capacity ahead of time, thereby amplifying the mismatch between the cadence of supply and the cadence of demand.

Third, enterprise adoption is spreading. Adoption is increasing in industries such as medical diagnostics and financial modeling, and enterprises are beginning to migrate to larger-scale compute clusters.

At the same time, server architectures themselves are becoming more complex. Liquid cooling, higher-density interconnects, and custom silicon are becoming standard configurations for next-generation systems, directly driving up manufacturing costs and average selling prices. Another reason for high average prices is supply constraints: production capacity for advanced chips and specialized cooling systems remains limited.

Market Impact

A measurable structural adjustment is taking place on the supply side. The report notes that the share of ARM-based CPUs in data center servers has risen from about 5% in 2020 to nearly 20%, with improved energy efficiency being the main reason for their rapid share gains.

The interconnect segment is upgrading in tandem. Optical interconnect suppliers are accelerating their migration to 1.6T optical modules, with the timeline pointing to 2026, in order to meet continuously growing data traffic demand. Targeted products are also emerging at the component level, such as the NVLink spine cartridge developed by Amphenol, which is designed for more compact server architectures.Supply chain bottlenecks are concentrated in advanced packaging and memory. The report notes that shortages in the advanced packaging industry and the memory supply chain are putting pressure on prices, but the overall pricing trend remains relatively stable. This means that in the short term, price increase pressure comes more from upstream capacity than from overheated demand.

In terms of deployment models, cloud still dominates. Cloud deployment models for generative AI workloads clearly exceed on-premises deployments, with on-premises and hybrid cloud forms coexisting but accounting for smaller shares. The direct impact of this landscape on enterprises is that compute acquisition remains primarily based on leasing elastic capacity, while self-built clusters mainly appear in scenarios with clear sovereignty or compliance requirements.

Procurement drivers are shifting from training to inference. The report explicitly points out that agentic AI workloads have become a driving force in application procurement, and the scale of inference workloads is surpassing the training-centric era. This change imposes requirements on hardware configuration, memory bandwidth, and network topology that differ from those of training clusters.

Competitive Landscape

GPUs remain dominant, but the challenge from custom silicon is becoming substantive. The report views Google Ironwood TPU becoming generally available in 2026 as a key milestone, believing it gives hyperscalers a more attractive option beyond traditional GPU procurement. Google Cloud also launched ARM-based Axion instances, interpreted by the report as a signal that hyperscalers are trying to establish an alternative path beyond NVIDIA.

AMD is entering through product segmentation. The Instinct MI400 series is described by the report as representative of next-generation architecture evolution: MI430X targets sovereign AI and high-performance computing, while MI440X and MI455X target AI-specific precision workloads. The series is said to be among the first graphics processors to support the new UALink scale-up interconnect standard. The significance of UALink is that it attempts to establish an open standard beyond proprietary interconnects, thereby reducing lock-in risk in cluster scaling.

Network architecture is becoming a new competitive dimension. The report mentions that Google's Virgo Network can connect 134,000 TPUs into a single network fabric within a single data center, and supports connecting more than 1 million TPUs across multiple data centers; notably, the same network fabric architecture is also used for NVIDIA hardware. This interoperability design means the network layer is no longer solely tied to a single accelerator vendor.

In-house chip development is deepening vertically. The report mentions that Google's eighth-generation chips include TPU 8t, a training chip designed by Broadcom, indicating that custom silicon has entered training, the most core workload domain, rather than being limited to inference.

• Taken together:- Beneficiaries: GPU suppliers, custom ASIC design partners, high-density interconnect and optical module suppliers, liquid cooling solution providers, and holders of advanced packaging capacity.
• Under pressure: buyers that rely on a single accelerator supplier, and traditional general-purpose server vendors lacking differentiated capabilities for generative AI workloads.
• Potential followers: other hyperscale cloud providers and sovereign AI projects, whose common move is to expand procurement structures featuring custom silicon and parallel multi-vendor sourcing.

Enterprise Implications

First, multi-vendor procurement should be elevated from a cost issue to an infrastructure strategy. The report’s core judgment is that splitting procurement orders between GPUs and custom ASICs can effectively isolate the risks arising from allocation bottlenecks and enhance supply chain resilience. At the same time, maintaining collaborative coordination with multiple silicon design companies is viewed as a necessary action in the post-2026 supply environment.

Second, power availability has become the primary strategic constraint. The report notes that nearly all hyperscale projects announced in 2026 list energy availability as a primary strategic constraint. This means site selection, power supply agreements, and long-term electricity pricing structures are now on par with chip selection.

Third, procurement, power strategy, and supplier diversification must be synchronized. The report observes that enterprises that have advanced these three in a coordinated manner are better prepared to respond to fluctuations in the demand curve. For enterprises that still split infrastructure planning across three separate processes—procurement, IT, and finance—this is a clear organizational capability gap.

Fourth, the scale of inference workloads will redefine the cost structure. When inference overtakes training to become the primary workload, the importance of unit inference cost, energy efficiency ratio, and network latency will surpass peak computing power, and enterprises’ evaluation metrics need to be adjusted accordingly.

Fifth, risk is concentrated upstream. Advanced packaging and storage will remain bottleneck links for the foreseeable future, and enterprises should incorporate them into delivery-cycle assumptions rather than treat them as short-term disruptions.

Outlook

12 months (2026). The market is expected to reach US$135.34 billion. Key industry events include the deployment of Ironwood TPU and Google Cloud Axion instances, the ramp-up of AMD Instinct MI400 series production, large-scale delivery of 1.6T optical modules, and the initial ecosystem formation of the UALink standard. Power constraints will be most evident in project approval.

24 months. Inference and agentic workloads will dominate procurement decisions, rack-scale systems will become the main competitive battleground, and liquid cooling will shift from a high-end option to a default configuration. Differentiation among hyperscale providers will be reflected more in network fabric and interoperability than in any single chip model.3 years and above. A 34% compound annual growth rate trajectory from 2026 to 2035 means that heterogeneous computing architectures—a mix of GPUs, custom ASICs, and ARM CPUs—will become the standard form of data centers, rather than a transitional solution. Sovereign AI deployments will persist as an independent demand category and drive regional investment in compute self-sufficiency. On pricing, before constraints on advanced packaging and memory capacity are lifted, average server selling prices are unlikely to decline rapidly.

Conclusion

The generative AI server market currently presents a triple-layered state: extremely high demand certainty, clear supply constraints, and rapidly increasing dimensions of competition. For enterprises and investment institutions, what truly needs to be tracked is not annual scale figures, but three structural signals: whether procurement structures continue to diversify, whether power becomes a decisive variable for project implementation, and whether inference workloads have already surpassed training in procurement decisions.