The race to reduce costs and scale AI computing is pushing major cloud providers—and their ecosystem of partners—into a new phase: more racks with custom accelerators (ASICs) and less reliance on standard hardware. The clear thesis is: as inference volumes grow, efficiency per dollar and watt becomes paramount; and in this context, specialized designs (TPUs, internal chips, custom accelerators) become more attractive compared to general-purpose solutions.
In this environment, several industry sources point to a strong rebound in ASIC shipments for cloud in 2026, with Broadcom securing large-scale production projects alongside multiple cloud service providers (CSPs). Meanwhile, Taiwanese design and backend companies such as MediaTek, Alchip, and GUC are launching new products into volume production. The goal: accelerate the deployment of “ASIC-first” racks without waiting for longer cycles typical of traditional hardware.
The turning point: the business shifts from training to inference (and agents)
The fundamental change isn’t just technological—it’s economic. TrendForce describes how, after a period dominated by training large models using GPU servers + HBM, from the second half of 2025 the market is shifting toward inference as a service (Copilot, Gemini, and LLaMA-based applications, among others), with AI agents gaining increasing monetization importance. This transition means demand is no longer solely focused on “pure” AI racks: there’s also growing pressure on general-purpose servers performing pre/post-inference and storage tasks.
Simultaneously, capital expenditure is rising. TrendForce estimates that the total capex for the five major North American CSPs (Google, AWS, Meta, Microsoft, and Oracle) will grow by 40% year-over-year in 2026, partly due to infrastructure expansion and partly due to renewing servers acquired during the 2019–2021 boom.
ASIC market share reaches recent highs… but the bottleneck lies elsewhere
The most telling data is shipment distribution: TrendForce projects that ASIC-based AI servers will account for 27.8% of units in 2026, the highest since 2023, while GPU-based systems will still lead with 69.7%.
Within the ASIC landscape, Google stands out as the most advanced case: TrendForce emphasizes that Google’s commitment to proprietary ASICs surpasses many competitors, with their TPUs (which power Google Cloud Platform) also being marketed to external clients like Anthropic.
Up to this point, demand seems on track. The challenge—and operational risk for 2026—comes on the supply side: memory.
Why memory has become the limiting factor
In current AI racks, processing capacity only delivers real performance if the system can feed data with sufficient bandwidth and low latency. This places memory at the center:
- High-performance DRAM (especially HBM on GPU platforms) to handle parameters and activations.
- Enterprise SSDs for data pipelines, caches, and vector storage (RAG), with more random and demanding access patterns.
TrendForce’s industry forecast indicates that sustained demand for AI servers and enterprise storage will continue to drive the memory market upward through 2027, with annual growth rates exceeding 50% and peak revenues projected at $842.7 billion in that year (following a record $551.6 billion in 2026). (Data published by TrendForce in January 2026 analyzes the DRAM/NAND market.)
The key warning for the ASIC sector is that, although project volumes and deployment readiness are clearer than a year ago, memory capacity and scheduling for 2026 are becoming the most unstable elements. In other words: you can have the ASIC, the board, the network, and the rack; but if the necessary memory “budget” isn’t available, production rollout slows down.
Implications for 2026: more “tailored” designs and long-term contracts
With memory as a strategic resource, CSPs and integrators are adjusting their strategies in two ways:
- Securing supply: multi-year contracts for 2027–2028 and capacity agreements that mitigate short-term volatility.
- Optimizing architecture: designs that reduce memory pressure without degrading SLAs (cache hierarchies, compression, inference batching, RAG tuning, or modifications in prompting and context management).
The consequence for the supplier ecosystem is twofold. On one hand, ASIC developers and their partners (EDA, packaging, substrates, validation) enter a clear growth window. On the other hand, memory and storage are becoming the “toll” that will determine who deploys first and who waits.
Table 1 — AI rack value chain with ASICs (where the pace is gained or lost)
| System Layer | What it provides | Most common risk in 2026 | How to mitigate |
|---|---|---|---|
| ASIC (accelerator) | Cost/performance optimized for specific workloads | Volume ramp-up and time-to-yield | Co-design with CSP, short iterations, pre-validation |
| CPU/host | Orchestration, pre/post-inference | Overload due to inference growth | Fleet renewal, load balancing, offload |
| Memory (DRAM/HBM) | Bandwidth and latency | Insufficient or expensive allocation | Contracts, prioritization, redesigning profiles |
| Storage (SSD) | Datasets, vectors, caches | IOPS and enterprise SSD availability | JBOF/JBOD, layered scaling, tiering |
| Network (Ethernet/InfiniBand) | Scale and east-west traffic | Tr bottlenecks in inference traffic | Specific topologies, 400G/800G, traffic engineering |
Table 2 — Executive summary: why memory “risk” matters even if the ASIC is ready
| Signal | What it indicates | Direct impact |
|---|---|---|
| Increase in ASIC purchase plans | Demand is already decided | More pressure on advanced nodes and backend |
| Deployment delays “without demand reason” | Supply problem | Memory dictates the actual timeline |
| Contracts for 2027–2028 signed | Short-term scarcity is assumed | It’s “compensated” later, but 2026 slows down |
Frequently Asked Questions
What is a cloud ASIC, and why is its adoption accelerating in 2026?
A cloud ASIC is a dedicated accelerator designed for specific workloads (e.g., model inference), usually promoted by a CSP to optimize cost, power, and performance compared to general-purpose hardware. The expansion of inference and AI agents makes this efficiency more valuable than ever.
How much market share could AI servers based on ASICs reach in 2026?
TrendForce projects that ASIC-based AI servers will account for around 27.8% of shipments in 2026—a recent high—though GPUs will continue to dominate the majority of the market.
Why is memory (DRAM/HBM and enterprise SSDs) considered the main risk for deployment?
Because modern AI is intensive in bandwidth and data access. If memory doesn’t arrive in volume and on schedule, the rack won’t deliver expected performance, delaying production even if the accelerator has already been validated.
What implications could this have for prices and availability of AI cloud services?
If memory and storage supply constrain deployment, the effective cost per token or query may take longer to decrease. Meanwhile, CSPs will prioritize workloads with clearer returns and tighten governance through quotas, throttling, and service levels.
Source: Jukan on X