AI infrastructure is now limited more by the network than by compute. The supporting layers like switches, optics, NICs, and ASICs cannot scale at the pace GPUs are accelerating. and then add this right after it. Together, they point to a clear industry trend: AI infrastructure is entering a multi-year phase where networking capacity — not compute — will determine what can actually be deployed.
During our recent conversation with Claus Aasholm, Founder of Semiconductor Business Intelligence and one of the semiconductor industry’s most trusted independent analysts, he described the memory market as being in a period of “structural shortage” driven by AI’s unrelenting demand. That prediction has played out exactly as he outlined.
Now, in his latest analysis, Aasholm is making a similar call — this time on AI datacenter networking. The upstream signals he’s tracking point to accelerating demand, lagging infrastructure investment, and early indications of a supply imbalance that will not correct quickly.
GPU performance continues to accelerate each generation, but networking throughput is not keeping pace. Clusters are increasingly compute-rich but bandwidth-limited.
Switches, optical transceivers, NICs, DPUs, and link ASICs must move dramatically more data across larger and denser clusters. Instead, these components are hitting manufacturing limits, foundry constraints, and uneven ordering cycles that are already constraining supply.
The result is a widening gap between what AI systems can compute and what the network can actually move.
Claus points to a shift that is less visible than GPU shortages but far more consequential: networking is now the true determinant of cluster performance and deployment speed.
As hyperscalers deploy larger clusters, each rack requires more bandwidth, more switching capacity, and more advanced optics. Even slight constraints in any of these components can delay entire deployments.
This is the same dynamic that pushed DDR4 and DDR5 into persistent scarcity: profitable upstream products take priority, and every other segment competes for what is left.
Networking is now entering that same structural scarcity cycle.
While leading-edge nodes remain in high demand for compute, the networking stack depends heavily on mature and mid-range nodes, which are already strained.
Optical transceivers, control ICs, and many networking ASICs rely on these nodes. Foundries are running at high utilization, and geopolitical factors make relocating this production extremely challenging.
Claus emphasized this imbalance in his discussion with Fusion: companies want to diversify away from China, yet a significant portion of mature-node networking capacity is still located there. The result is a tension between supply resiliency and supply reality.
Adding to this imbalance, Marvell’s own networking lead times have begun extending, reportedly tied to tightening conditions at advanced TSMC nodes. Pressure is rising across both ends of the process stack, and it is beginning to show in networking availability.
One of Claus’s most important signals is the rise of Broadcom inside datacenter networking.
Broadcom is the only major semiconductor vendor outside the GPU category gaining meaningful datacenter market share today, driven entirely by demand for:
At the same time, Nvidia is no longer selling standalone GPUs. It is selling full rack-scale systems such as the GB200 and GB300, a shift accelerated by Nvidia’s acquisition of Mellanox and its tightly integrated networking portfolio. These systems require enormous quantities of networking hardware — often more than the GPUs themselves.
Performance gains in AI systems now depend less on semiconductor node advancements and more on internal network architecture. As Nvidia moves further into complete system design, networking has become the determining factor in how much AI compute customers can actually deploy.
This shift is already making networking the rate limiter of AI expansion.
Several major patterns are converging:
This mirrors the same structural shift Claus identified in memory, where upstream profitability and constrained tooling created persistent scarcity.
Networking is entering that same multi-year cycle.
Organizations building AI infrastructure will need to adapt sourcing and planning strategies quickly. Based on the signals Claus highlighted:
Those who anticipate the tightening will secure timelines and pricing ahead of competitors.
The networking bottleneck will define AI infrastructure for the next several years. Fusion Worldwide helps close this gap by providing:
Our role is simple: remove sourcing friction where bottlenecks appear first.
Browse Fusion Worldwide’s networking inventory to see what’s available now, and catch our newest conversation with Claus Aasholm for insight into why networking is emerging as AI’s next constraint.
You can also follow Claus on LinkedIn and Substack for more of his research.
Is networking really the next AI bottleneck?
Yes. Multiple indicators—ASIC demand, optics tightness, switching silicon constraints—show networking is scaling slower than compute.
Is this similar to what happened with memory?
Very. Just as HBM reshaped DRAM allocation, rack-scale networking demands are reshaping ASIC, optics, and switching availability.
Why can’t mature-node production simply scale?
Relocation is difficult due to tooling, geography, and geopolitical risk. The majority of capacity remains where companies are trying to reduce reliance.
Are shortages already happening?
Yes, especially in optics and key ASIC-driven components. Lead times are extending and allocation patterns are emerging.