The weight of memory costs

The rise of artificial intelligence has rekindled the appetite for memory, driving prices to levels that now penalize even well-established suppliers. CyberTAN, a long-standing Taiwanese manufacturer of networking equipment, is a case in point: the company is trying to carve out a place in the new markets for AI and Wi-Fi 7, but is colliding with the high cost of DRAM and HBM chips and demand that struggles to gain traction. For organizations deploying on-premise infrastructure, this news is a warning sign: the cost of memory, an essential component for servers, GPUs, and edge appliances, is becoming a dominant factor in TCO, capable of slowing or even freezing local projects.

The vertical integration of AI and networking that CyberTAN is pursuing requires next-generation hardware: Wi-Fi 7 access points with on-board inferencing capability and appliances for self-hosting Large Language Models. In both cases, bandwidth and memory capacity requirements are stringent. HBM prices, in particular, have skyrocketed due to concentrated demand from data center GPUs, while DDR5 has followed a similar, albeit less explosive, trajectory. The result is a higher per-card cost that erodes margins and makes it harder for small and mid-sized enterprises to fund the transition to Wi-Fi 7 or to deploy local AI infrastructure.

Risky diversification in a shifting market

CyberTAN’s strategy reflects a broader trend: networking hardware makers are turning to AI as a new growth area, driven by a cooling traditional market and the opportunities offered by distributed inferencing. Wi-Fi 7, with its reduced latency and high throughput, is the ideal vehicle to bring AI compute capacity close to users in an edge logic. However, both markets require costly components just as businesses are more cautious. The weak demand for networking gear reported by the company might be a sign of a cyclical slowdown after two years of hybrid-work-driven investments, or of a reallocation of spending toward cloud infrastructure.

For an outsider like CyberTAN, high memory prices widen the gap with established AI incumbents. Large integrators can negotiate long-term supply contracts or absorb costs while waiting for volume. A specialized networking provider, on the other hand, must manage a rising bill of materials precisely when the market offers no room to pass costs downstream. This case shows how component inflation can act as a competitive filter, slowing innovation across the industry.

On-premise AI: the memory factor

Those who manage on-premise infrastructure are well aware of memory’s weight: in a multi-GPU server with 8 high-VRAM cards, the cost of HBM alone can exceed 40% of the machine’s price, and even CPU-based configurations with large RAM capacity for quantized model inference are sensitive to DDR5 pricing. In this scenario, the TCO of a local deployment becomes heavily dependent on the semiconductor market, a factor often overlooked in initial financial assessments. The rise in memory prices is no isolated event: AI-driven demand is likely to keep HBM supply tight throughout 2024 and beyond, influencing purchasing decisions.

The analytical frameworks discussed on AI-RADAR (see the /llm-onpremise section) help map the trade-off between upfront CapEx and operational cost: when memory prices climb, the break-even point against the cloud shifts forward, making on-premise less economical for intermittent or scalable workloads. Yet, data sovereignty and latency requirements continue to push many organizations toward self-hosted solutions. The choice of hardware configurations, quantization levels, and model distribution becomes critical to contain memory costs without sacrificing performance.

Outlook and choices for AI investors

CyberTAN’s situation is emblematic of a broader shift: memory cost has become a strategic risk factor for the entire AI value chain, from consumer hardware to large training clusters. Memory makers are investing in new lines for HBM4 and in expanding DDR5 capacity, but supply-side adjustment takes time. Meanwhile, AI memory demand will remain robust because each new model tends to require more parameters and thus more VRAM for inference.

For those designing on-premise architectures, this means that procurement planning must account for price volatility. Choices such as adopting lighter models, distributing inference across multiple lower-power nodes, or using dedicated NPUs can ease memory pressure, but they do not alter the underlying reality: without control over component costs, on-premise AI will remain a financial gamble. The CyberTAN case suggests that even established suppliers will need to redefine their strategies, prioritizing partnerships and modular designs that can absorb price spikes without halting innovation.