Accelerated Computing Solutions Analyst

Marvell strengthens AI memory infrastructure portfolio. Article by: Marvell Technology Inc. New storage, CXL memory, and photonic technologies address growing memory bottlenecks in large-scale AI inference. Marvell Technology Inc. has expanded its AI memory infrastructure portfolio with new storage, memory expansion, and photonic interconnect technologies designed to improve the efficiency of agentic AI inference. The portfolio targets hyperscalers and cloud providers seeking to scale memory independently of compute as AI models become larger and more memory intensive. The announcement reflects a growing industry shift toward memory-disaggregated architectures, where memory capacity, bandwidth, and connectivity are increasingly viewed as critical factors alongside compute performance. As large language models adopt longer context windows and larger key-value (KV) caches, conventional server-attached memory architectures are becoming a limiting factor for AI inference performance. Marvell's latest portfolio spans three infrastructure layers: server-level AI storage, rack-scale CXL memory expansion and pooling, and pod-level optical shared memory. Together, these technologies aim to improve GPU utilization by reducing memory bottlenecks, minimizing data movement, and enabling AI systems to generate more inference tokens within existing power and data center constraints. At the server level, the new Bravera SC6 PCIe 6.0 SSD controller is designed to accelerate AI inference by allowing larger KV caches to be offloaded from high-bandwidth memory (HBM) to SSDs. Marvell said the controller delivers twice the performance of its previous-generation Bravera SC5 PCIe 5.0 SSD controller while reducing write amplification and improving NAND endurance. Support for NAND from multiple suppliers also provides hyperscalers with greater deployment flexibility. For rack-scale deployments, Marvell introduced its Structera X memory expansion platform, which enables hyperscalers to expand and optimize memory resources using Compute Express Link (CXL). The platform supports larger shared memory pools, helping improve infrastructure utilization and reducing the total cost of ownership while providing a migration path toward more flexible memory-sharing architectures for AI inference. The company also unveiled its Photonic Fabric architecture for pod-level optical shared memory. Comprising photonic memory modules, network interface controllers, and chiplets, the solution creates a shared-memory tier spanning multiple AI accelerators and racks over distances of up to 50 meters. Marvell said the architecture supports as much as 32 TB of warm KV cache while delivering high bandwidth and low latency, enabling up to two to three times higher token throughput without increasing data center footprints or power consumption. "AI infrastructure is moving beyond isolated servers to systems where compute, memory and connectivity operate seamlessly together," said Will Chu, executive vice president and general manager of Custom Cloud Solutions at Marvell. "As AI scales, memory must scale more independently of compute so resources can be deployed where they deliver the greatest value. With the industry's broadest AI memory infrastructure portfolio, Marvell is helping customers improve utilization, boost token efficiency and scale AI without compromising performance, power or cost." Alan Weckel, co-founder and technology analyst at 650 Group, said memory is becoming a primary constraint on AI performance as workloads continue to increase in size and complexity. "As AI workloads grow larger and more complex, memory capacity, bandwidth, latency and data movement are becoming primary constraints on AI performance," Weckel said. "Marvell's memory and storage portfolio gives hyperscalers and cloud providers a strong foundation for building scalable, efficient AI systems capable of supporting increasingly advanced workloads."

Marvell Technology's communications and other business segment is recovering as customer inventories normalise. In the first quarter of fiscal 2027, communications revenues increased 29% year over year to $585 million, driven by its AI data centre networking business. The company expects its interconnect business to grow more than 70% year over year in fiscal 2027. Management forecasts TIAs and drivers to exceed a $1 billion annualised run rate in coming quarters, with DCI module revenues reaching about $1 billion during fiscal 2028. However, Marvell's shift towards lower-margin custom silicon is causing gradual gross margin decline. The company expects its communications end market to decline mid-single digits sequentially in Q2 whilst growing high-single digits year over year. Marvell shares have rallied 157.3% year to date.

Marvell Technology's stock surged 25% over five trading days, prompting investor interest. However, analysis reveals the semiconductor company's returns largely mirror the broader market rather than offering a distinct investment opportunity. Over five years, Marvell demonstrated a 0.63 correlation with the S&P 500, indicating significant overlap with standard index funds. The company's revenue concentration reinforces this pattern, with 76% of its $2.418 billion in fiscal Q1 2027 revenue coming from data centre operations serving a handful of AI infrastructure buyers. Marvell's volatility presents additional considerations. Over the past year, the stock captured approximately 356% of the market's gains on up days and 307% of losses on down days, essentially amplifying market movements threefold in both directions. Its annualised volatility reached 64%, compared with 17.2% for the S&P 500.