MarketCatalyst
← Back to Resources
Semiconductors

Marvell, Penguin Bet on Recycled DDR4 and Light to Ease AI's Memory Crunch

Marvell and Penguin Solutions will put Marvell's Structera X controllers in Penguin's SMART CXL cards so data centers can reuse old DDR4, then use Marvell's Photonic Fabric to share petabyte-scale memory pools across racks. The goal: fuller GPUs, less strain on scarce HBM and DDR5, and cheaper AI inference.

Published Nov. 16, 2026 · 9:50 a.m. ET · 3 min read·MarketCatalyst

Marvell Technology (MRVL) and Penguin Solutions (PENG) announced a strategic collaboration on Monday to expand memory in AI data centers, pairing CXL-based cards that reuse older DDR4 modules with photonic memory pools that can be shared across racks. The companies are targeting three linked problems: idle GPUs waiting on memory, tight supplies of high-bandwidth memory (HBM) and DDR5, and the rising cost of running AI inference.

End-2026
Structera X-powered CXL cards ready for qualification and customer deployment
H2 2027
Photonic memory expansion servers expected to sample
Petabyte
Scale of the multi-rack memory pools the companies are targeting

Near term: reusing DDR4 through CXL

  • The product. Marvell's Structera X memory-expansion controllers will go into Penguin's SMART CXL add-in cards.
  • How it works. The cards let data centers take high-density DDR4 modules from decommissioned servers and attach them directly to current servers over PCIe 5.0 and CXL 2.0, adding memory capacity without replacing the servers.
  • Pooled appliances. The same cards can also be used to build dedicated memory appliances that pool recycled DDR4 across multiple hosts.
  • Timing. The cards are expected to be available for qualification and enterprise deployment by the end of 2026.

Longer term: memory over light

  • Photonic Fabric. Marvell's Photonic Fabric memory modules and network cards will go into Penguin's MemoryAI memory appliances and ComputeAI systems, with the goal of high-bandwidth, low-latency memory pools that reach petabyte scale across multiple racks for agentic AI workloads.
  • Inside the module. Each Photonic Fabric memory module pairs two HBM3e stacks totaling 72 GB with eight DDR5 channels and a silicon photonic interposer rated at 7.2 Tbps of bidirectional bandwidth.
  • The appliance. Penguin's first systems hold 16 modules, about 1 TB of HBM, and can serve up to 16 servers with 57.6 Tbps of aggregate bandwidth. Penguin is showing the MemoryAI Photonic Appliance at the SC26 supercomputing conference in Chicago this week.
  • Inference use. Marvell pitches the appliance as a faster place to offload the key-value caches that inference models build up, instead of pushing them to storage.
  • Timing. Servers integrating Photonic Fabric are expected to sample in the second half of 2027.

Why it matters for AI economics

AI models increasingly need hundreds of gigabytes, and in some cases terabytes, of memory per server. Buyers often add accelerators just to get more attached HBM, which leaves compute underused. Pooling memory lets capacity grow separately from GPUs, which the companies say should raise accelerator utilization, ease pressure on scarce HBM and DDR5, and lower the cost per token without a major redesign of existing infrastructure.

Where the technology came from. Marvell got Photonic Fabric through its acquisition of startup Celestial AI, which closed in February 2026 for about $3.25 billion upfront in cash and stock. The Penguin deal is one of the first named customer paths for that technology.

What the companies said

  • Marvell. Preet Virk, who runs Marvell's Photonic Fabric business, said scaling inference “requires delivering tokens at a higher economic efficiency.”
  • Penguin. CEO Kash Shaikh said memory is becoming central to both the performance and the economics of AI factories, and that the partnership advances Penguin's memory roadmap.

What to watch

  • Qualification. Whether SMART CXL cards with Structera X clear customer qualification on schedule by year-end.
  • Photonic samples. Whether the photonic memory servers sample in H2 2027 and how latency compares with local memory.
  • Memory prices. How long HBM and DDR5 stay tight, which drives demand for reuse and pooling.
  • Celestial payoff. Whether Photonic Fabric starts adding meaningful revenue for Marvell.

This article is informational only and is not investment advice.

Readers also read

MarketCatalyst is not a registered investment advisor and does not manage client assets. Content on this platform is provided for informational and educational purposes only. It is not investment advice, and MarketCatalyst is not a stock-picking or trade-alert service. Trading stocks and options involves risk, including the possible loss of principal. Consider your own goals, time horizon, and risk tolerance, and consult a qualified financial advisor before making any investment decision.

Marvell, Penguin Bet on Recycled DDR4 and Light