Skip to report
Next edition in 679 min Sep 18, 2026, 12:45-23:00 Lisbon
Premarket Alpha Post-Market Alpha

Daily Alpha · Substack

Two items carried signal: an inference architecture that points DRAM/NVMe demand higher, and dedicated SiGe capacity for optical products.

10letters

Themes on this desk

Inference architecture

The Engram design extends standard token embeddings with learned multi-token lookups so recurring local patterns retrieve vectors directly, reducing the need to reconstruct them through attention and feed-forward layers — a mechanism-level

Optical capacity

Marvell and GlobalFoundries expanded a multi-year manufacturing agreement, with GFS adding silicon-germanium capacity for Marvell's optical products.

Substack - SemiAnalysis

SemiAnalysis flags DRAM/NVMe TAM implications from new inference architecture

SemiAnalysis previews a piece on 'Engram' architecture that extends standard token embeddings with learned multi-token lookups, where recurring local patterns retrieve vectors directly, reducing the need to reconstruct them through attention and feed-forward layers. The preview explicitly frames this as having 'New Model Architecture Implications for TAM of DRAM/NVMe,' alongside references to DeepSeek V4.1 Flash, AgentX, InferenceX and NVMe experiments.

If embedding-lookup architectures shift inference workloads toward memory and storage rather than compute, the addressable market for DRAM and NVMe could be repriced independently of GPU compute demand narratives.

Watch Whether the full piece quantifies a DRAM/NVMe TAM delta and whether subsequent model releases (e.g., DeepSeek V4.1 Flash) adopt embedding-lookup designs at scale.

Source →
public excerpt 396 chars read
Substack - Gaetano

Marvell–GlobalFoundries multi-year agreement adds SiGe capacity for optical products

Gaetano Research notes that MRVL and GFS expanded their multi-year manufacturing agreement, with GFS adding more silicon-germanium (SiGe) capacity for Marvell's optical products, and frames this as 'another reason to look at $GFS.'

Dedicated SiGe capacity tied to a multi-year Marvell commitment supports the optical/interconnect demand thesis and gives GFS a specialty foundry revenue stream less exposed to commodity logic pricing.

Watch Disclosed capacity volumes, timing of SiGe line ramp, and whether Marvell optical revenue growth tracks the added capacity in subsequent quarters.

Source →
public excerpt 225 chars read