ReportGem ReportGem 中文

REAL-TIME GLOBAL RESEARCH

Weekly Chip Chat, Weekly HW Updates from Joe, WSTS Readthroughs, VSH, Accton, Nanya Tech: US Hardware / Semiconductors Sector Specialist Commentary

Published: 2026-07-06Institution: JPMorganPages: 7Original language: EnglishEvidence page: 3

Research evidence excerpt

Weekly Chip Chat, Weekly HW Updates from Joe, WSTS Readthroughs, VSH, Accton, Nanya Tech: US Hardware / Semiconductors Sector Specialist Commentary

Joshua Meyers - Specialist Sales - US TMT AC (1-617) 310 0767 North America Specialist Sales J P M O R G A Njoshua.meyers@jpmorgan.com

J.P. Morgan Securities LLC 06 July 2026

Secretary Bessent’s bullish stance (he’s touted $750B in AI infrastructure spending and called competition with China the real

risk) — a Treasury spokesperson dismissed the findings as unvetted, though the document is reportedly finished and awaiting

approval before eventual public release -NOTUS

• Meta Compute: Everyone Wants To Be A Neocloud.The META selloff seems overdone. The company has contracted 5GW+

of capacity in H1 2026 alone, with capex set to surge further in 2027 — driven by four high-margin use cases (frontier AI

training, RecSys scaling, a Bedrock-style Claude resale deal with Anthropic, and SpaceX-style on-demand compute sales),

meaning Coreweave, Nebius, and other Neoclouds should see continued RPO growth rather than displacement. The RecSys

opportunity is particularly hardware-intensive: Meta’s new GEM/HSTU architecture unlocks LLM-like compute scaling laws

for ad recommendation (previously impossible), suggesting a profitable >10x increase in AdRec GPU demand is coming, with

each incremental gigawatt directly translating to measurable ad revenue gains. Compute deals — priced at ~$50B/GW

annually with 90-day cancellation clauses — represent a structurally new high-margin revenue model that only Oracle and

Meta can realistically execute at scale, creating a powerful incentive loop to keep building out hardware aggressively

regardless of whether Meta Superintelligence Labs succeeds or fails - SemiAnalysis

The English excerpt is extracted automatically from the cited source page and may contain layout or recognition errors. It is never batch translated.

Open report viewer