AMD abandons HBM for inferior LPDDR5x as AI monster devours precious high-bandwidth memory
THE SO WHAT
HBM scarcity is now reshaping chip design—AMD shifting some SoCs to LPDDR5x means bandwidth becomes a first-order constraint for many workloads. If you're building around non-HBM parts, revisit model sizes, batch strategies, and latency SLOs rather than assuming drop-in parity.
READ THE SOURCE
MORE FROM THE WIRE
Deep & Emerging TechMicrosoft Reports a Massive 25 Percent Jump in Emissions
A 25% emissions jump tied to data center power use makes AI buildout a carbon story, not just a capex story. If you’re scaling AI workloads, your board will start asking for a carbon budget alongside your compute budget—get ahead of it with location, energy mix, and efficiency assumptions baked into planning.
Deep & Emerging TechSK Hynix Debut Is a Bet That AI Breaks Boom-and-Bust Chip Cycle
A record US listing and a 13% day-one pop for a memory player is a bet that AI demand can smooth the historic DRAM/NAND whiplash. If you’re a large AI buyer, assume memory pricing power persists longer than past cycles—lock in supply and revisit TCO models that assumed reversion to old norms.
Deep & Emerging TechNuclear Energy Services Firm Holtec Joins Sector’s IPO Rush
Data center power demand is creating a financing window for nuclear services players like Holtec to tap public markets. If your AI roadmap assumes cheap, abundant grid power, start mapping exposure to nuclear buildout timelines and regulatory risk, not just GPU supply.
Deep & Emerging TechShionogi-Backed Apnimed Files for IPO to Fund Sleep Apnea Drug
A late-stage sleep apnea drug filing for IPO signals public markets are still open to focused, near-commercial biotech stories. Healthtech and AI-in-drug-discovery teams should note: clear path-to-market and specialty focus still price better than broad platform narratives.