🍪 TWiC: TPU v8, Intel Rises Up, Marvell-Polariton

Google TPU v8: training vs inference hardware split, Virgo networking

—First TPU generation with two distinct SKUs: v8t for training, v8i for inference — signals that training and inference have fundamentally different hardware requirements
—TPU 8i carries 3x more on-chip SRAM than v8t, enabling fast decode without specialized inference chips like Groq LPU; Vik argues SRAM is the key, not exotic architectures
—TPU 8i HBM capacity is higher than training chip — memory bandwidth is critical as more users and agents hit the same inference endpoint
—Virgo scale-out architecture replaces Jupiter network using Optical Circuit Switches (OCS) with high switch radix, reducing networking layers and latency
—Boardfly scale-up: 4 TPU 8i per board (copper) → 8 boards per rack (AEC) → 36 racks per pod (OCS) = 1,152 TPUs per pod; OCS is the structural backbone throughout

$GOOGL ↑看多中期

TPU v8 dual-SKU design shows architectural maturity; Virgo + Boardfly networking innovations position Google's own infrastructure ahead of relying on merchant silicon

AI算力 ↑看多中期

Training vs inference hardware bifurcation is becoming structural — inference chips need more SRAM and HBM, training chips can trade memory for more compute tiles #催化劑

光通訊 ↑看多中期

OCS (Optical Circuit Switches) becoming the structural networking substrate for hyperscale AI pods; Google's Virgo architecture makes this a long-term shift #催化劑

Intel的復甦：agentic AI CPU需求 + EMIB封裝 + 18A良率回升

—Agentic AI surge is a windfall for Intel — demand for x86 CPUs for inference far exceeds supply, precisely aligning with Intel's core competence
—Lip-Bu stated on earnings call that AI-driven business is now 60% of revenue and grew 40% YoY
—Intel quietly hiring Eric Demers (ex-Qualcomm GPU lead) as Chief GPU architect — early-stage but signals ambition in datacenter GPU

$INTC ↑看多中期

TSMC CoWoS bottleneck is driving order backlog to Intel's EMIB advanced packaging — structural beneficiary of AI基建 capex surge 18A yields reportedly improving; if execution holds, Intel has three converging assets: x86 franchise, EMIB packaging, and fab network AI-driven revenue = 60% of total, +40% YoY per latest earnings #催化劑 Hiring ex-Qualcomm GPU lead Eric Demers as Chief GPU architect — nascent but directionally significant for long-term datacenter GPU ambitions

半導體 ↑看多中期

Intel's recovery is a case study in how an unexpected demand catalyst (agentic AI CPU) can reverse a company's trajectory when it aligns with existing capabilities #比較

AI基建 ↑看多中期

$INTC EMIB packaging demand rising as $TSM CoWoS remains bottlenecked — advanced packaging outside TSMC is becoming a real alternative #比較

Marvell收購Polariton：電漿光子調製器推進3.2T光連接

—Polariton's plasmonic electro-optic modulators use Surface Plasma Polaritons (SPPs) — light interacts with electrons in metal electrodes through a dielectric waveguide
—Key advantages: extremely small footprint → low parasitic capacitance → high-speed modulation; and much lower power dissipation vs conventional modulators
—Marvell continues M&A spree in photonics as it pushes toward next-gen 3.2T optical interconnects

$MRVL ↑看多長期

Acquiring Polariton (Swiss plasmonic startup) adds novel EO modulator IP to its photonics stack — deepens moat for 3.2T 光通訊 networking Consistent M&A pattern: Marvell systematically acquiring photonics capabilities rather than developing everything in-house

光通訊 ↑看多長期

Plasmonic EO modulators represent the next wave of optical innovation — smaller, faster, lower power; Marvell's acquisition brings this into commercial AI基建 roadmaps